Evaluating Sensor Data in the Context of Mobile Crowdsensing [original]

Faculty of

Engineering, Com-

puter Science and

Psychology

Institute of Databases

and Information Sys-

tems

Evaluating Sensor Data in the Context of

Mobile Crowdsensing

Master Thesis at Ulm University

Submitted by:

Maximilian Blasi

maximilian.blasi@uni-ulm.de

904922

Reviewers:

Prof. Dr. Manfred Reichert

Prof. Dr. Rüdiger Pryss

Supervisor:

Robin Kraft

2022

Version October 24, 2022

Composition: PDF-L

EX2

Acknowledgement

IwanttothankProf.Dr.ManfredReichertandProf.Dr.RüdigerPryssforreviewing

my work. My sincere thanks goes to Daniel Zöllner and Jan-Phillip Stöhr, as well as

the rest of my friends and family, for the help and support they provided.

Furthermore, I want to thank my supervisor Robin Kraft for his advice, guidance and

the possibility to work on this topic.

iii

Abstract

With the recent rise of the Internet of Things the prevalence of mobile sensors in our

daily life experienced a huge surge. Mobile crowdsensing (MCS) is a new emerging

paradigm that realizes the utility and ubiquity of smartphones and more precisely their

incorporated smart sensors. By using the mobile phones and data of ordinary citizens,

many problems have to be solved when designing an MCS-application. What data is

needed in order to obtain the wanted results? Should the calculations be executed

locally or on a server? How can the quality of data be improved? How can the

data best be evaluated? These problems are addressed by the design of a streamlined

approach of how to create an MCS-application while having all these problems in mind.

In order to design this approach, an exhaustive literature research on existing MCS-

applications was done and to validate this approach a new application was designed

with its help. The procedure of designing and implementing this application went

smoothly and thus shows the applicability of the approach.

Contents

1 Introduction 1

2 Handling Sensor Data in the Context of Mobile Crowdsensing 3

2.1 Methodology.............................. 3

2.2 MCS Applications - Systematic Research . . . . . . . . . . . . . . . 5

2.2.1 Goals.............................. 6

2.2.2 Sensor Utilization . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.3 Quality Assurance . . . . . . . . . . . . . . . . . . . . . . . 13

2.2.4 TimeConstraint........................ 14

2.2.5 Processing Device . . . . . . . . . . . . . . . . . . . . . . . 15

2.2.6 Evaluation Method . . . . . . . . . . . . . . . . . . . . . . . 16

2.2.7 ReportingMetric........................ 19

2.3 Advanced Techniques for Data Management . . . . . . . . . . . . . 21

3 Problem Approach 23

3.1 PlanningPhase............................. 23

3.2 Runtime and Post-Runtime Phase . . . . . . . . . . . . . . . . . . . 29

4 Application of the Approach 33

4.1 Dataset................................. 33

4.2 Implementation ............................ 34

5 Conclusion 47

A Appendix 49

Bibliography 60

1 Introduction

With the recent rise of the Internet of Things the prevalence of mobile sensors in

our daily life experienced a huge surge. Inspired by this trend mobile crowdsensing

(MCS) [39] is proposed in order to realize a cheaper and more flexible data acquisition

and analysis than with traditional sensor networks [69]. Guo et al. [46] define MCS as

follows: "A new sensing paradigm that empowers ordinary citizens to contribute data

sensed or generated from their mobile devices and aggregates and fuses the data in the

cloud for crowd intelligence extraction and human-centric service delivery." In other

words MCS tackles specific problems by utilizing large groups of individuals. These

individuals each have a mobile device with the problem specific sensors (e.g., GPS,

accelerometer,gyroscope,andmagnetometer)andaretaskedwithexecuting

some simple sensing tasks. The different habits of the participating individuals cause

heterogeneity in the data and the "mobile"-aspect of the participants requires the data

coverage and the data quality to be controlled.

An example for the application of MCS is traffic monitoring. The intuitive way of

monitoring the traffic would be the installation of traffic monitoring cameras or some-

thing similar. But this approach would require a lot of money investment in these

cameras. With the help of MCS this task can be done way cheaper and easier, all

that is needed are participants that are willing to share their GPS (Global Positioning

System) data while driving in a car. If a lot of participants in close proximity are only

advancing slowly, the possibility of a traffic jam in this region is high.

Designing an MCS-application is no easy task and comes with many difficulties. The

first problem to consider is what sensor data is necessary to achieve the wanted result.

As the participants of such an application are in most cases volunteers sensing with

their own mobile phone, it is advisable to keep the used sensors to a minimum. More

sensors used mean a higher resource cost for the participant and a higher variety of

data which can possibly lead to concerns regarding the data privacy of the participants.

1Introduction

The next problem stems from the question "Should the calculations be executed locally

or after an upload on the server?", the choice of the processing device. This problem

can be boiled down to pretty similar considerations as the previous one. More local

calculations lead to a higher resource cost for the participant, but more server-side

calculations expose more data privacy issues as the raw data is uploaded and at the

same time it also increases the network requirements. These conflicting points lead to

no clear solution and instead require a consideration between the trade-offs each time.

The next problem concerns the nature of crowdsensed data itself. MCS provides the

opportunity to sense data at all times and at all places which leads to huge amounts

of data. But all the sensing occurs in completely unsupervised environments which

means that a sizeable amount of this data is probably due to unforeseen circumstances

faulty or even missing. There is a multitude of possible approaches to this problem but

the decision how to best solve it in a specific case can be highly complex. After these

fundamental decisions have been made the last problem is finally how to evaluate the

data in order to get the wanted results. This problem is extremely application specific

which is why there is no universal answer to this problem.

These problems can be summarized in the following problem statements:

•What data is needed in order to obtain the wanted results?

•Where should the calculations be executed?

•How can the quality of data be improved?

•How can the data best be evaluated?

In order to tackle these questions an exhaustive literature research in the form of a

systematic review is made in Chapter 2. In Chapter 3, the knowledge gained from this

research is used to create a problem approach in the form of a step-by-step guide for

how to develop an MCS-application. To validate this approach it is then applied in

Chapter 4. Finally, Chapter 5 concludes this work.

2 Handling Sensor Data in the

Context of Mobile Crowdsensing

This chapter focuses on already existing literature on the topic of mobile crowdsensing.

First it is explained how the research is conducted and what criteria papers need to

fulfill in order to be valid for this work. Then the papers are exhaustively studied and all

relevant information are extracted. And after that some advanced data management

techniques for MCS are presented.

2.1 Methodology

The sources for papers describing MCS-applications were selected in the manner of

a systematic review [118], based on the PRISMA Statement [85]. The evaluation of

these papers can be seen in Section 2.2.

The main research questions are:

•What are goals for MCS-applications?

•Which sensors are MCS-applications utilizing and how?

•What time constraints do MCS-applications have?

•On which processing device are MCS-applications performing their calculations?

•How are MCS-applications evaluated?

•How can the findings of the paper be best presented?

In order to fulfill these research questions the following inclusion criteria (IC) are

defined.

2 Handling Sensor Data in the Context of Mobile Crowdsensing

•IC1: Any paper explaining its own MCS-application.

•IC2: Any paper explaining systems using one or more mobile devices as sensors.

Furthermore the following exclusion criteria (EC) are used to further enhance the

applicability and quality of the chosen papers:

•EC1: Systems not using any mobile sensors.

•EC2: Systems not describing how the data is sensed.

•EC3: Any paper older than 2007.

•EC4: Articles without full (English) text availability.

•EC5: Evaluation methods need to be applicable to MCS-sensed data.

•EC6: Any not peer reviewed paper.

In order to get papers fulfilling all these criteria the following search strategy is used on

multiple scientific databases (ACM Digital Library [1], IEEE Xplore [51], PubMed [96]):

Abstract :(crowdsensú)AND(AllField :(application)ORAllField :(app))

(2.1)

Awordbeginningwithcrowdsens has to be in the abstract or title and anywhere in

the paper the word application or app needs to occur, additionally a filter is applied

showing only papers submitted from 2007 and onwards.

This returns 641 possible papers. After adding 31 additional papers through manual

search and removing all duplicate results 661 papers remain. These papers are screened

by their abstract and title in order to eliminate papers that are not relevant for this

study. After screening only 172 papers remain and these are further filtered by their

inclusion and exclusion criteria after screening the full text of the paper, leaving 117

papers (see Figure 2.1).

Independent from the systematic review, in Section 2.3 some papers are chosen to

further explore techniques for the management of data in the context of MCS.

2 Handling Sensor Data in the Context of Mobile Crowdsensing

Figure 2.1: PRISMA [85] flow diagram of studies’ screening and selection

2.2 MCS Applications - Systematic Research

This section extracts the data of the previously selected papers and compares them in

regards to the following criteria:

•Goals: The goals and subgoals the paper fulfills.

•Sensor utilization: Which sensors are utilized in the paper and how/in which

way/for what purpose are they used (e.g., GPS used for localization or to mea-

sure the electron density in the atmosphere).

•Quality assurance: Which methods have been implemented in order to assure

2 Handling Sensor Data in the Context of Mobile Crowdsensing

the quality of information.

•Time constraint: Have there been any time constraints on the evaluation of the

data in this paper (i.e., were the results needed in (near) real-time?)?

•Processing device: What parts of the data processing were done on which device

and why (e.g., smartphone, server)?

•Evaluation method: What methods were used in order to evaluate the data?

(e.g., machine learning techniques)

•Reporting metric: How were the results of the paper presented (e.g., accuracy)?

Please note that not necessarily all applicable papers are mentioned in the citations in

order to improve the readability of the text. The full list of aspects each paper fulfills

can be found in Appendix A in Table A.3 - A.11.

2.2.1 Goals

This subsection compares the different goals and subgoals of the papers. Subgoals

are smaller goals the paper accomplishes (e.g., map matching or location match-

ing)andgoalsareusedasabroadertermthatencompassesmultiplesubgoals(e.g.,

localization). An arbitrary number of goals and subgoals can be achieved by each pa-

per. For reasons of clarity and comprehensibility the papers are in this section further

divided by their specific application area.

The first area, Urban Sensing,comprisestechnologiestosenseandobtaindataabout

physical areas and objects in urban spaces and how humans interact with them. This

includes techniques to analyze the public infrastructure like roads [131, 49, 86] or the

WiFi density of a city [37], the waiting time for specific services [148, 147, 13] and

other specific applications like an online reposting system for fliers [45].

The next area, Indoor Localization,focusesonlocalizationtechniquesforindoor

environments. This is a non-trivial problem as normal localization techniques have

many problems due to sensor inaccuracies within buildings leading to imprecise data.

Indoor localization techniques includes localization on an indoor map [101, 146], the

reconstruction of indoor maps [40, 19, 146], or other applications like generating a

map of the WiFi coverage of a floor [100] and collecting fingerprints of a specific

location [138].

2 Handling Sensor Data in the Context of Mobile Crowdsensing

The third area encompasses technologies regarding Environmental Monitoring.En-

vironmental Monitoring is normally done by using Wireless Sensor Networks (WSN),

but the installation and maintenance of these is expensive and thus MCS is often

used to circumvent these costs. These applications can range from analysing nightlife

behaviour of participants [108], detecting beautiful places in the city [78], and even

measuring the electron count in the ionosphere [88].

The last area, Social, Security and Healthcare (SSH),includesallapplications

regarding the physical and mental well-being of participants [95, 48], as well as appli-

cations for disaster relief [21, 123], detecting diseases [68, 31], help to observe huge

crowds [132, 15, 137], letting people report events they witness [84], or determining

the relationship between two people [28].

The number of papers per area can be seen in Table 2.1 and common goals of MCS-

applications can be seen in Table 2.2. Furthermore in Table A.1 in Appendix A all

papers can be found with their appropriate area.

Urban Indoor Environment SSH

Papers per area 51 (44%) 16 (14%) 19 (16%) 31 (26%)

Table 2.1: This table shows the 117 studies included in this research by their catego-

rized area. Urban,Indoor,Environment,andSSH are used as abbrevi-

ations for the areas Urban Sensing,Indoor Localization,Environmen-

tal Monitoring,andSocial, Security, and Healthcare.Thenumberin

brackets denounces the percentage of papers in that area compared to the

overall number of considered papers.

Urban Sensing: The most common goal in urban sensing techniques is localization.

Map matching [131, 49, 86] matches the current position to a road on existing maps.

In some cases the number of possible routes can be restricted in order to have more

options to achieve this goal. One specific example case can be route matching [148],

where a list of possible routes is known. Another subgoal is simply obtaining the

position of the user, location matching [37, 13, 147]. This knowledge can be used to

extract features at a given location [49] or to get the time spent at location [147].

Another common goal is street observation.Thisincludeseverythingrelatingtoroads

like inferring new roads [131], intersection classification [49], detecting traffic anoma-

2 Handling Sensor Data in the Context of Mobile Crowdsensing

Urban Indoor Environment SSH

Localization 49 (96%) 16 (100%) 15 (79%) 22 (71%)

Street Observation 22 (43%) 0 (0%) 1 (5%) 0 (0%)

Activity Recognition 16 (31%) 4 (25%) 1 (5%) 6 (19%)

Image Analysis 9 (18%) 3 (19%) 6 (32%) 4 (13%)

Map Generation 14 (27%) 7 (43%) 4 (21%) 1 (3%)

Sound Analysis 5 (10%) 0 (0%) 3 (16%) 4 (13%)

Air Pollution 2 (4%) 0 (0%) 7 (37%) 0 (0%)

Data Collection 2 (4%) 0 (0%) 3 (16%) 9 (29%)

Table 2.2: This table shows the goals of the papers by their categorized area. Urban,

Indoor,Environment,andSSH are used as abbreviations for the areas

Urban Sensing,Indoor Localization,Environmental Monitoring,and

Social, Security, and Healthcare.Thenumberinbracketsdenounces

the percentage of papers in that area achieving this goal. The different

ways localization was achieved can be seen in Table 2.3 below and further

common goals can be seen in Table A.2 in the Appendix A.

lies [86], determine parking spaces [27, 133, 20], and road surface monitoring [6,

114].

One of the many uses of recognition is to reduce the amount of false data (i.e., sensing

data at wrong moments). Therefore activity recognition [13, 148] can be done, this is

sometimes aided by the use of sound recognition [148]. Activity recognition is often

done to help achieve other subgoals like road surface monitoring [6, 114] or turn

detection [17].

Urban Indoor Environment SSH

Location Matching 46 (90%) 3 (19%) 14 (74%) 20 (64%)

Fingerprinting 13 (25%) 13 (81%) 1 (5%) 2 (6%)

Tracking 4 (8%) 8 (50%) 0 (0%) 0 (0%)

Event Detection 4 (8%) 1 (6%) 5 (26%) 6 (19%)

Table 2.3: This table shows in which way the papers used localization, categorized

by their area. Urban,Indoor,Environment,andSSH are used as ab-

breviations for the areas Urban Sensing,Indoor Localization,Environ-

mental Monitoring,andSocial, Security, and Healthcare.Thenum-

ber in brackets denounces the percentage of papers in that area achieving

localization in that way.

2 Handling Sensor Data in the Context of Mobile Crowdsensing

Time prediction is also often done in order to give an estimated time to enhance

the user-experience. This can be done by predicting the arrival time, arrival time

prediction [148], or the waiting time, waiting time prediction [147, 13], of specific

services.

Many urban sensing applications aim to achieve map generation.Thesetobegen-

erated maps can range from WiFi coverage maps [37, 117, 64, 135] over cellular

coverage maps [128, 38] to maps highlighting the state of the road surface [6, 114]

and free parking spaces on streets [27].

Other less common subgoals of urban sensing contain determine photo quality [45],

photo tagging [45], and photo grouping [45]. A common type of collected data is

GPS traces and splitting GPS traces [49, 86] is sometimes done in order to analyze

the data for specific information.

Indoor Localization: Again, the biggest goal in indoor localization is localization.

As the usage of GPS in an indoor environment is very error-prone, a common way to

achieve this goal is by using fingerprinting. Fingerprinting obtains the users current

location by comparing current sensor readings with previously recorded sensor readings

with a corresponding location. This can be done by either WiFi-fingerprinting [101,

100], where a list of WAPs (Wireless Access Points) and their location is stored, or

Magnetic-fingerprinting [146], where the user needs to walk a bit in order to get the

location as the magnetic fingerprints is just a 3D vector and thus needs a temporal

dimension. Another way of achieving the goal is by using tracking [101, 100, 40, 19,

146, 90, 57, 60] using the accelerometer,gyroscope and sometimes magnetome-

ter to track the movement patterns of the user. One of these tracking-techniques is

Pedestrian Dead Reckoning (PDR) [101, 100]. PDR tracks the users movement by

knowing the starting location and estimating the travelled distance and direction. As

PDR continuously estimates an estimation error accumulates over time and thus a

combination of PDR and another indoor localization technique proves very beneficial.

Simple location matching [19] is also done in order to detect the rough location.

Map generation is another common goal of indoor localization. Reconstructing a floor

plan [40, 19, 146] tries to build a map of an indoor floor by using a PDR similar ap-

proach [146], estimating the travelled distance and direction, or by letting participants

record videos or photos of the environment, which are used for extracting information

2 Handling Sensor Data in the Context of Mobile Crowdsensing

from picture [40, 19] and picture concatenation [40], connect the adjoining wall seg-

ments of photos into continuous boundaries in order to obtain hallway connectivity,

orientation and room sizes. Some MCS-applications also aim to map WiFi coverage

of an indoor floor [100].

Other indoor localization subgoals are navigation [146, 90], navigating in an indoor

environment, activity recognition [101, 100], fingerprint collection [138], and QR code

forgery detection [138].

Environmental Monitoring: Localization is also the most common goal in envi-

ronmental monitoring, but most localization-tasks in this area are pretty simple, like

location matching [11, 26, 77, 62] is sufficient as only the location of the user is

needed and event detection [78, 67, 140, 79], detecting the location of a physical

event (e.g., flowering cherry blossoms).

Detecting air pollution is one of the biggest challenges in environmental monitoring,

as normal mobile phones usually do not have the necessary sensors to tackle this prob-

lem. Most applications dealing with this problem have an external mobile sensor that

connects to measure the necessary data [67, 11, 62, 97, 134], but some applications

try to deal with this by analysing images in order to detect pollution in the air [47,

66].

Image analysis is not only used to help with the detection of air pollution, but also for

other purposes, like analyze the brightness level of a video [108], analyze the loudness

level of a video [108], or simply extracting information from picture [78, 47, 26, 66]

in order to detect a specific feature in the photos.

Other subgoals include conduct a questionnaire [108], detect point of interest [78],

expand area of interest [78], location matching [78], and measure electron count in

ionosphere [88].

Social, Security, and Healthcare (SSH): As in all other areas localization is the

most common goal in SSH applications. Location matching [48, 28, 84, 132, 31,

50] is often required to simply get the current location of the participant. In some

cases no exact location is needed, but just the knowledge if the participant is within a

certain area is enough, which is called geofencing [15]. Also, other event detection [84,

2 Handling Sensor Data in the Context of Mobile Crowdsensing

68, 21, 123, 34, 56] methods like swipe localization [84], where multiple participants

indicate a direction in which an event is taking place, are often executed.

The second most frequent goal in SSH applications is data collection.Thisgoalis

rather simple but can be achieved by many different methods like conduct a question-

naire [95, 104] or simply data collection [95, 9, 104]

Other subgoals of this area include activity recognition [48, 15], detect nearby peo-

ple [28], infer relationship [28], determine swipe direction [84], detect nearby Bluetooth

devices [132], and estimate crowd density [132].

2.2.2 Sensor Utilization

Before explaining how the sensors are used in MCS-applications the sensors are first

introduced.

Sensor Introduction: Nowadays mobile phones are equipped with a multitude of

sensors which can be used in order to sense all kinds of data. Most currently available

mobile phones come with GPS,gyroscope,accelerometer,magnetometer,cam-

era,microphone,fingerprint,barometer,ambient light,andtouchscreen [72].

The GPS [74] sensor sends signals to satellites in order to know its current distance to

multiple satellites. This distance is used to calculate the exact location of the device

in the form of longitude, latitude, and altitude. The gyroscope [126] measures the

angular velocity around the three axis within a local coordinate system defined by the

device. In a similar fashion the accelerometer [124] measures the acceleration along

the three axis within a local coordinate system. The magnetometer [127] measures

the strength of the magnetic field around the phone by recording the magnetic field

values about the corresponding phone axis. This information can be used to obtain

the phones absolute direction related to the earths geomagnetic field and thus this

sensor can function as a compass. The camera can be used to take photos or videos

and the microphone is necessary to record sounds. The fingerprint [72] sensor is a

type of biometric recognition system and is used to identify the user. Changes in the

atmospheric pressure in the surroundings are detected by the barometer [72]. The

ambient light [125] sensor detects the illuminance, current ambient light level around

the phone, in lux. Even the touchscreen [72] of a phone is a sensor in itself and most

2 Handling Sensor Data in the Context of Mobile Crowdsensing

touchscreens distinguish between three types of input, namely tapping (clicking any

location on the screen), multitouch (clicking multiple locations on the screen at once),

and gesture (drawing a certain pattern on the touchscreen).

Sensor Application: The GPS sensor can be used for map matching [131, 49, 86]

and location matching [147, 37, 19, 78, 48, 84, 132], but simply using GPS can have

errors if the exact location is relevant and thus [147] proposed a possible solution by

using the center of consecutive GPS readings. In order to measure electron count in

ionosphere [88] dual-frequency GPS can be used. For this, GPS signals are sent

on two different frequencies to the receiver and the delay between the arrival of these

two signals can be used to calculate the electron count.

Another option used for location matching [13] is the usage of WiFi to detect WAP-

locations or directly detecting a specific WAP. The WiFi sensor can also be used for

detecting WiFi density [37], getting the currently detected WAPs, for route match-

ing [148], which fingerprints cell tower IDs, and for normal WiFi-fingerprinting [101,

100], associating a list of WAPs with a specific location and using this for localization.

GPS and WiFi can also be used together for location matching [28] or geofencing [15]

to achieve even more accurate results.

The magnetometer can be used for Magnetic-fingerprinting [146], which works just

like the WiFi equivalent with the one exception that a temporal dimension is needed,

i.e., the participant needs to walk the path for a little while in order to get the location.

WiFi and the magnetometer can also be combined to form a combined fingerprint

for fingerprint collection [138].

Activity recognition is most often done by using the accelerometer [148, 13, 100, 48,

15], and this can be aided by utilizing the microphone for sound recognition [148].

Another way to use the accelerometer is to determine the tilt angle of the phone, this

information combined with the magnetometer can be used for swipe localization [84].

A multitude of sensors can be used for movement tracking. With just the gyroscope

it can be detected if the participant makes a turn [146]. Accelerometer and gyro-

scope can be used together to measure distances and orientation between start and

finish [40]. Accelerometer,magnetometer and optionally gyroscope can all be

utilized together for PDR [101, 100].

Other sometimes used sensors are the power sensor to detect whether a phone is

2 Handling Sensor Data in the Context of Mobile Crowdsensing

charging [49], the camera can be used to make photos [45, 40] and videos [19, 108,

78], the microphone can record ambient sound [95], Bluetooth can detect nearby

Bluetooth devices [132], and a combination of the accelerometer,themagnetome-

ter and the ambient light sensor can be used to determine photo quality [45].

2.2.3 Quality Assurance

In order to lessen the impact of outliers on the data, the accumulated data from all

participants can be aggregated [19, 84], averaged [147], or clustered [131, 49]. These

clusters can then be used for further computations like automatically assigning the

best possible parameters [49].

This also applies to all cases of fingerprinting [101, 100, 146, 138], the more partic-

ipants upload their data the better the fingerprints in the dataset get and the effect

of outliers can be further reduced. This can be improved even further by choosing

multiple representatives for the fingerprints by using e.g., affinity propagation [146].

Localization tasks can in most cases minimize localization errors by using multiple

localization techniques (e.g., GPS and fingerprinting) at the same time [100, 15].

Complex localization tasks, like PDR [100], can be aided even further. Previous

knowledge of the map or activity recognition can provide the opportunity to limit the

error accumulation caused by noisy sensors.

In order to avoid false data, activity recognition can be done to detect if the participants

are doing the to be detected activity (e.g., standing in line) [148, 13]. A simple way

to be able to retrace possible errors is by documenting the GPS-location error [37]. If

participants are supposed to execute tasks at specific locations the current detected

fingerprint can be compared to by other participants previously detected fingerprint

at the same location [138] as GPS-spoofing can be easily done. If Bluetooth devices

in the vicinity should be detected [132] the signal strength can also be considered in

order to better be able to extrapolate the number of devices in a bigger area than the

sensor can detect. To reduce the noise of sensors a median filter [48] can be applied

to the sensor readings.

Manual confirmation of the participant [19, 84] can also be a good way to confirm

the correctness of the data. Moreover, not all papers mentioned any steps taken in

order to account for data quality [86, 45, 40, 108, 78, 95, 28].

2 Handling Sensor Data in the Context of Mobile Crowdsensing

2.2.4 Time Constraint

Most MCS-applications have no time constraints [49, 147, 45, 37, 100, 40, 19, 138]

as they are used to gather non-time sensitive information, e.g., to update maps or get

information just for data analysis reasons.

Other MCS-applications aim to relay the gathered information, e.g., estimated current

waiting times, as fast as possible to the participants [148, 13, 78, 48, 15].

Many MCS-applications have a combination of real-time components and non-time

sensitive components [131, 86, 101, 146]. In these cases the reason for the time

constraints of the specific components is in most cases the same as mentioned above,

anon-timesensitiveinformationisrequiredinordertofurtherprocessthetimesensitive

information, e.g., calculating the typical routing behaviour in order to detect anomalies

in real-time [86]. In the case of GROPING [146], the non-time sensitive component

is an update to the fingerprint map, this is at first time-sensitive as the fingerprint

database needs to first be filled to execute the other components, but later on it is

not important how quick this database is updated with newer fingerprints. In order to

enable real-time calculation, CrowdAtlas [131] omits the clustering step in their phone

application.

Table 2.4 shows the time constraints per area. It can be seen that most applications

either have no time constraint at all or only some components are time relevant.

Only about 20% of all the considered papers present their application as completely

real-time dependant.

Urban Indoor Environment SSH

Real-Time 10 (20%) 1 (6%) 4 (21%) 10 (32%)

No Time Constraint 25 (49%) 10 (62%) 10 (53%) 15 (48%)

Mixed 16 (31%) 5 (31%) 5 (26%) 6 (19%)

Table 2.4: This table shows the time constraints of the applications, categorized by

their area. Urban,Indoor,Environment,andSSH are used as abbrevi-

ations for the areas Urban Sensing,Indoor Localization,Environmen-

tal Monitoring,andSocial, Security, and Healthcare.Thenumberin

brackets denounces the percentage of papers in that area with the corre-

sponding time constraint.

2 Handling Sensor Data in the Context of Mobile Crowdsensing

2.2.5 Processing Device

The choice of the processing device is a really important decision and the choice made

in the reviewed papers are highlighted here. The one thing all applications have in

common however, is that the sensing of data is in always done by the phone or external

sensors connected to the phone.

Most applications [131, 49, 86, 148, 13, 101, 19] preprocess the data locally on the

phone in order to lessen the amount of data that needs to be uploaded and in order

to lessen the burden on the participants phones, in cases where data transfer can

be avoided at all calculations can be executed on locally if avoiding data transfer is

ahigherprioritythanavoidingcomputations. Thesepreprocessingandcalculations

subgoals can include route matching [148], splitting GPS traces [49, 86], sound recog-

nition [148], conduct a questionnaire [95, 108], record ambient sound [95], determine

swipe direction [84], and detect nearby Bluetooth devices [132].

The server executed calculations are usually the more expensive calculations and can

include detecting traffic anomalies [86], arrival time prediction [148], time spent at

location [147], waiting time prediction [147, 13], determine photo quality [45], photo

tagging [45], photo grouping [45], detecting WiFi density [37], reconstructing a floor

plan [40, 19, 146], picture concatenation [40], map WiFi coverage of an indoor

floor [100], navigation [146], QR code forgery detection [138], analyze the bright-

ness level of a video [108], analyze the loudness level of a video [108], detect point of

interest [78], expand area of interest [78], measure electron count in ionosphere [88],

swipe localization [84], detect nearby people [28], and estimate crowd density [132].

Multiple subgoals are often executed on either the phone or the server. These include

map matching [131, 49, 86], location matching [37, 13, 147, 19, 78, 48, 28, 84,

132], inferring new roads [131], extract features at a given location [49], extracting

information from picture [40, 19, 78], intersection classification [49], activity recogni-

tion [13, 148, 101, 100, 48, 15], fingerprinting [101, 100, 146, 138], PDR [101, 100],

and geofencing [15]. Especially activity recognition [13, 148, 101, 100, 48, 15] is most

of the time executed in order to check for prerequisites for the sensing of data (e.g.,

standing in line) and thus often done locally.

Table 2.5 shows an overview of the processing devices per area and surprisingly about

50% of the regarded MCS-applications have done no local preprocessing before up-

loading the data to the server. This is often the case when the main purpose of the

2 Handling Sensor Data in the Context of Mobile Crowdsensing

application is data collection [145, 10] as that goal needs no processing, or a goal of

the application is to not impede with the normal usage of the phone and thus not

requiring many computational resources [100, 40, 146].

CrowdAtlas [131] uses an interesting hybrid concept where, in addition to the process-

ing on the server after uploading the preprocessed data, all calculations can be done

locally on the phone with less accuracy in order to enable real-time calculations. In

addition, SmartRoad [49] tested different data aggregation schemes where they dif-

fer between calculations done locally on the phone and calculations of the aggregated

data on the server. The findings were as expected, the earlier the data was aggregated

(i.e., uploaded to the server) the better the results.

Urban Indoor Environment SSH

Local Preprocess 28 (55%) 7 (44%) 8 (42%) 15 (48%)

Direct Upload 23 (45%) 9 (56%) 11 (58%) 16 (52%)

Table 2.5: This table shows the processing device of the applications, categorized by

their area. Urban,Indoor,Environment,andSSH are used as abbrevi-

ations for the areas Urban Sensing,Indoor Localization,Environmen-

tal Monitoring,andSocial, Security, and Healthcare.Thenumberin

brackets denounces the percentage of papers in that area with the corre-

sponding processing device.

2.2.6 Evaluation Method

The collected data needs to be processed differently depending on the to be achieved

goals and subgoals of the application. Due to the broad field of application of MCS a

huge variety of techniques are utilized in MCS-applications.

One very commonly needed technique is classification. Classification techniques have

multiple outcome sets and matches a new observation in the set it belongs to, e.g.,

deciding whether something is large or small. More complex classification tasks (e.g.,

intersection classification [49], photo tagging [45], infer relationship [28], ...)areof-

ten done by using existing machine learning techniques (Random Forest, AdaBoost,

Support Vector Machine, ...). The less complex tasks can often be achieved without

the use of machine learning and using a simple algorithm to evaluate sensor data.

For activity recognition [148, 13, 101, 100, 48, 15] it is often enough to analyze the

2 Handling Sensor Data in the Context of Mobile Crowdsensing

accelerometer readings to discern between very few different cases, like walking vs.

idle or steady motion vs. unsteady motion, and this can be easily done by simple

algorithms like Signal Magnitude Area or Signal Magnitude Vector. Sound recogni-

tion [148] is in this regard very similar to activity recognition with the one exception

that the per microphone recorded sound needs to be converted into frequency first,

this can be done by Fast Fourier transform. If the recorded data can vary slightly each

time a top-k matching algorithm, like Smith-Waterman algorithm, proves also very

helpful when making classifications, e.g., route matching [148].

Another often needed function in MCS-applications is clustering, the ability to group

multiple similar objects with each other. A pretty standard use of clustering can be

seen when inferring new roads [131]. This works by gathering multiple GPS-locations

and clustering them according to Single-Linkage clustering, a hierarchical clustering

approach that clusters the data points closest to each other. In order to remove

duplicate images clustering can also be used for photo grouping [45], a widely used

way to achieve this is by retrieving near duplicate images via Scale-invariant feature

transform.

MCS-applications frequently ask the user to record photos or videos and therefore

techniques for image analysis are needed. Analyzing these kinds of data often tends

to be a fairly complex problem, e.g., extracting information from picture [40, 19, 78],

but thankfully there is already a huge amount of existing computer vision techniques

(e.g., Structure from motion, Vanishing Line Detection, Shape Matching, Wavelet

Decomposition, Color Histograms, Fractal analysis, ...) that can be used to tackle

these problems. There are some simpler cases like analyze the brightness level of a

video [108], simply extract the brightness (intensity of the luminance channel) per

frame and make the mean of all frames. If the distance and orientation between

photos is relevant, a similar approach to picture concatenation [40], measuring the

distance and orientation between photos and use maximum likelihood estimations to

gain the relative coordinates between different photos, can be made.

Similarly sound analysis is also a fairly complex problem, which many MCS-applications

need to tackle. In almost all cases the first step towards sound analysis is by converting

the time domain signal to the frequency domain signal by using Fourier transform [148,

144, 50, 34]. Speech recognition [23] can be handled by the open source CMU Sphinx

recognizer and extracting audio as features mel-frequency cepstral coefficients [23, 137]

for further processing are other ways to handle sound data. When trying to analyze

2 Handling Sensor Data in the Context of Mobile Crowdsensing

the loudness level of a video [108] simply extracting the audio power (calculate using

the audio channel) per frame and calculate the mean of all frames is sufficient.

Most indoor localization techniques require the use of fingerprinting, sensing signals

of the local environment and using these to fill a map of the area with the recorded

signal in order to determine the current location of a user in the future. This is

mostly done by utilizing WiFi for WiFi-fingerprinting [101, 100] or magnetometer

for Magnetic-fingerprinting [146], these fingerprinting techniques can also be combined

for an increased accuracy. A good way to use WiFi-fingerprinting is recording a

vector of the five strongest Access Points (APs) of a location, retrieving the closest

matching entry chosen by euclidean distance and returning the weighted mean of the

top-k results [100]. Magnetic-fingerprinting works in a very similar way, the ambient

magnetic field is sensed and assigned to a location on a map. However, the recorded

magnetic fingerprint is just a 3D vector and while matching this vector to the map a

misalignment could very easily happen due to different walking speeds. The Dynamic-

Time-Warping algorithm can be used to tackle this problem, furthermore due to the

nature of magnetic fingerprints the localization needs a temporal dimension [146] (i.e.,

the users need to walk a bit). This method of localization also provides a good way

to avoid and detect GPS-Spoofing, the act of altering your own GPS signal to fake

your location. QR code forgery detection [138] uses this by comparing the currently

detected fingerprints and comparing them to the centroid of previous fingerprints,

should they differ too much GPS-Spoofing is suspected.

Generating maps, locating or tracking the user, and event detection is a goal of many

MCS-applications. Mapping the WiFi density to a map (detecting WiFi density [37],

map WiFi coverage of an indoor floor [100]) is extremely similar to the mapping

process of WiFi-fingerprinting and is sometimes [100] even combined. In order to

visualize such a coverage map the Inverse Distance Weight based spatial interpolation

can be used. Reconstructing a floor plan [40, 19, 146] is a very complex undertaking

and thus has many different approaches. A common approach is by using user recorded

photos or videos and analyze these in order to gain information of the map layout [40,

19]. These information can then be used to connect the adjoining wall segments of

landmarks into continuous boundaries by using combinatorical optimization and use

probabilistic occupancy maps to obtain hallway connectivity, orientation and room

sizes. Another approach is stitching sensed fingerprints, mapped to a segment of a

map, together when parts of their segments overlaps [146]. These indoor floor plans

2 Handling Sensor Data in the Context of Mobile Crowdsensing

can be used to help the user with navigation [146], this can be done by tracking

the users current location by any means (e.g., fingerprinting, PDR, ...) and using

that knowledge combined with the map to execute any pathfinding algorithms, like

Dijkstra’s algorithm.

Other interesting, but less often used evaluation methods include the need to deter-

mine the swiping angle on the phone and methods to estimate waiting time. The

swiping angle can be calculated by Equation 2.2. This calculation only works if the

phone is held parallel to the ground, which can be confirmed by using the accelerom-

eter and checking if the condition Equation 2.3 is satisfied. Waiting time can be

estimated by many different approaches. One approach is to use regression to quan-

tify the releationship between wait time and the time of day and using this to calculate

a weighted euclidian distance and average the top-k closest results [13]. Another ap-

proach is recording waiting times at a specific time of day at a specific location. If

there is enough data at a location simply making the average yields a solid result.

But if the data is sparse at a location the knowledge from the other locations can be

used and the problem can be compared to a recommender system, thus context aware

collaborative filtering and factorization can be used to utilize contextual features to

improve the waiting time prediction [147].

3(swiped angle on the touchscreen with respect to the phones x-axis)≠

(angle between phones y-axis and magnetic north)4mod 2ﬁ

(2.2)

abs(mean(acck)) Æ0.05g,std(acck)Æ0.1g,k œ{x, y}(2.3)

2.2.7 Reporting Metric

The way the results of an MCS-application are reported is very dependent on the

to be reported subgoals of the application. Subgoals that cluster, classify or sim-

ply discern between two cases (extract features at a given location [49], intersection

classification [49], route matching [148], activity recognition [100, 148, 15], sound

recognition [148], photo grouping [45], photo tagging [45], detecting traffic anoma-

lies [86], QR code forgery detection [138], detect point of interest [78], extracting

2 Handling Sensor Data in the Context of Mobile Crowdsensing

information from picture [78] used to classify the image, infer relationship [28], es-

timate crowd density [132] which estimates by classification, geofencing [15]) can

evaluate their results by reporting their accuracy. The same principle can be applied

when reporting the correctness of predicted shapes (Reconstructing a floor plan [40,

19]). The accuracy can be reported as Precision [49, 148, 45, 100, 40, 19], Recall [86,

45, 40, 19, 138, 78] or F-Measure [45, 40, 19]. Precision depicts the number of cor-

rect class predictions that actually belong to the correct class and Recall depicts the

number of correct class predictions out of all positive outcomes, while the F-Measure

provides a statistical number that regards the Precision and the Recall. The accuracy

of swipe localization [84] was reported as reporting rate and detection rate according

to Equation 2.4.

reportingRate =(#reportedEvents/#trueEvents)

detectionRate =(#reportedTrueEvents/#trueEvents)(2.4)

Sound recognition [148] can be further evaluated by reporting the distance to the

target while reporting the accuracy and the position of the phone (e.g., in hand or in

bag).

All estimation techniques, like localizations (PDR [101, 100], fingerprinting [146], lo-

cation matching [84]), extract information from video/photo (extracting information

from picture [40], picture concatenation [40]), or time estimations (arrival time predic-

tion [148], waiting time prediction [147, 13]), are best reported by the estimation error.

This is most often done by simply plotting the estimation error [101, 100, 40, 146, 84]

(in the respective unit, e.g., meter). For more specific evaluation the mean absolute

error [148, 13, 147], median absolute error [13], and the standard deviation [147] can

be used. The mean absolute error represents the sum of absolute errors divided by the

sample size and the median absolute error represents the median of all errors, while

the standard deviation indicates how accurately the mean represents sample data.

An indirect way to evaluate a systems component is by reporting the end results

with and without the usage of said component. This is often done for activity recogni-

tion [13]. In cases where statistical evaluation may be hard due to the lack of a ground

truth, e.g., generating a new map [131, 37, 100, 40, 19, 146] or service quality [146,

138, 108, 95], a manual comparison can also be made.

2 Handling Sensor Data in the Context of Mobile Crowdsensing

2.3 Advanced Techniques for Data Management

In this section different ways of managing crowdsensed data are explored. This section

names some more advanced techniques that can be applicable to MCS-applications.

The network requirements for real-time data management can increase quickly and

thus Sarkar et al. [109] propose an approach for this problem by assigning a device

to a variably designed cluster. Each device sends their data to a cluster head which

aggregates the data in order to reduce the transmission cost of the data to the server.

The different sensors used for mobile crowdsensing can cause a significant imbalance in

the sensed data. Patel et al. [89] propose different solutions for this problem. The first

way proposed to approach this problem is a "Data-Level Solution", balancing the data

to alter the original distribution of data to achieve better classification for imbalanced

datasets. The next approach is a so called "Algorithmic-Level Solution", altering

the used processing algorithms to achieve better results. Finally "Cost-Sensitive and

Ensemble Solutions" are explored, merges different approaches in order to achieve

better accuracy and more reliable results. For each of the different solution approaches

many different algorithms are presented in [89].

In crowdsensing vast amounts of data can be gathered and thus a high data dimension-

ality with high computing complexity can be a problem for many MCS-applications.

AwaytotacklethisproblemisbyusingdimensionreductiontechniqueslikePrin-

cipal Component Analysis (PCA) [55]. Less dimensions in the data lead to simpler

computations which makes PCA a good choice for data exploration and data pre-

processing before applying more complex statistical or machine learning tools [55].

Possible advantages of PCA include:

•Obtaining better insight into the data.

•Identifying potential issues with the data such as artifacts or outliers.

•Reduce the number of predictors of a linear regression model, therefore avoiding

the multicollinearity problem and reducing the risk of overfitting.

•Obtain higher accuracy and efficiency for all the methods based on computing

distances between data points (e.g., K-nearest neighbors, K-means, Support

Vector Machine).

2 Handling Sensor Data in the Context of Mobile Crowdsensing

•Enable the algorithm applied to the data to run faster, which provides additional

time for parameters optimization or model benchmarking.

•Less storage space requirement.

The quality of data can also be a problem in MCS-applications, as the data is almost

exclusively sensed in unsupervised environments. Therefore Kong et al. [58] provide

an in-depth look at many methods to improve the quality of data. These methods

include missing data reconstruction, fault data detection, data privacy preservation,

multidimensional data conversion, and efficient task allocation.

Wireless Sensor networks are a vastly more explored topic than MCS and due to the

similarity in terms of data collection and evaluation most techniques for data man-

agement in wireless sensor networks work for MCS. Akyildiz et al. [5] provide a huge

in-depth analysis to wireless sensor networks and many transferable techniques includ-

ing error control in transmission [5, Chapter 6], data compression before transmission

and how to query sensors [5, Chapter 9], and time synchronization [5, Chapter 11] are

explained in great detail.

3 Problem Approach

When considering to employ an MCS-application many options need to be regarded.

This section gives an overview of the aspects to consider in the different phases an

MCS-application can have. Getting the most out of your application is possible when

planning the app from scratch (called planning phase), but this also means that the

most aspects need to be considered. An already existing app can be modified to further

enhance the wanted outcome after the app was deployed (called runtime phase).

Many previously existing MCS-application have already collected vast amounts of data

and not completely explored the knowledge hidden in it, thus evaluating already sensed

data can also yield new insights (called post-runtime phase).

3.1 Planning Phase

When planning to develop an MCS-application from scratch a lot of things need to

be considered, including what goals the application should fulfill, the choice of the

processing device, and what methods should be taken to ensure the quality of the

data and the result.

Goals: The first thing to consider is what goals and subgoals the application should

fulfill and in what way these should be accomplished. Many goals need other goals or

subgoals in order to be achieved, these connections are described in the following and

can also be seen in Figure 3.1.

If the location of the user is needed for the application localization needs to be ex-

ecuted. This can be done in many different ways, the standard approach in big and

open areas (e.g., a city) is location matching by using GPS and in some cases WiFi

or cell tower signals. If the application is supposed to work in an indoor environment

3 Problem Approach

GPS is very unreliable and fingerprinting and tracking methods are preferable. In

some cases the location of the user is not to be identified by coordinates but another

concept, e.g., is the user on a train. This can be a very application specific problem,

but the most commonly used techniques to solve this problem are activity recognition

(e.g., detect the movement patterns of the specific vehicle) and sound analysis (e.g.,

detect the sound of the IC card reader when boarding a bus [148]).

Street observation tries to determine different states of the road and in order to do this

the current location of the participant is always relevant. Thus localization of some

kind is always required and depending on the specific subgoals (e.g., road surface

monitoring)activity recognition is often required to detect the specifically observed

road state.

Trying to measure air pollution with MCS is no easy task to fulfill by using smartphones

as mobile sensors. When using only smartphone intern sensors this goal is almost

always achieved by using image analysis to detect pollution in the air, but the standard

solution is to use an external sensor that connects to the smartphone via Bluetooth. In

this way even more fine-grained information regarding the air pollution, like what kind

of substance is polluting the air to what degree, can be collected. This measurement is

usually always coupled with the location of the sensed air pollution and thus localization

is required.

Map generation also always needs some form of localization to infer the coordinates

of the to be mapped objects or events. Typical application are WiFi/cellular coverage

maps of cities, maps containing information generated through street observation or

air pollution,orreconstructing a floor plan.

If the main purpose of the application is data collection the way to achieve this is com-

pletely dependent on the data the application wants to collect, theoretically recording

all sensor readings or conducting a questionnaire is enough to fulfill this goal.

Activity recognition,image analysis,andsound analysis are mostly used as utility func-

tions in order to aid another goal or subgoal of the application (see Section 2.2.6).

Processing Device: The second important consideration when designing an MCS-

application is the choice of the processing device, in other words whether the appli-

cation executes the calculations locally (i.e., on the smartphone) or on the server.

This decision can be made for the whole application as a whole, but in most cases

3 Problem Approach

Figure 3.1: Typical goals of MCS-applications and how they commonly relate to each

other. Squares represent goals, rounded squares represent subgoals, cir-

cles represent possibly needed sensors, diamonds represent decisions, and

dotted lines represent common but not necessary connections. (Note that

this picture only depicts common connections and not every possible con-

nection.)

it is more sensible to make this decision for each component of the app separately

as some components may have, for example, less security-relevant data or only some

components are required to provide their results in real-time. In the following the most

common deciding aspects for this decision are explained, which can also be seen in

Figure 3.2.

The first and arguably most important aspect to consider for this decision is whether

the observed component is supposed to work in (near) real-time or it has no time

constraints at all. In case the results of this component are not time sensitive this

aspect can be ignored. But if it tries to present the results in (near) real-time to the

3 Problem Approach

user the data either needs to be locally processed or a constant network connection is

required (i.e., no opportunistic uploading when connected to hotspots possible). Most

applications are not time sensitive or only have some time sensitive parts, meaning

only those parts of the application need to be especially considered when planning the

processing device.

The second aspect to consider is whether the components of the application can

even be feasibly executed locally. Components that only use the local data can be

executed locally without many problems but components requiring data from multiple

phones/users at once (e.g., clustering GPS locations of multiple users) would need to

download the rest of the data from the server while still uploading their own data for

the other users to use. In most cases there is not much reason to do this instead of

just uploading the own data to the server, letting the server make the calculations,

and just downloading the calculation results from the server. If this aspect was already

considered as not feasible the other decisions can be skipped for this component of

the application.

Another concern is the matter of data privacy, the more data of a user is upload to

the servers the more issues with data privacy can occur. Executing as many of the

calculations as possible locally helps to reduce the amount of data privacy issues.

Power and resource consumption are another aspect in terms of user-friendliness.

Users will not be content with the application if all the computational resources of

their phones are occupied by the application and the battery life of their phones is

noticeable shortened by using the application. In this aspect uploading to the server

as soon as possible would be preferable, as the expensive calculations can impact the

battery life noticeably [24].

Yet another aspect is the network requirements. Most users will not have an unlimited

amount of mobile network data and thus considerations need to be made. If the

data should possibly be uploaded anywhere and everywhere it would be preferable to

execute tasks that compress the data for further calculations locally (e.g., classification

tasks where multiple types of data are used as input and the outcome is just a class

label). Another possibility is to opportunistically upload the data when the user is

connected to a WiFi hotspot, this approach circumvents the problem of data usage

for the participant but does not allow for real-time results.

These aspects need to be considered for every component of the MCS-application

3 Problem Approach

in order to decide the processing device of the components. These decisions are

everything but trivial and it boils down to a case-to-case consideration what aspects

are prioritized over others.

Figure 3.2: Decision diagram for deciding the processing device of a component of

an MCS-application. Squares represent the considered concerns, rounded

squares represent the outcome of the previous decisions, and diamonds

represent decisions.

Quality Assurance: The ubiquitous deployment of mobile sensors for the use of

MCS means that the data is sensed in all sorts of unsupervised situations. This

provides MCS-applications with the possibility to accumulate huge amounts of data,

however due to the unsupervised characteristic of the data collection process the quality

of the collected data is everything but guaranteed. In order to improve or ensure the

3 Problem Approach

quality of the sensed data and results many precautions can be taken when developing

MCS-applications. The typical pipeline for MCS-applications considers these problems

by breaking them up in four major components, which can be seen in Figure 3.3. First,

the data needs to be sensed in the so called data acquisition stage. Second, the sensed

data is corrected or validated in the data processing stage, before missing or incorrect

data is estimated in the data imputation stage. And then finally the gathered data

from the different sensors is united and used for data analysis.

The first typical steps to improve the quality of the data can be applied when sensing

the data in the data acquisition stage. Different mobile phones have different built-in

sensor-hardware and therefore the sensor readings can differ from the expected values.

As such, some MCS-applications provide their participants with the phones used for

the sensing tasks, in order to avoid different sensor-hardware and have coherent results.

This approach ensures the homogeneity of the sensed data, but reduces the potential

amount of collected data greatly and increases the required budget for the project.

Another possible way to improve the quality of the sensed data is by carefully selecting

the participants for the data sensing tasks. By only allowing trusted participants the

quality of the sensed data can be improved dramatically, but doing so greatly reduces

the potential amount of collected data. In some cases the sensing of data is only

desired if certain conditions are met. This condition is often the presence of the

participant on a specific type of transportation vehicle, e.g., train, and thus already

existing activity recognition methods are often a good way to validate the condition.

The ubiquitousness of MCS-applications in all possible situations means that the sensed

data is not always sensed correctly, under the correct conditions, or even sensed at

all. For this reason the data processing stage focuses on validating and correcting

the sensed data. For many MCS-applications the exact location of the participant is

extremely important and the most commonly used technique to detect this location

is location matching via GPS.TheGPS readings can potentially be quite a way off

of the correct location which can cause problems for some applications. However,

the severity of this problem can easily be lessened by recording the location errors as

validation for the data. Besides just tracking the error many already existing methods

can be applied to validate the correctness of the data such as support vector machines,

artificial neural networks, Bayes classifiers, and K-nearest neighbour.

The data imputation stage tries to estimate missing or invalid data in order to improve

the results of the application, however this approach can potentially lead to even worse

3 Problem Approach

results. This threat occurs most often when the amount of correct data is sparse, but

due to the nature of MCS the available data should be more than enough to promise

an improvement. Data imputation can be a very complex field, but luckily there

already exist many techniques exactly for this problem, including Imputation Tree, K-

nearest neighbour techniques and many variants (i.e., K-nearest neighbour imputation,

sequential K-nearest neighbour method-based imputation, and K-nearest neighbour

imputation method based on Mahalanobis distance), K-means-based imputation, fuzzy

C-means clustering imputation, singular value decomposition, back propagate-based

neural networks, random recursive partitioning, maximum likelihood, and Bayesian

estimation.

In the data analysis stage the sensed and corrected data is finally analyzed in the way

the application wants. The here used methods are highly dependant on the application

and stem most often from either machine learning, pattern recognition, or data mining.

Many commonly used techniques are described in Section 2.2.6. Another often used

method to improve the results of the application is by having multiple components

fulfill the same goal by different subgoals, e.g., in order to minimize the localization

error multiple means of localization like location matching via GPS and fingerprinting

can be used in conjunction. This approach leads to better, more trustworthy results,

however this comes with the downside of more resource cost for the participant.

3.2 Runtime and Post-Runtime Phase

Runtime Phase: In many cases MCS-applications are already running and are sup-

posed to be expanded. When working with an application during its runtime mostly

the same aspects need to be considered as in the planning phase, but some options

are lost.

The first step when trying to work with an already running MCS-application should

be to explore the already collected data to get a better understanding of the already

collected data.

After getting to know the dataset the next consideration should be how the application

should be expanded, i.e., what new goals should be achieved and can these new goals

retrospectively be accomplished with the already sensed data. Figure 3.1 can be used

to help with this question, as the connection between the goals and what sensors

3 Problem Approach

Figure 3.3: Typical quality assurance methods. Dotted squares represent the pipeline

stages, squares represent the quality assurance methods, rounded squares

represent the pros (left) and cons (right) of those methods.

they use is always the same. If the new goals cannot be achieved with the currently

used sensors the application can be expanded to include more sensors in future sensing

instances. When additional components are added or existing components are changed

all the considerations for the processing device mentioned in Figure 3.2 need to be

kept in mind.

And just like in the planning phase, quality assurance is still an important aspect. The

pipeline ensuring the quality of data and the results is almost identical to the planning

phase, shown in Figure 3.3. The only two differences are in the data acquisition stage,

namely the participants can no longer be chosen and the phones can no longer be

3 Problem Approach

provided to ensure sensor equality, as the application is already distributed.

This shows that the steps that need to be taken to plan an MCS-application from

scratch and to expand an existing one are almost identical. A step by step list of what

to consider for expanding an application can be seen in Figure 3.4.

Figure 3.4: Possible considerations when expanding an MCS-application during the

runtime-phase. Squares represent the considerations and circles represent

the the steps taken for this consideration.

Post-Runtime Phase: During runtime MCS-applications can collect vast amounts

of data from multiple sensors. These amounts of data and their connection to each

other are so expansive that there is always more knowledge to be gained within the

collected data. For that reason many applications do not sense at all and only use

already collected data from other MCS-applications in order to analyze them even

further.

When developing such a post-runtime application the considerations are almost exactly

the same as for the runtime phase. The main difference is that there is no longer a

data acquisition stage, meaning that components can no longer be added or changed

to the mobile application, resulting in no new processing device considerations, the

new goals need to be achieved with the already gathered data, the sensing location can

3 Problem Approach

no longer be restricted to improve the quality of the sensed data, and error tracking

is no longer an option.

This chapter highlighted the most important aspects to consider when planning to

work with MCS-applications. The most established use cases for such applications

are covered and by following the mentioned steps the process of developing such an

application can be somewhat streamlined. When working with an already launched

or finished application it is especially important to extensively explore the already

gathered data and many of the previous aspects become irrelevant.

4 Application of the Approach

In this Chapter the approach explained in Chapter 3 is demonstrated. To do so a

dataset of already gathered data is used and the application of the post-runtime

phase approach is illustrated. The used dataset is introduced and the implementation

of the approach is shown.

4.1 Dataset

The used data was gathered by the TrackYourStress (TYS) application [93, 94, 52],

amHealthcrowdsensingplatformthatissupposedtotracktheindividualstresslevel

of users. This is tracked by sensing the environmental sound level and the GPS po-

sition, furthermore questionnaires about the current stress level and other possibly

related parameters are conducted. The progression of the TYS pipeline has two ma-

jor components [93] (see Figure 4.1), the registration procedure and the continuous

mobile crowdsensing procedure. The first component starts with a registration on

the TYS website [52] or mobile application [94]. After that the participants have to

fill in a registration questionnaire, before subscribing to the appropriate studies. The

continuous mobile crowdsensing procedure starts by choosing a notification schema,

which determines the frequency and types of the questionnaires. Then questionnaires

popup in the given time interval and while they are filled out the application senses

the environmental sound level and the GPS position. This component is looped for

as long as the participants wishes to continue in the study. For the original analysis

78 participants were included and these each have conducted 4.87 questionnaires on

average [93].

4 Application of the Approach

Figure 4.1: TrackYourStress [93, 94, 52] progression pipeline.

4.2 Implementation

The implementation is made in the Jupyter Notebook [92] environment, a web-based

integrated development environment for python 3.9.13 [99] and offers a simple, stream-

lined, and document-centric experience.

The implementation follows the approach for the post-runtime phase as mentioned

in Section 3.2. Accordingly the data is first explored, then the to be achieved goals

are defined, and then the quality assurance pipeline is executed in order to achieve the

results (see Figure 4.2).

Figure 4.2: Possible considerations when expanding an MCS-application during the

post-runtime phase. Squares represent the considerations and circles

represent the the steps taken for this consideration.

4 Application of the Approach

Data Exploration: The data is received in the form of an Excel file, a screenshot

of this file can be seen in Figure 4.3. The data contained are id,user_id,question-

naire_id,sensordata,client,andcollected_at.Theid column contains the id of the

sensing run, multiple values can be sensed within a single sensing run. The user_id

denounces the user, while the questionnaire_id denounces the type of questionnaire it

describes. The sensed data can be found in the sensordata column and is in the form of

an array containing multiple JSON strings, the sensed data contains GPS and micro-

phone readings. Information about the used client can be found in a JSON string in

the client column. And finally collected_at denounces the moment the questionnaire

was finished.

Figure 4.3: Excerpt of the TYS data file.

To explore the data Python pandas version 1.5.0[87] is used, as it has many ready-

made methods for data exploration. A quick analysis shows that there are 3070 sensing

runs, 150 unique users, and 4different values for the questionnare id. Of the sensing

runs, 2933 successfully gathered data and all 3070 collected information on the client

and the time of the questionnaire completion.

The contents of the client JSON string are explained in Listing 4.1 and further analysis

shows that about two thirds of all the collected data was sensed by iOS devices.

Listing 4.1: Structure of the client JSON string.

2"device":"name of the device",

3"name":"version number of TYS",

4 Application of the Approach

4"os":"operating system on the device"

The main content modules of sensordata are explained in Listing 4.2. The sensordata

column contains an array of multiple JSON strings and describes all data sensed

from a single sensing run. Each JSON array contains at most one GPS reading

JSON string, this reading is marked with "name" as "gps" and has a corresponding

"longitude" and "latitude" value. Furthermore it also contains an "altitude" value that

is sometimes also called "altidude", which is most likely just a typing mistake. Besides

the GPS reading, sensordata also contains any number of microphone readings,

which are marked with "name" as "microphone". Each of those JSON strings has

a measurement of the ambient noise as either "loudness" or "amplitude". All JSON

strings in sensordata also have an assigned "collected_at" value, marking the exact

time of this sensor reading. An analysis shows that within the 3070 sensing runs only

2511 GPS readings have been recorded, but a whole of 62023 microphone readings

from 2669 different sensing runs has been sensed.

Listing 4.2: Structure of the sensordata Array of JSON strings.

3"name" = "gps",

4"longitude" = "longitude value",

5"latitude" = "latitude value",

6"altitude / altidude" = "altitude value",

7"collected_at" = "time stamp of this sensor reading"

8},

10 "name" = "microphone",

11 "loudness / amplitude" = "ambient noise value",

12 "collected_at" = "time stamp of this sensor reading"

13 },

14 ...

15 ]

4 Application of the Approach

Goals: With the dataset explored it can now be considered what new goals should

be achieved by using the existing data and how these goals should be achieved.

By following Figure 3.1 from the sensors upwards, we can see that with the GPS read-

ing we can execute some form of localization,tobemorespecificlocation matching

and with the microphone data we can apply some form of sound analysis. Another

Possible goal that can be achieved with these inputs is map generation,therequired

localization is given and the data we can map is the location of the microphone.

Another possible goal that can be achieved would be some kind of data analysis that

combines the location and noise level for which there are many possibilities. The form

of this analysis can be anything, e.g., find hotspots where the noise is notably higher

than in others (which could perfectly be matched with map generation)ordetermine

if there is a relation between the mean distance a person travels and the mean noise

difference that person experiences. This thought process according to Figure 3.1 is

visualized in Figure 4.4.

The goals chosen for this work were map generation in order to map the location of

the microphone readings and data analysis to answer the question "Do people that

travel a lot experience a bigger difference in the ambient noise level than people that

do not?" by determining if there is a relation between the mean distance travelled and

the mean noise difference per user.

Quality Assurance: The quality assurance is done according to the pipeline pre-

sented in Chapter 3, but with the according changes resulting from the post-runtime

phase. This adjusted pipeline is shown in Figure 4.5.

Starting with the data processing false data is identified and corrected. For this

the obvious "altidude" error is adjusted by utilizing methods from pandas [87] (see

Listing 4.3). The data contained in sensordata is loaded into a pandas DataFrame

called sensordataValues. All the data contained in "altidude" is copied into "altitude"

and afterwards every data contained in "altidude" is deleted from the DataFrame.

4 Application of the Approach

Figure 4.4: The achievable goals according to Figure 3.1. The green squares and

rounded squares represent the achievable goals, the green circles show the

available sensor data, and the green lines represent the traceable connec-

tions. Data analysis was added as this is no typical main goal, but is

thinkable for this occasion.

Listing 4.3: Fixing of the "altidude" typo.

1#Retrieve all occurrences of "altidude" and copy these

Òævalues from "altidude" into "altitude".

2sensordataValues.loc[sensordataValues["altidude"] >= 0,

Òæ"altitude"] = sensordataValues["altidude"]

4#Drop every occurrence of "altidude" from the DataFrame.

5sensordataValues = sensordataValues.drop(["altidude"],

Òæaxis = 1)

4 Application of the Approach

Figure 4.5: The quality assurance pipeline according to Figure 3.3 adjusted to the

post-runtime phase.

The next data processing step taken is the merging of the different ambient noise

measures. The "amplitude" value denounces the recorded ambient noise as a value

between 0and 1, while "loudness" describes a value between 0and 100.Bothofthese

sensor readings can be interpreted as the percentage of detected volume. To merge

those two, the amplitude value is multiplied by 100 and both values are copied to a

new column called "noiselevel" (see Listing 4.4).

Listing 4.4: Merging "amplitude" and "loudness".

1#A new empty list is generated.

2noiseList = []

4 Application of the Approach

4#Each sensor reading is looped and checked whether it

Òæcontains an "amplitude" value. If such a value is

Òædetected it is multiplied by 100 and added to the

Òælist. If no such value was detected the value

Òæcontained in "loudness" is added to the list.

5for iin range (sensordataValues.shape[0]):

6if (sensordataValues["amplitude"][i] >= 0):

7noiseList.append(sensordataValues["amplitude"][i

Òæ]*100)

8else:

9noiseList.append(sensordataValues["loudness"][i

Òæ])

11 #The sensordataValues DataFrame is expanded by

Òæ"noiselevel" with the values previously contained

Òæin the list.

12 sensordataValues["noiselevel"] = noiseList

After that the data processing is done and because the amount of data available for

this experiment is quite sparse for an MCS-application it is decided against any data

imputation methods in order to avoid a distortion of the data.

At last in the data analysis stage the methods to analyze the data are executed. To

achieve the first goal of map generation alistofallsensingrunscontainingboth,a

GPS and a microphone reading, is needed in order to display the locations where

an ambient noise reading was made on a map. This list is then used to filter the

sensordataValues DataFrame for these sensing runs and this is then further filtered

to only contain the GPS readings in the now called geo DataFrame. Afterwards this

DataFrame is converted to a GeoDataFrame that contains a GeometryArray created

from the "longitude" and "latitude" by using geopandas version 0.11.1[43]. This

GeoDataFrame can now be plotted over a world map. The code for this process can

be seen in Listing 4.5.

Listing 4.5: Generating the map with the locations of the microphone readings.

1#Get a list with all the sensing run id’s that contain

Òæboth, GPS and microphone readings.

4 Application of the Approach

2allSensesWithBoth = set(sensordataValues.loc[

ÒæsensordataValues["name"] == "gps"]["id"].unique())

Òæ&set(sensordataValues.loc[sensordataValues[

Òæ"name"] == "microphone"]["id"].unique())

3#The sensordataValued DataFrame is filtered for sensing

Òæruns contained in the new list.

4combined = sensordataValues.loc[sensordataValues["id"].

Òæisin(allSensesWithBoth)]

5#Further filtered to only contain GPS readings.

6geo = combined.loc[combined["name"] == "gps"]

7geo = geo.reset_index(drop=True)

9#The DataFrame is transformed into a GeoDataFrame with

ÒæGeometryArray created from longitude and latitude.

10 geoGDF = gpd.GeoDataFrame(geo, geometry=gpd.

Òæpoints_from_xy(geo.longitude, geo.latitude))

12 #The GeoDataFrame gets mapped over a world map

13 world = gpd.read_file(gpd.datasets.get_path(

Òæ’naturalearth_lowres’))

14 ax = world.plot(color=’white’, edgecolor=’black’)

15 geoGDF.plot(ax=ax, color=’red’)

17 plt.show()

The data analysis goal tries to find a relation between the mean distance travelled and

the mean noise difference per user. To achieve this a data structure containing all

the values per user is needed, therefore iterating over all users with the corresponding

executed sensing tasks. The DataFrame is then filtered by the "user_id" and by

iterating over all of its values a list with all the users GPS readings in the form of

latitude and longitude tuples or all the users microphone readings is retrieved. By

comparing all the values in this list with each other, a new list is generated containing

all the distances or noise differences. This new list is then used as the value for the

current user in a dictionary. The resulting dictionary is a key-value data structure

containing the users as keys and their distances or noise differences as values. The

4 Application of the Approach

process of generating this dictionary can be seen in Listing 4.6.

Listing 4.6: Generation of the distances dictionary (The noise difference dictionary is

generated respectively).

1#Iterate over all Users with the corresponding sensor

Òæreading.

2for iin userList:

3#Filter the current user and create empty lists.

4userData = location.loc[location["user_id"] == i]

5userData = userData.reset_index(drop=True)

6gpsOfUser = []

7listOfDistancesPerUser = []

9#Iterate over all readings from this user and fill

Òæthe list with latitude and longitude tuples.

10 for jin range (userData.shape[0]):

11 gpsOfUser.append((userData.latitude[j], userData

Òæ.longitude[j]))

12 #print(gpsOfUser)

14 #If this user has more than one reading, compare

Òæeach GPS location with each other, calculate

Òæthe distances of this user and add those to

Òæthe dictionary. Else add an empty list in the

Òædictionary.

15 if(len(gpsOfUser) > 1):

16 for jin range (len(gpsOfUser)):

17 for kin range(j + 1, len(gpsOfUser)):

18 listOfDistancesPerUser.append(geopy.

Òædistance.distance(gpsOfUser[j],

ÒægpsOfUser[k]).km)

19 userDistanceDict[i] = listOfDistancesPerUser

20 else:

21 userDistanceDict[i] = []

After that, the mean is calculated by simply iterating over these dictionaries and calcu-

4 Application of the Approach

lating the mean of all the values for each specific key, this is again saved in a dictionary

called meanDistanceDict or meanNoiseDiffDict. After filtering these dictionaries to

only contain the keys ("user_id") occurring in both dictionaries and converting all

values in a pandas Series, the data can now be analyzed. This final step before the

analysis is shown in Listing 4.7.

Listing 4.7: Calculating the mean and converting to pandas Series.

1#Iterate over the dictionary in order to create a new

Òædictionary containing the mean of each user (only

Òædo this if the value is not an empty list).

2for key, value in userNoiseDict.items():

3if(len(value) > 0):

4meanNoiseDiffDict[key] = np.mean(value)

6#Filtering the dictionaries by intersecting keys.

7keysInBothDicts = list(set(meanNoiseDiffDict.keys()) &

Òæset(meanDistanceDict.keys()))

8meanNoiseDiffDictInstersection = {}

9meanDistanceInstersection = {}

10 for key in keysInBothDicts:

11 meanNoiseDiffDictInstersection[key] =

ÒæmeanNoiseDiffDict[key]

12 meanDistanceInstersection[key] = meanDistanceDict[

Òækey]

14 #Convert to pandas Series.

15 noiseSeries = pd.Series(list(

ÒæmeanNoiseDiffDictInstersection.values()))

16 distanceSeries = pd.Series(list(

ÒæmeanDistanceInstersection.values()))

Finally the data is analyzed (see Listing 4.8) in the form of the covariance, by using

apandasmethod,andthedifferenttypesofcorrelation,byutilizingstatisticalfunc-

tions provided by the stats [112] module of scipy version 1.9.1[110]. In the end the

correlation is plotted by using pyplot [98] from the matplotlib version 3.6.0[73].

4 Application of the Approach

Listing 4.8: Calculating the covariance and correlation. Plotting the correlation

1#Calculate covariance.

2meanDataFrame = pd.concat([noiseSeries, distanceSeries],

Òæaxis = 1)

3print(meanDataFrame.cov())

5#Calculate the three types of correlation with P-value.

6print(scipy.stats.pearsonr(noiseSeries, distanceSeries))

7print(scipy.stats.spearmanr(noiseSeries, distanceSeries)

Òæ)

8print(scipy.stats.kendalltau(noiseSeries, distanceSeries

Òæ))

10 #Plot the correlation

11 plt.title(’Correlation’)

12 plt.scatter(noiseSeries , distanceSeries)

13 plt.plot(np.unique(noiseSeries), np.poly1d(np.polyfit(

ÒænoiseSeries , distanceSeries , 1)) (np.unique(

ÒænoiseSeries)), color=’red’)

14 plt.xlabel(’mean noise diff’)

15 plt.ylabel(’mean distance’)

16 plt.show()

Results: Aquickanalysisofthedatashowsthatofthetotal150 participants there

are 112 that executed both sensor readings at one point. Furthermore there are 2247

sensing runs that contained both, one GPS and at least one microphone reading.

The map generation result of mapping the location of these sensing runs on a world

map can be seen in Figure 4.6. It shows that the absolute majority of microphone

readings was done in Europe with some outliers in different parts of the world.

Atotalof73 participants made at least two GPS and microphone readings while

using the TYS application. This number is relevant because it is the minimum num-

ber required to calculate the mean of distances a user travelled and the mean noise

differences the user experienced. By trying to find a relation between those means

4 Application of the Approach

Figure 4.6: Map generation showing the location of all sensed microphone readings.

the data analysis shows a variance in the mean noise differences of 27.05% and in

the mean of distances travelled of 66651.61 km and a covariance between these two

variables of 32.22%km.Thepositivecovarianceindicatesthatbothvariablesmovein

the same direction. However, the strength of this relation cannot be inferred from this

value. To get this strength, the different correlation coefficients and their P-values

are calculated. All correlation coefficients are in a range between ≠1and +1,with

the sign indicating the direction and the value indicating the strength of a relation.

The P-value indicates the probability of the results being a random outcome of this

dataset, i.e., a lower P-value denounces the trustworthiness of the results. Conven-

tionally, for P<0.05 the correlation coefficient is called statistically significant [29].

The resulting correlation is plotted in Figure 4.7. The Pearson correlation is highly

influenced by outliers and thus works best for normally distributed data [81]. For this

reason, the value of the Pearson correlation in this study is extremely low at 0.024

and the P-value is extremely high at 0.840. The Spearman correlation is appropriate

when the variables are skewed and thus robust for outliers [81]. This results in a better

correlation value of 0.209 and a P-value of 0.076.Finally,theKendallrankcorrelation

describes not the relation between the values of the variables, but the rank of these

values [111]. This also leads to high robustness for outliers and results in a correlation

value of 0.155,butwiththebestP-valueof0.052. All these correlation values indicate

that there is a weak relation between the mean distance travelled and the mean noise

4 Application of the Approach

difference experienced.

Figure 4.7: Plot showing the correlation between the mean distance travelled and the

mean noise difference experienced.

This chapter visualized the application of the approach suggested in Chapter 3. The

streamlined process made the strategy to develop this post-runtime phase MCS-

application simple and helped with the execution.

5 Conclusion

MCS is an emerging topic that tries to improve the applicability and cost efficiency

of conventional Wireless Sensor Networks by utilizing the sensors embedded in the

mobile phones of ordinary citizens. However, most ordinary citizens will not use an

application that notably drains their resources and network capacity and exposes their

personal data while not gaining any benefits from it. For this reason there are many

concerns when creating an MCS-application in order to minimize these problems for

the users.

The first focus of this work was an exhaustive literature research on the topic of

MCS-applications and compared 117 different works in terms of different key aspects.

The first aspect were the goals an application fulfills. Next the sensor utilization,

which sensors were utilized and in which way they were used, was examined. As the

third aspect, methods ensuring the quality of information were regarded. After that,

the time constraints of an application were investigated. The processing device, the

device executing the calculations for the application, was regarded as the next aspect.

After that, an investigation on the used evaluation methods was conducted. What

reporting metrics were used in what cases was regarded as the last aspects to compare

the works by. The knowledge gained through this research was used to create a

streamlined approach for developing MCS-applications. This approach focuses mainly

on what typical MCS goals are and how those goals are achieved, what processing

device the calculations should be executed on, and how to efficiently develop the

application in order to ensure the quality of the sensed data and the results. The

effectiveness of this approach was then displayed by creating a new MCS-application

while following the steps mentioned in it. The application was developed in python

3.9.13 and utilizes data previously sensed by the TrackYourStress app and analyzes

the connection between GPS and microphone readings.

The procedure of designing and implementing this application went smoothly and thus

5 Conclusion

shows the applicability of the approach. Further research on this topic can be executed

to expand this approach to other areas this work did not focus on. A few examples

of these areas would be the incentive mechanism, how are ordinary citizens persuaded

to participate in the application, or task allocation, how are sensing tasks efficiently

allocated to the participants.

A Appendix

Area Papers in that Area

Urban [131, 49, 86, 148, 147, 13, 45, 37, 27, 129, 7, 30, 8, 142, 117, 23,

22, 12, 116, 36, 75, 128, 102, 38, 130, 136, 133, 139, 33, 105,

149, 6, 18, 114, 64, 76, 20, 135, 121, 25, 122, 63, 150, 32, 35,

113, 17, 120, 61, 91, 14]

Indoor [101, 100, 40, 19, 146, 138, 90, 59, 70, 65, 103, 57, 60, 54, 119,

42]

Environment [108, 88, 78, 67, 140, 79, 115, 144, 141, 107, 83, 47, 11, 26, 77,

62, 97, 66, 134]

SSH [95, 48, 28, 84, 132, 15, 2, 68, 41, 82, 4, 44, 16, 145, 10, 21, 53,

137, 24, 80, 106, 9, 104, 31, 50, 143, 123, 34, 56, 3, 71]

Table A.1: This table shows the area of the application described in the papers for

the systematic research. Urban,Indoor,Environment,andSSH are

used as abbreviations for the areas Urban Sensing,Indoor Localization,

Environmental Monitoring,andSocial, Security, and Healthcare.

A Appendix

Urban Indoor Environment SSH

Navigation 4 (8%) 2 (13%) 0 (0%) 1 (3%)

Detect Nearby

BT Devices

0 (0%) 0 (0%) 1 (5%) 3 (10%)

Estimate

Crowd Den-

sity

2 (4%) 0 (0%) 1 (5%) 1 (3%)

Time Estima-

tion

6 (12%) 0 (0%) 0 (0%) 0 (0%)

Table A.2: This table shows other less often occurring goals of the papers by their

categorized area. Urban,Indoor,Environment,andSSH are used as

abbreviations for the areas Urban Sensing,Indoor Localization,En-

vironmental Monitoring,andSocial, Security, and Healthcare.The

number in brackets denounces the percentage of papers in that area achiev-

ing this goal.

A Appendix

[131] [49] [86] [148] [147] [13] [45] [37] [27] [129] [7] [30] [8] [142]

Localization X X X X X X X X X X X X X

Street Observation X X X X X X

Activity Recognition X X X

Image Analysis X

Map Generation X X

Sound Analysis X

Detect Air Pollution

Data Collection

Navigation

Detect Nearby BT Devices

Estimate Crowd Density

Time Estimation X X X

Location Matching X X X X X X X X X X X

Fingerprinting X X

Tracking

Event Detection X X

Real-Time X X X X

No Time Constraint X X X X X X X

Mixed X X X

Local Preprocess X X X X X X X X

Direct Upload X X X X X X

Table A.3: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 1 of 9)

A Appendix

[117] [23] [22] [12] [116] [36] [75] [128] [102] [38] [130] [136] [133]

Localization X X X X X X X X X X X X X

Street Observation X X X X X

Activity Recognition X X X

Image Analysis X X X

Map Generation X X X

Sound Analysis X

Detect Air Pollution

Data Collection

Navigation X X

Detect Nearby BT Devices

Estimate Crowd Density

Time Estimation X X

Location Matching X X X X X X X X X X X X X

Fingerprinting X X X X X

Tracking X

Event Detection X

Real-Time X

No Time Constraint X X X X X X X X

Mixed X X X X

Local Preprocess X X X X X X X

Direct Upload X X X X X X

Table A.4: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 2 of 9)

A Appendix

[139] [33] [105] [149] [6] [18] [114] [64] [76] [20] [135] [121] [25] [122]

Localization X X X X X X X X X X X X X

Street Observation X X X X X X X X

Activity Recognition X X X X X X

Image Analysis X X X

Map Generation X X X X X

Sound Analysis X X

Detect Air Pollution

Data Collection

Navigation X X

Detect Nearby BT Devices

Estimate Crowd Density X X

Time Estimation X

Location Matching X X X X X X X X X X X X

Fingerprinting X X X X

Tracking X

Event Detection X

Real-Time X X X

No Time Constraint X X X X X X

Mixed X X X X X

Local Preprocess X X X X X X X

Direct Upload X X X X X X X

Table A.5: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 3 of 9)

A Appendix

[63] [150] [32] [35] [113] [17] [120] [61] [91] [14] [101] [100] [40] [19]

Localization X X X X X X X X X X X X X X

Street Observation X X X

Activity Recognition X X X X X X

Image Analysis X X X X

Map Generation X X X X X X X

Sound Analysis X

Detect Air Pollution X X

Data Collection X

Navigation

Detect Nearby BT Devices

Estimate Crowd Density

Time Estimation

Location Matching X X X X X X X X X X X

Fingerprinting X X X X

Tracking X X X X X X

Event Detection

Real-Time X X

No Time Constraint X X X X X X X

Mixed X X X X X

Local Preprocess X X X X X X X X

Direct Upload X X X X X X

Table A.6: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 4 of 9)

A Appendix

[146] [138] [90] [59] [70] [65] [103] [57] [60] [54] [119] [42] [108] [88]

Localization X X X X X X X X X X X X

Street Observation

Activity Recognition X X

Image Analysis X X

Map Generation X X X X X

Sound Analysis X

Detect Air Pollution

Data Collection X

Navigation X X

Detect Nearby BT Devices

Estimate Crowd Density

Time Estimation

Location Matching X X

Fingerprinting X X X X X X X X X X X

Tracking X X X X

Event Detection X

Real-Time X

No Time Constraint X X X X X X X X X

Mixed X X X X

Local Preprocess X X X X X X

Direct Upload X X X X X X X X

Table A.7: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 5 of 9)

A Appendix

[78] [67] [140] [79] [115] [144] [141] [107] [83] [47] [11] [26] [77] [62]

Localization X X X X X X X X X X X X X

Street Observation

Activity Recognition

Image Analysis X X X X

Map Generation X

Sound Analysis X

Detect Air Pollution X X X X

Data Collection X X

Navigation

Detect Nearby BT Devices X

Estimate Crowd Density X

Time Estimation

Location Matching X X X X X X X X X X X

Fingerprinting X

Tracking

Event Detection X X X X X

Real-Time X X X

No Time Constraint X X X X X X X X

Mixed X X X

Local Preprocess X X X X X X

Direct Upload X X X X X X X X

Table A.8: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 6 of 9)

A Appendix

[97] [66] [134] [95] [48] [28] [84] [132] [15] [2] [68] [41] [82] [4]

Localization X X X X X X X X X X X X X

Street Observation X

Activity Recognition X X X X

Image Analysis X X

Map Generation X X X

Sound Analysis X

Detect Air Pollution X X X

Data Collection X

Navigation

Detect Nearby BT Devices X

Estimate Crowd Density X

Time Estimation

Location Matching X X X X X X X X X X X X X

Fingerprinting X

Tracking

Event Detection X X

Real-Time X X X X X X

No Time Constraint X X X X X X

Mixed X X

Local Preprocess X X X X X X X X X

Direct Upload X X X X X

Table A.9: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part7 of 9)

A Appendix

[44] [16] [145] [10] [21] [53] [137] [24] [80] [106] [9] [104] [31] [50]

Localization X X X X X X

Street Observation

Activity Recognition X

Image Analysis X X X

Map Generation

Sound Analysis X X

Detect Air Pollution

Data Collection X X X X X

Navigation

Detect Nearby BT Devices X X

Estimate Crowd Density

Time Estimation

Location Matching X X X X X

Fingerprinting X

Tracking

Event Detection X

Real-Time X X

No Time Constraint X X X X X X

Mixed X X X X X X

Local Preprocess X X X X

Direct Upload X X X X X X X X X X

Table A.10: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 8 of 9)

A Appendix

[143] [123] [34] [56] [3] [71]

Localization X X X X X X

Street Observation

Activity Recognition X X

Image Analysis

Map Generation

Sound Analysis X X

Detect Air Pollution

Data Collection X

Navigation X

Detect Nearby BT Devices

Estimate Crowd Density

Time Estimation

Location Matching X X X X X

Fingerprinting

Tracking

Event Detection X X X

Real-Time X X X

No Time Constraint X X X

Mixed

Local Preprocess X X X

Direct Upload X X X

Table A.11: This table (Table A.3 - A.11) shows the goals, localization techniques, time constraint, and processing device of

each paper. (Part 9 of 9)

Bibliography

[1] ACM Digital Library.Accessed:2022-09-14.url:https://dl.acm.org/.

[2] Mahmoud A. Abdo, Ayman A. Abdel-Hamid, and Hesham A. Elzouka. “A

Cloud-based Mobile Healthcare Monitoring Framework with Location Privacy

Preservation”. In: 2020 International Conference on Innovation and Intelligence

for Informatics, Computing and Technologies (3ICT).2020,pp.1–8.doi:10.

1109/3ICT51146.2020.9311999.

[3] Dhruv Agarwal et al. “Understanding Driver-Passenger Interactions in Vehicular

Crowdsensing”. In: Proc. ACM Hum.-Comput. Interact. 5.CSCW2 (2021). doi:

10.1145/3479869.url:https://doi.org/10.1145/3479869.

[4] Adrian Aiordchioae, Daniel Furtun, and Radu-Daniel Vatavu. “Aggregating

Life Tags for Opportunistic Crowdsensing with Mobile and Smartglasses Users”.

In: Proceedings of the 6th EAI International Conference on Smart Objects and

Technologies for Social Good. GoodTechs ’20. Antwerp, Belgium: Association

for Computing Machinery, 2020, 66–71. isbn:9781450375597.doi:10.1145/

3411170.3411237.url:https://doi.org/10.1145/3411170.3411237.

[5] Ian F Akyildiz and Mehmet Can Vuran. Wireless sensor networks. John Wiley

& Sons, 2010. isbn:978-0-470-03601-3.

[6] Heba Aly, Anas Basalamah, and Moustafa Youssef. “Map++: A crowd-sensing

system for automatic map semantics identification”. In: 2014 Eleventh Annual

IEEE International Conference on Sensing, Communication, and Networking

(SECON).2014,pp.546–554.doi:10.1109/SAHCN.2014.6990394.

[7] Theodoros Anagnostopoulos, Theodoros Xanthopoulos, and Yannis Psaromilig-

kos. “A Smartphone Crowdsensing System Enabling Environmental Crowd-

sourcing for Municipality Resource Allocation with LSTM Stochastic Predic-

tion”. In: Sensors (Basel) 20.14 (2020). doi:https://doi.org/10.3390/

s20143966.

Bibliography

[8] Luka Baljak et al. “A system for crowdsensing vibration in public transporta-

tion”. In: 2019 International Conference on Artificial Intelligence: Applications

and Innovations (IC-AIAI).2019,pp.30–304.doi:10.1109/IC-AIAI48757.

2019.00012.

[9] Erin K. Barrett et al. “Mobile Sensing: Leveraging Machine Learning for Effi-

cient Human Behavior Modeling”. In: 2020 Systems and Information Engineer-

ing Design Symposium (SIEDS).2020,pp.1–7.doi:10.1109/SIEDS49339.

2020.9106648.

[10] Felix Beierle et al. “Corona Health-A Study- and Sensor-Based Mobile App

Platform Exploring Aspects of the COVID-19 Pandemic”. In: Int J Environ

Res Public Health 18.14 (July 2021). doi:https://doi.org/10.3390/

ijerph18147395.

[11] Michael Bosello, Giovanni Delnevo, and Silvia Mirri. “On Exploiting Gamifica-

tion for the Crowdsensing of Air Pollution: A Case Study on a Bicycle-Based

System”. In: Proceedings of the 6th EAI International Conference on Smart

Objects and Technologies for Social Good. GoodTechs ’20. Antwerp, Belgium:

Association for Computing Machinery, 2020, 205–210. isbn:9781450375597.

doi:10.1145/3411170.3411256.url:https://doi.org/10.1145/

3411170.3411256.

[12] Raj Bridgelall. “Characterizing Ride Quality With a Composite Roughness In-

dex”. In: IEEE Transactions on Intelligent Transportation Systems (2022),

pp. 1–10. doi:10.1109/TITS.2021.3140177.

[13] Muhammed Fatih Bulut, Murat Demirbas, and Hakan Ferhatosmanoglu. “LineK-

ing: Coffee Shop Wait-Time Monitoring Using Smartphones”. In: IEEE Trans-

actions on Mobile Computing 14.10 (2015), pp. 2045–2058. doi:10.1109/

TMC.2014.2384032.

[14] Chu Cao et al. “Walkway Discovery from Large Scale Crowdsensing”. In: 2018

17th ACM/IEEE International Conference on Information Processing in Sensor

Networks (IPSN).2018,pp.13–24.doi:10.1109/IPSN.2018.00009.

[15] Giuseppe Cardone et al. “Crowdsensing in Urban Areas for City-Scale Mass

Gathering Management: Geofencing and Activity Recognition”. In: IEEE Sen-

sors Journal 14.12 (2014), pp. 4185–4195. doi:10 . 1109 / JSEN . 2014 .

2344023.

Bibliography

[16] Chieh-ming Chang, Szu-Chuang Li, and Yennun Huang. “Building Bluetooth

Beacon-Based Network for Spatial-Temporal Data Collection”. In: Proceedings

of the 2016 International Conference on Communication and Information Sys-

tems. ICCIS ’16. Bangkok, Thailand: Association for Computing Machinery,

2016, 91–95. isbn:9781450347914.doi:10.1145/3023924.3023935.url:

https://doi.org/10.1145/3023924.3023935.

[17] Dongyao Chen and Kang G. Shin. “TurnsMap: Enhancing Driving Safety at

Intersections with Mobile Crowdsensing and Deep Learning”. In: Proc. ACM

Interact. Mob. Wearable Ubiquitous Technol. 3.3 (2019). doi:10 . 1145 /

3351236.url:https://doi.org/10.1145/3351236.

[18] Huihui Chen, Bin Guo, and Zhiwen Yu. “Measures to Improve Outdoor Crowd-

sourcing Photo Collection on Smart Phones”. In: 2019 IEEE SmartWorld, Ubiq-

uitous Intelligence & Computing, Advanced & Trusted Computing, Scalable

Computing & Communications, Cloud & Big Data Computing, Internet of

People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBD-

Com/IOP/SCI).2019,pp.907–915.doi:10.1109/SmartWorld-UIC-ATC-

SCALCOM-IOP-SCI.2019.00183.

[19] Si Chen et al. “Crowd Map: Accurate Reconstruction of Indoor Floor Plans from

Crowdsourced Sensor-Rich Videos”. In: 2015 IEEE 35th International Confer-

ence on Distributed Computing Systems.2015,pp.1–10.doi:10 . 1109 /

ICDCS.2015.9.

[20] Jim Cherian et al. “Poster: ParkGauge: Gauging the Congestion Level of Park-

ing Garages with Crowdsensed Parking Characteristics”. In: Proceedings of the

13th ACM Conference on Embedded Networked Sensor Systems. SenSys ’15.

Seoul, South Korea: Association for Computing Machinery, 2015, 395–396.

isbn:9781450336314.doi:10 . 1145 / 2809695 . 2817881.url:https :

//doi.org/10.1145/2809695.2817881.

[21] Yoonjo Choi et al. “Critical Image Identification via Incident-Type Definition

Using Smartphone Data during an Emergency: A Case Study of the 2020

Heavy Rainfall Event in Korea”. In: Sensors (Basel) 21.10 (2021). doi:https:

//doi.org/10.3390/s21103562.

[22] Yohan Chon, Yunjong Kim, and Hojung Cha. “Autonomous place naming sys-

tem using opportunistic crowdsensing and knowledge from crowdsourcing”. In:

Bibliography

2013 ACM/IEEE International Conference on Information Processing in Sensor

Networks (IPSN).2013,pp.19–30.doi:10.1145/2461381.2461388.

[23] Yohan Chon et al. “Automatically Characterizing Places with Opportunistic

Crowdsensing Using Smartphones”. In: Proceedings of the 2012 ACM Con-

ference on Ubiquitous Computing. UbiComp ’12. Pittsburgh, Pennsylvania:

Association for Computing Machinery, 2012, 481–490. isbn:9781450312240.

doi:10.1145/2370216.2370288.url:https://doi.org/10.1145/

2370216.2370288.

[24] Yohan Chon et al. “Crowdsensing-Based Smartphone Use Guide for Battery Life

Extension”. In: Proceedings of the 2016 ACM International Joint Conference

on Pervasive and Ubiquitous Computing. UbiComp ’16. Heidelberg, Germany:

Association for Computing Machinery, 2016, 958–969. isbn:9781450344616.

doi:10.1145/2971648.2971728.url:https://doi.org/10.1145/

2971648.2971728.

[25] Yohan Chon et al. “Sensing WiFi Packets in the Air: Practicality and Im-

plications in Urban Mobility Monitoring”. In: Proceedings of the 2014 ACM

International Joint Conference on Pervasive and Ubiquitous Computing.Ubi-

Comp ’14. Seattle, Washington: Association for Computing Machinery, 2014,

189–200. isbn:9781450329682.doi:10 . 1145 / 2632048 . 2636066.url:

https://doi.org/10.1145/2632048.2636066.

[26] Andrea Coletta et al. “Optimal Deployment in Crowdsensing for Plant Disease

Diagnosis in Developing Countries”. In: IEEE Internet of Things Journal 9.9

(2022), pp. 6359–6373. doi:10.1109/JIOT.2020.3002332.

[27] Vladimir Coric and Marco Gruteser. “Crowdsensing Maps of On-street Parking

Spaces”. In: 2013 IEEE International Conference on Distributed Computing in

Sensor Systems.2013,pp.115–122.doi:10.1109/DCOSS.2013.15.

[28] Justin Cranshaw et al. “Bridging the Gap between Physical Location and Online

Social Networks”. In: Proceedings of the 12th ACM International Conference on

Ubiquitous Computing. UbiComp ’10. Copenhagen, Denmark: Association for

Computing Machinery, 2010, 119–128. isbn:9781605588438.doi:10.1145/

1864349.1864380.url:https://doi.org/10.1145/1864349.1864380.

[29] Matt David. Fundamentals of analysis.Accessed:2022-10-21.url:https:

//dataschool.com/fundamentals-of-analysis/.

Bibliography

[30] Salvatore Distefano et al. “A Stack4Things-based platform for mobile crowd-

sensing services”. In: 2016 ITU Kaleidoscope: ICTs for a Sustainable World

(ITU WT).2016,pp.1–8.doi:10.1109/ITU-WT.2016.7805722.

[31] Thierry Edoh. “Risk Prevention of Spreading Emerging Infectious Diseases Us-

ing a HybridCrowdsensing Paradigm, Optical Sensors, and Smartphone”. In: J

Med Syst 42.5 (2018), p. 91. doi:https://doi.org/10.1007/s10916-

018-0937-2.

[32] Amr S. El-Wakeel et al. “Towards a Practical Crowdsensing System for Road

Surface Conditions Monitoring”. In: IEEE Internet of Things Journal 5.6 (2018),

pp. 4672–4685. doi:10.1109/JIOT.2018.2807408.

[33] Amr S. El-Wakeel et al. “iDriveSense: Dynamic Route Planning Involving

Roads Quality Information”. In: 2018 IEEE Global Communications Confer-

ence (GLOBECOM).2018,pp.1–6.doi:10.1109/GLOCOM.2018.8648009.

[34] Viktor Erdélyi et al. “Sonoloc: Scalable Positioning of Commodity Mobile De-

vices”. In: Proceedings of the 16th Annual International Conference on Mobile

Systems, Applications, and Services. MobiSys ’18. Munich, Germany: Associ-

ation for Computing Machinery, 2018, 136–149. isbn:9781450357203.doi:

10.1145/3210240.3210324.url:https://doi.org/10.1145/3210240.

3210324.

[35] Xiaochen Fan et al. “Towards System Implementation and Data Analysis for

Crowdsensing Based Outdoor RSS Maps”. In: IEEE Access 6(2018),pp.47535–

47545. doi:10.1109/ACCESS.2018.2867578.

[36] Karoly Farkas et al. “Crowdsending based public transport information service

in smart cities”. In: IEEE Communications Magazine 53.8 (2015), pp. 158–

165. doi:10.1109/MCOM.2015.7180523.

[37] Arsham Farshad, Mahesh K. Marina, and Francisco Garcia. “Urban WiFi char-

acterization via mobile crowdsensing”. In: 2014 IEEE Network Operations and

Management Symposium (NOMS).2014,pp.1–9.doi:10 . 1109 / NOMS .

2014.6838233.

[38] Pawe≥Foremski et al. “Energy-Efficient Crowdsensing of Human Mobility and

Signal Levels in Cellular Networks”. In: Sensors (Basel) 15.9 (2015), pp. 22060–

22088. doi:https://doi.org/10.3390/s150922060.

Bibliography

[39] Raghu K. Ganti, Fan Ye, and Hui Lei. “Mobile crowdsensing: current state and

future challenges”. In: IEEE Communications Magazine 49.11 (2011), pp. 32–

39. doi:10.1109/MCOM.2011.6069707.

[40] Ruipeng Gao et al. “Jigsaw: Indoor Floor Plan Reconstruction via Mobile

Crowdsensing”. In: Proceedings of the 20th Annual International Conference on

Mobile Computing and Networking. MobiCom ’14. Maui, Hawaii, USA: Asso-

ciation for Computing Machinery, 2014, 249–260. isbn:9781450327831.doi:

10.1145/2639108.2639134.url:https://doi.org/10.1145/2639108.

2639134.

[41] Zhigang Gao et al. “A Student Attendance Management Method Based on

Crowdsensing in Classroom Environment”. In: IEEE Access 9(2021),pp.31481–

31492. doi:10.1109/ACCESS.2021.3060256.

[42] Juan A. Álvarez García et al. “Vision and Crowdsensing Technology for an

Optimal Response in Security: Project results”. In: 2021 IEEE International

Mediterranean Conference on Communications and Networking (MeditCom).

2021, pp. 82–87. doi:10.1109/MeditCom49071.2021.9647661.

[43] Geopandas 0.11.0¶.Accessed:2022-10-20.url:https://geopandas.org/

en/stable/.

[44] Avik Ghose, Chirabrata Bhaumik, and Tapas Chakravarty. “BlueEye: A System

for Proximity Detection Using Bluetooth on Mobile Phones”. In: Proceedings

of the 2013 ACM Conference on Pervasive and Ubiquitous Computing Adjunct

Publication. UbiComp ’13 Adjunct. Zurich, Switzerland: Association for Com-

puting Machinery, 2013, 1135–1142. isbn:9781450322157.doi:10.1145/

2494091.2499771.url:https://doi.org/10.1145/2494091.2499771.

[45] Bin Guo et al. “FlierMeet: A Mobile Crowdsensing System for Cross-Space

Public Information Reposting, Tagging, and Sharing”. In: IEEE Transactions

on Mobile Computing 14.10 (2015), pp. 2020–2033. doi:10.1109/TMC.

2014.2385097.

[46] Bin Guo et al. “Mobile Crowd Sensing and Computing: The Review of an

Emerging Human-Powered Sensing Paradigm”. In: ACM Comput. Surv. 48.1

(2015). issn:0360-0300.doi:10.1145/2794400.url:https://doi.org/

10.1145/2794400.

Bibliography

[47] Pengqi Hao et al. “Fine-Grained PM2.5 Detection Method based on Crowd-

sensing”. In: 2020 IEEE International Conference on Consumer Electronics -

Taiwan (ICCE-Taiwan).2020,pp.1–2.doi:10.1109/ICCE-Taiwan49838.

2020.9258279.

[48] Yi He, Ye Li, and Shu-Di Bao. “Fall detection by built-in tri-accelerometer of

smartphone”. In: Proceedings of 2012 IEEE-EMBS International Conference

on Biomedical and Health Informatics.2012,pp.184–187.doi:10.1109/

BHI.2012.6211540.

[49] Shaohan Hu et al. “SmartRoad: Smartphone-Based Crowd Sensing for Traf-

fic Regulator Detection and Identification”. In: ACM Trans. Sen. Netw. 11.4

(2015). issn:1550-4859.doi:10.1145/2770876.url:https://doi.org/

10.1145/2770876.

[50] Xiping Hu et al. “SAfeDJ: A Crowd-Cloud Codesign Approach to Situation-

Aware Music Delivery for Drivers”. In: ACM Trans. Multimedia Comput. Com-

mun. Appl. 12.1s (2015). issn:1551-6857.doi:10.1145/2808201.url:

https://doi.org/10.1145/2808201.

[51] IEEE Xplore.Accessed:2022-09-14.url:https://ieeexplore.ieee.org/

Xplore/home.jsp.

[52] Tinnitus Research Initiative. Track Your Stress.Accessed:2022-10-20.url:

https://www.trackyourstress.org/home.

[53] Katia Jaffrès-Runser et al. “Crowdsensing mobile content and context data:

Lessons learned in the wild”. In: 2017 IEEE International Conference on Perva-

sive Computing and Communications Workshops (PerCom Workshops).2017,

pp. 311–315. doi:10.1109/PERCOMW.2017.7917579.

[54] Ryoma Kawajiri, Masamichi Shimosaka, and Hisashi Kashima. “Steered Crowd-

sensing: Incentive Design towards Quality-Oriented Place-Centric Crowdsens-

ing”. In: Proceedings of the 2014 ACM International Joint Conference on Per-

vasive and Ubiquitous Computing. UbiComp ’14. Seattle, Washington: Asso-

ciation for Computing Machinery, 2014, 691–701. isbn:9781450329682.doi:

10.1145/2632048.2636064.url:https://doi.org/10.1145/2632048.

2636064.

Bibliography

[55] Ferath Kherif and Adeliya Latypova. “Chapter 12 - Principal component anal-

ysis”. In: Machine Learning. Ed. by Andrea Mechelli and Sandra Vieira. Aca-

demic Press, 2020, pp. 209–225. isbn:978-0-12-815739-8.doi:https://

doi.org/10.1016/B978-0-12-815739-8.00012-2.url:https://www.

sciencedirect.com/science/article/pii/B9780128157398000122.

[56] Keunseo Kim et al. “TrailSense: A Crowdsensing System for Detecting Risky

Mountain Trail Segments with Walking Pattern Analysis”. In: Proc. ACM Inter-

act. Mob. Wearable Ubiquitous Technol. 1.3 (2017). doi:10.1145/3131893.

url:https://doi.org/10.1145/3131893.

[57] Yungeun Kim, Yohan Chon, and Hojung Cha. “Mobile Crowdsensing Frame-

work for a Large-Scale Wi-Fi Fingerprinting System”. In: IEEE Pervasive Com-

puting 15.3 (2016), pp. 58–67. doi:10.1109/MPRV.2016.50.

[58] Linghe Kong, Bowen Wang, and Guihai Chen. When Compressive Sensing

Meets Mobile Crowdsensing. Springer, 2019. doi:https://doi.org/10.

1007/978-981-13-7776-1.

[59] Johannes Kässinger et al. “GreenMap: Approximated Filtering Towards Energy-

Aware Crowdsensing for Indoor Mapping”. In: 2018 IEEE 15th International

Conference on Mobile Ad Hoc and Sensor Systems (MASS).2018,pp.451–

459. doi:10.1109/MASS.2018.00069.

[60] Nicholas D. Lane et al. “Piggyback CrowdSensing (PCS): Energy Efficient

Crowdsourcing of Mobile Sensor Data by Exploiting Smartphone App Oppor-

tunities”. In: Proceedings of the 11th ACM Conference on Embedded Net-

worked Sensor Systems. SenSys ’13. Roma, Italy: Association for Computing

Machinery, 2013. isbn:9781450320276.doi:10.1145/2517351.2517372.

url:https://doi.org/10.1145/2517351.2517372.

[61] Mohamed Laraki, Giovanni De Nunzio, and Laurent Thibault. “Vehicle speed

trajectory estimation using road traffic and infrastructure information”. In:

2020 IEEE 23rd International Conference on Intelligent Transportation Sys-

tems (ITSC).2020,pp.1–7.doi:10.1109/ITSC45102.2020.9294337.

[62] Chiara Leonardi et al. “SecondNose: An Air Quality Mobile Crowdsensing Sys-

tem”. In: Proceedings of the 8th Nordic Conference on Human-Computer In-

teraction: Fun, Fast, Foundational. NordiCHI ’14. Helsinki, Finland: Associa-

tion for Computing Machinery, 2014, 1051–1054. isbn:9781450325424.doi:

Bibliography

10.1145/2639189.2670273.url:https://doi.org/10.1145/2639189.

2670273.

[63] Zan Li et al. “SoiCP: A Seamless Outdoor–Indoor Crowdsensing Positioning

System”. In: IEEE Internet of Things Journal 6.5 (2019), pp. 8626–8644. doi:

10.1109/JIOT.2019.2921561.

[64] Emanuel Lima, Ana Aguiar, and Paulo Carvalho. “Offloading Surrogates Char-

acterization via Mobile Crowdsensing”. In: Proceedings of the First ACM Work-

shop on Mobile Crowdsensing Systems and Applications. CrowdSenSys ’17.

Delft, Netherlands: Association for Computing Machinery, 2017, 7–12. isbn:

9781450355551. doi:10.1145/3139243.3139253.url:https://doi.

org/10.1145/3139243.3139253.

[65] Chang Liu, Chun Wang, and Juan Luo. “Large-Scale Deep Learning Framework

on FPGA for Fingerprint-Based Indoor Localization”. In: IEEE Access 8(2020),

pp. 65609–65617. doi:10.1109/ACCESS.2020.2985162.

[66] Liang Liu et al. “Third-Eye: A Mobilephone-Enabled Crowdsensing System for

Air Quality Monitoring”. In: Proc. ACM Interact. Mob. Wearable Ubiquitous

Technol. 2.1 (2018). doi:10.1145/3191752.url:https://doi.org/10.

1145/3191752.

[67] Tong Liu et al. “ALC2: When Active Learning Meets Compressive Crowdsens-

ing for Urban Air Pollution Monitoring”. In: IEEE Internet of Things Journal

6.6 (2019), pp. 9427–9438. doi:10.1109/JIOT.2019.2939552.

[68] Jacob Livingston and Robert Steele. “A crowdsensing algorithm for imputing

Zika outbreak location data”. In: 2017 IEEE 8th Annual Ubiquitous Comput-

ing, Electronics and Mobile Communication Conference (UEMCON).2017,

pp. 334–340. doi:10.1109/UEMCON.2017.8249065.

[69] Huadong Ma, Dong Zhao, and Peiyan Yuan. “Opportunities in mobile crowd

sensing”. In: IEEE Communications Magazine 52.8 (2014), pp. 29–35. doi:

10.1109/MCOM.2014.6871666.

[70] Sumudu Hasala Marakkalage et al. “Identifying Indoor Points of Interest via

Mobile Crowdsensing: An Experimental Study”. In: 2019 IEEE VTS Asia Pacific

Wireless Communications Symposium (APWCS).2019,pp.1–5.doi:10 .

1109/VTS-APWCS.2019.8851651.

Bibliography

[71] Sumudu Hasala Marakkalage et al. “Understanding the Lifestyle of Older Pop-

ulation: Mobile Crowdsensing Approach”. In: IEEE Transactions on Compu-

tational Social Systems 6.1 (2019), pp. 82–95. doi:10.1109/TCSS.2018.

2883691.

[72] Mohammad Masoud et al. “Sensors of Smart Devices in Internet of Everything

(IoE) Era: Big Opportunities and Massive Doubts”. In: Journal of Sensors 2019

(May 2019). doi:10.1155/2019/6514520.

[73] Matplotlib - Visualization with python.Accessed:2022-10-20.url:https:

//matplotlib.org/.

[74] Jules G. McNeff. “The global positioning system”. In: IEEE Transactions on

Microwave Theory and Techniques 50.3 (2002), pp. 645–652. doi:10.1109/

22.989949.

[75] Jernej Mihelj et al. “Crowdsourced Traffic Event Detection and Source Repu-

tation Assessment Using Smart Contracts”. In: Sensors (Basel) 19.15 (2019).

doi:https://doi.org/10.3390/s19153267.

[76] Silvia Mirri, Catia Prandi, and Paola Salomoni. “Personalizing Pedestrian Ac-

cessible way-finding with mPASS”. In: 2016 13th IEEE Annual Consumer Com-

munications & Networking Conference (CCNC).2016,pp.1119–1124.doi:

10.1109/CCNC.2016.7444946.

[77] Jayantrao Mohite et al. “RuPS: Rural participatory sensing with rewarding

mechanisms for crop monitoring”. In: 2015 IEEE International Conference on

Pervasive Computing and Communication Workshops (PerCom Workshops).

2015, pp. 378–383. doi:10.1109/PERCOMW.2015.7134067.

[78] Shigeya Morishita et al. “SakuraSensor: Quasi-Realtime Cherry-Lined Roads

Detection through Participatory Video Sensing by Cars”. In: Proceedings of

the 2015 ACM International Joint Conference on Pervasive and Ubiquitous

Computing. UbiComp ’15. Osaka, Japan: Association for Computing Machin-

ery, 2015, 695–705. isbn:9781450335744.doi:10.1145/2750858.2804273.

url:https://doi.org/10.1145/2750858.2804273.

[79] Davide Moroni et al. “A mobile crowdsensing app for improved maritime secu-

rity and awareness”. In: 2022 IEEE International Conference on Pervasive Com-

puting and Communications Workshops and other Affiliated Events (PerCom

Bibliography

Workshops).2022,pp.103–105.doi:10.1109/PerComWorkshops53856.

2022.9767516.

[80] Moad Mowafi, Fahed Awad, and Fida’a Al-Quran. “Distributed Visual Crowd-

sensing Framework for Area Coverage in Resource Constrained Environments”.

In: Sensors (Basel) 22.15 (2022). doi:https : / / doi . org / 10 . 3390 /

s22155467.

[81] M. M. Mukaka. “Statistics corner: A guide to appropriate use of correlation

coefficient in medical research”. eng. In: Malawi medical journal : the journal

of Medical Association of Malawi 24.3 (2012). 23638278[pmid], pp. 69–71.

issn:1995-7270.url:https://pubmed.ncbi.nlm.nih.gov/23638278.

[82] Xiaoguang Niu et al. “A hierarchical-learning-based crowdedness estimation

mechanism for crowdsensing buses”. In: 2017 IEEE 36th International Perfor-

mance Computing and Communications Conference (IPCCC).2017,pp.1–8.

doi:10.1109/PCCC.2017.8280471.

[83] Klimis S. Ntalianis, Andreas Kener, and Jahna Otterbacher. “Feelings’ Rating

and Detection of Similar Locations, Based on Volunteered Crowdsensing and

Crowdsourcing”. In: IEEE Access 7(2019),pp.90215–90229.doi:10.1109/

ACCESS.2019.2926812.

[84] Robin Wentao Ouyang et al. “If You See Something, Swipe towards It: Crowd-

sourced Event Localization Using Smartphones”. In: Proceedings of the 2013

ACM International Joint Conference on Pervasive and Ubiquitous Comput-

ing. UbiComp ’13. Zurich, Switzerland: Association for Computing Machinery,

2013, 23–32. isbn:9781450317702.doi:10.1145/2493432.2493455.url:

https://doi.org/10.1145/2493432.2493455.

[85] PRISMA Statement.Accessed:2022-09-14.url:https://prisma-statement.

org/.

[86] Bei Pan et al. “Crowd Sensing of Traffic Anomalies Based on Human Mobility

and Social Media”. In: Proceedings of the 21st ACM SIGSPATIAL International

Conference on Advances in Geographic Information Systems. SIGSPATIAL’13.

Orlando, Florida: Association for Computing Machinery, 2013, 344–353. isbn:

9781450325219. doi:10.1145/2525314.2525343.url:https://doi.

org/10.1145/2525314.2525343.

[87] Pandas.Accessed:2022-10-20.url:https://pandas.pydata.org/.

Bibliography

[88] Victor Pankratius et al. “Mobile crowd sensing in space weather monitoring:

the mahali project”. In: IEEE Communications Magazine 52.8 (2014), pp. 22–

28. doi:10.1109/MCOM.2014.6871665.

[89] Harshita Patel et al. “A review on classification of imbalanced data for wireless

sensor networks”. In: International Journal of Distributed Sensor Networks 16.4

(2020), p. 1550147720916404. doi:10.1177/1550147720916404.eprint:

https://doi.org/10.1177/1550147720916404.url:https://doi.org/

10.1177/1550147720916404.

[90] Lambros Petrou et al. “Demonstration Abstract: Crowdsourced Indoor Local-

ization and Navigation with Anyplace”. In: Proceedings of the 13th Interna-

tional Symposium on Information Processing in Sensor Networks. IPSN ’14.

Berlin, Germany: IEEE Press, 2014, 331–332. isbn:9781479931460.

[91] Catia Prandi et al. “Walking with Geo-Zombie: A pervasive game to engage

people in urban crowdsourcing”. In: 2016 International Conference on Comput-

ing, Networking and Communications (ICNC).2016,pp.1–5.doi:10.1109/

ICCNC.2016.7440545.

[92] Project jupyter.Accessed:2022-10-20.url:https://jupyter.org/.

[93] Rüdiger Pryss et al. “Exploring the Time Trend of Stress Levels While Using

the Crowdsensing Mobile Health Platform, TrackYourStress, and the Influence

of Perceived Stress Reactivity: Ecological Momentary Assessment Pilot Study”.

In: JMIR Mhealth Uhealth 7.10 (2019), e13978. issn:2291-5222.doi:10.

2196/13978.url:http://mhealth.jmir.org/2019/10/e13978/.

[94] Rüdiger Pryss et al. “Machine Learning Findings on Geospatial Data of Users

from the TrackYourStress mHealth Crowdsensing Platform”. In: 2019 IEEE

20th International Conference on Information Reuse and Integration for Data

Science (IRI).2019,pp.350–355.doi:10.1109/IRI.2019.00061.

[95] Rüdiger Pryss et al. “Mobile Crowd Sensing in Clinical and Psychological Trials

– A Case Study”. In: 2015 IEEE 28th International Symposium on Computer-

Based Medical Systems.2015,pp.23–24.doi:10.1109/CBMS.2015.26.

[96] National Center for Biotechnology Information PubMed. PubMed.Accessed:

2022-09-14. url:https://pubmed.ncbi.nlm.nih.gov/.

Bibliography

[97] Zuber Purahoo and Sudha Cheerkoot-Jalim. “SenseAPP: An IoT-Based Mo-

bile Crowdsensing Application for Smart Cities”. In: 2020 3rd International

Conference on Emerging Trends in Electrical, Electronic and Communications

Engineering (ELECOM).2020,pp.47–52.doi:10 . 1109 / ELECOM49001 .

2020.9297018.

[98] Pyplot tutorial - matplotlib 3.6.0 documentation.Accessed:2022-10-20.url:

https://matplotlib.org/stable/tutorials/introductory/pyplot.

html.

[99] Python. Python release python 3.9.13.Accessed:2022-10-21.url:https:

//www.python.org/downloads/release/python-3913/.

[100] Valentin Radu, Lito Kriara, and Mahesh K. Marina. “Pazl: A mobile crowd-

sensing based indoor WiFi monitoring system”. In: Proceedings of the 9th

International Conference on Network and Service Management (CNSM 2013).

2013, pp. 75–83. doi:10.1109/CNSM.2013.6727812.

[101] Anshul Rai et al. “Zee: Zero-Effort Crowdsourcing for Indoor Localization”. In:

Proceedings of the 18th Annual International Conference on Mobile Computing

and Networking. Mobicom ’12. Istanbul, Turkey: Association for Computing

Machinery, 2012, 293–304. isbn:9781450311595.doi:10.1145/2348543.

2348580.url:https://doi.org/10.1145/2348543.2348580.

[102] Tirathraj Ramburn, Deevash Badoreea, and Sudha Cheerkoot-Jalim. “DriveMU:

A Real-time Road-Traffic Monitoring Android Application for Mauritius”. In:

2019 Conference on Next Generation Computing Applications (NextComp).

2019, pp. 1–8. doi:10.1109/NEXTCOMP.2019.8883623.

[103] Patrice Raveneau, Stéphane D’Alu, and Hervé Rivano. “Localisation based on

Wi-Fi fingerprints: A crowdsensing approach with a device-to-device aim”. In:

2017 IEEE International Conference on Pervasive Computing and Communi-

cations Workshops (PerCom Workshops).2017,pp.321–325.doi:10.1109/

PERCOMW.2017.7917581.

[104] Vassili Rivron et al. “Refining smartphone usage analysis by combining crowd-

sensing and survey”. In: 2015 IEEE International Conference on Pervasive Com-

puting and Communication Workshops (PerCom Workshops).2015,pp.366–

371. doi:10.1109/PERCOMW.2015.7134065.

Bibliography

[105] João G. P. Rodrigues, João P. Pereira, and Ana Aguiar. “Impact of Crowd-

sourced Data Quality on Travel Pattern Estimation”. In: Proceedings of the

First ACM Workshop on Mobile Crowdsensing Systems and Applications.Crowd-

SenSys ’17. Delft, Netherlands: Association for Computing Machinery, 2017,

38–43. isbn:9781450355551.doi:10.1145/3139243.3139254.url:https:

//doi.org/10.1145/3139243.3139254.

[106] Robert Rusek, Joaquim Melendez Frigola, and Joan Colomer Llinas. “Influence

of occupant presence patterns on energy consumption and its relation to com-

fort: a case study based on sensor and crowd-sensed data”. In: Energy Sustain

Soc 12.1 (2022), p. 13. doi:https://doi.org/10.1186/s13705-022-

00336-6.

[107] Darshan Santani et al. “DrinkSense: Characterizing Youth Drinking Behav-

ior Using Smartphones”. In: IEEE Transactions on Mobile Computing 17.10

(2018), pp. 2279–2292. doi:10.1109/TMC.2018.2797901.

[108] Darshan Santani et al. “The Night is Young: Urban Crowdsourcing of Nightlife

Patterns”. In: Proceedings of the 2016 ACM International Joint Conference

on Pervasive and Ubiquitous Computing. UbiComp ’16. Heidelberg, Germany:

Association for Computing Machinery, 2016, 427–438. isbn:9781450344616.

doi:10.1145/2971648.2971713.url:https://doi.org/10.1145/

2971648.2971713.

[109] Joy Lal Sarkar et al. “A Novel Approach for Real-Time Data Management in

Wireless Sensor Networks”. In: Proceedings of 3rd International Conference

on Advanced Computing, Networking and Informatics.Ed.byAtulyaNagar,

Durga Prasad Mohapatra, and Nabendu Chaki. New Delhi: Springer India,

2016, pp. 599–607. isbn:978-81-322-2529-4.

[110] SciPy. SciPy.Accessed:2022-10-20.url:https://scipy.org/.

[111] SciPy. Scipy.stats.kendalltau - SciPy v1.9.1 Manual.Accessed:2022-10-22.

url:https : / / docs . scipy . org / doc / scipy - 1 . 9 . 1 / reference /

generated/scipy.stats.kendalltau.html.

[112] SciPy. Statistical Functions (scipy.stats) - SciPy v1.9.1 Manual.Accessed:

2022-10-20. url:https : / / docs . scipy . org / doc / scipy - 1 . 9 . 1 /

reference/stats.html.

Bibliography

[113] Lu Shao et al. “Traffic condition estimation using vehicular crowdsensing data”.

In: 2015 IEEE 34th International Performance Computing and Communications

Conference (IPCCC).2015,pp.1–8.doi:10.1109/PCCC.2015.7410307.

[114] Effrosyni Sigala, Efthimios Alepis, and Constantinos Patsakis. “Measuring the

Quality of Street Surfaces in Smart Cities through Smartphone Crowdsensing”.

In: 2020 11th International Conference on Information, Intelligence, Systems

and Applications (IISA.2020,pp.1–8.doi:10.1109/IISA50023.2020.

9284384.

[115] Phillip Smith et al. “A Novel Software Defined Radio for Practical, Mobile

Crowd-sourced Spectrum Sensing”. In: IEEE Transactions on Mobile Comput-

ing (2021), pp. 1–1. doi:10.1109/TMC.2021.3107409.

[116] Junji Takahashi et al. “Clustering for Road Damage Locations Obtained by

Smartphone Accelerometers”. In: Proceedings of the Second International Con-

ference on IoT in Urban Space. Urb-IoT ’16. Tokyo, Japan: Association for

Computing Machinery, 2016, 89–91. isbn:9781450342049.doi:10.1145/

2962735.2962741.url:https://doi.org/10.1145/2962735.2962741.

[117] Manoop Talasila, Reza Curtmola, and Cristian Borcea. “Alien vs. Mobile user

game: Fast and efficient area coverage in crowdsensing”. In: 6th International

Conference on Mobile Computing, Applications and Services.2014,pp.65–74.

doi:10.4108/icst.mobicase.2014.257779.

[118] Gehad Mohamed Tawfik et al. “A step by step guide for conducting a system-

atic review and meta-analysis with simulation data”. In: Trop Med Health 47

(2019), p. 46. doi:https://doi.org/10.1186/s41182-019-0165-6.

[119] Xiaohua Tian et al. “Toward a Quality-Aware Online Pricing Mechanism for

Crowdsensed Wireless Fingerprints”. In: IEEE Transactions on Vehicular Tech-

nology 67.7 (2018), pp. 5953–5964. doi:10.1109/TVT.2018.2805383.

[120] Yao Tong et al. “Vehicle Inertial Tracking via Mobile Crowdsensing: Experience

and Enhancement”. In: IEEE Transactions on Instrumentation and Measure-

ment 71 (2022), pp. 1–13. doi:10.1109/TIM.2022.3156993.

[121] Gianni Tumedei et al. “Promoting a Safe Return to University Campuses dur-

ing the COVID-19 Pandemic: Crowdsensing Room Occupancy”. In: Proceed-

ings of the Conference on Information Technology for Social Good.GoodIT

’21. Roma, Italy: Association for Computing Machinery, 2021, 145–150. isbn:

Bibliography

9781450384780. doi:10.1145/3462203.3475911.url:https://doi.

org/10.1145/3462203.3475911.

[122] Rohit Verma et al. “Smart-Phone Based Spatio-Temporal Sensing for Anno-

tated Transit Map Generation”. In: Proceedings of the 25th ACM SIGSPA-

TIAL International Conference on Advances in Geographic Information Sys-

tems. SIGSPATIAL ’17. Redondo Beach, CA, USA: Association for Computing

Machinery, 2017. isbn:9781450354905.doi:10.1145/3139958.3140005.

url:https://doi.org/10.1145/3139958.3140005.

[123] Aku Visuri et al. “Smartphone Detection of Collapsed Buildings during Earth-

quakes”. In: Proceedings of the 2017 ACM International Joint Conference on

Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM In-

ternational Symposium on Wearable Computers. UbiComp ’17. Maui, Hawaii:

Association for Computing Machinery, 2017, 557–562. isbn:9781450351904.

doi:10.1145/3123024.3124402.url:https://doi.org/10.1145/

3123024.3124402.

[124] W3C. Accelerometer.Accessed:2022-10-16.2022.url:https://www.w3.

org/TR/accelerometer/.

[125] W3C. Ambient light sensor.Accessed:2022-10-21.2022.url:https://www.

w3.org/TR/ambient-light/.

[126] W3C. Gyroscope.Accessed:2022-10-16.2021.url:https://www.w3.org/

TR/gyroscope/.

[127] W3C. Magnetometer.Accessed:2022-10-21.2021.url:https://www.w3.

org/TR/magnetometer/.

[128] Hai Wang et al. “CSMC: Cellular Signal Map Construction via Mobile Crowd-

sensing”. In: Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 5.4

(2022). doi:10 . 1145 / 3494959.url:https : / / doi . org / 10 . 1145 /

3494959.

[129] Xiaojie Wang et al. “A City-Wide Real-Time Traffic Management System: En-

abling Crowdsensing in Social Internet of Vehicles”. In: IEEE Communications

Magazine 56.9 (2018), pp. 19–25. doi:10.1109/MCOM.2018.1701065.

[130] Xiaoyan Wang et al. “Green Spectrum Sharing Framework in B5G Era by

Exploiting Crowdsensing”. In: IEEE Transactions on Green Communications

and Networking (2022), pp. 1–1. doi:10.1109/TGCN.2022.3186282.

Bibliography

[131] Yin Wang et al. “CrowdAtlas: Self-Updating Maps for Cloud and Personal

Use”. In: Proceeding of the 11th Annual International Conference on Mobile

Systems, Applications, and Services. MobiSys ’13. Taipei, Taiwan: Association

for Computing Machinery, 2013, 27–40. isbn:9781450316729.doi:10.1145/

2462456.2464441.url:https://doi.org/10.1145/2462456.2464441.

[132] Jens Weppner and Paul Lukowicz. “Bluetooth based collaborative crowd den-

sity estimation with mobile phones”. In: 2013 IEEE International Conference

on Pervasive Computing and Communications (PerCom).2013,pp.193–200.

doi:10.1109/PerCom.2013.6526732.

[133] Di Wu et al. “Human as a Service: Towards Resilient Parking Search System

With Sensorless Sensing”. In: IEEE Transactions on Intelligent Transportation

Systems 23.8 (2022), pp. 13863–13877. doi:10.1109/TITS.2021.3133713.

[134] Di Wu et al. “When Sharing Economy Meets IoT: Towards Fine-Grained Urban

Air Quality Monitoring through Mobile Crowdsensing on Bike-Share System”.

In: Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 4.2 (2020). doi:

10.1145/3397328.url:https://doi.org/10.1145/3397328.

[135] Xuangou Wu et al. “Privacy preserving RSS map generation for a crowdsensing

network”. In: IEEE Wireless Communications 22.4 (2015), pp. 42–48. doi:

10.1109/MWC.2015.7224726.

[136] Xuan Xiao et al. “How Many Bumps in Your City? Personalized Bump Seeker

With Mobile Crowdsensing”. In: IEEE Transactions on Instrumentation and

Measurement 71 (2022), pp. 1–12. doi:10.1109/TIM.2021.3135549.

[137] Chenren Xu et al. “Crowdsensing the speaker count in the wild: implications

and applications”. In: IEEE Communications Magazine 52.10 (2014), pp. 92–

99. doi:10.1109/MCOM.2014.6917408.

[138] Qiang Xu and Rong Zheng. “MobiBee: A Mobile Treasure Hunt Game for

Location-Dependent Fingerprint Collection”. In: Proceedings of the 2016 ACM

International Joint Conference on Pervasive and Ubiquitous Computing: Ad-

junct. UbiComp ’16. Heidelberg, Germany: Association for Computing Ma-

chinery, 2016, 1472–1477. isbn:9781450344623.doi:10.1145/2968219.

2968590.url:https://doi.org/10.1145/2968219.2968590.

Bibliography

[139] Kuldeep Yadav et al. “Human sensors: Case-study of open-ended community

sensing in developing regions”. In: 2013 IEEE International Conference on Per-

vasive Computing and Communications Workshops (PERCOM Workshops).

2013, pp. 389–392. doi:10.1109/PerComW.2013.6529523.

[140] Chaoqun Yang et al. “A Crowdsensing-Based Cyber-Physical System for Drone

Surveillance Using Random Finite Set Theory”. In: ACM Trans. Cyber-Phys.

Syst. 3.4 (2019). issn: 2378-962X. doi:10.1145/3342049.url:https:

//doi.org/10.1145/3342049.

[141] Guang Yang et al. “CEDAR: A Cost-Effective Crowdsensing System for De-

tecting and Localizing Drones”. In: IEEE Transactions on Mobile Computing

19.9 (2020), pp. 2028–2043. doi:10.1109/TMC.2019.2921962.

[142] Takuro Yonezawa et al. “Accelerating Urban Science by Crowdsensing with Civil

Officers”. In: Proceedings of the 2018 ACM International Joint Conference

and 2018 International Symposium on Pervasive and Ubiquitous Computing

and Wearable Computers. UbiComp ’18. Singapore, Singapore: Association for

Computing Machinery, 2018, 303–306. isbn:9781450359665.doi:10.1145/

3267305.3267641.url:https://doi.org/10.1145/3267305.3267641.

[143] Ana María Zambrano V et al. “SIGPRO: A Real-Time Progressive Notifica-

tion System Using MQTT Bridges and Topic Hierarchy for Rapid Location

of Missing Persons”. In: IEEE Access 8(2020),pp.149190–149198.doi:

10.1109/ACCESS.2020.3015183.

[144] W. Zamora et al. “Accurate Ambient Noise Assessment Using Smartphones”.

In: Sensors (Basel) 17.4 (2017).

[145] Mattia Zeni, Enrico Bignotti, and Fausto Giunchiglia. “Combining Crowdsourc-

ing and Crowdsensing to Infer the Spatial Context”. In: 2018 IEEE International

Conference on Pervasive Computing and Communications Workshops (PerCom

Workshops).2018,pp.28–33.doi:10.1109/PERCOMW.2018.8480312.

[146] Chi Zhang et al. “GROPING: Geomagnetism and cROwdsensing Powered In-

door NaviGation”. In: IEEE Transactions on Mobile Computing 14.2 (2015),

pp. 387–400. doi:10.1109/TMC.2014.2319824.

Bibliography

[147] Fuzheng Zhang et al. “Sensing the Pulse of Urban Refueling Behavior”. In:

Proceedings of the 2013 ACM International Joint Conference on Pervasive

and Ubiquitous Computing. UbiComp ’13. Zurich, Switzerland: Association for

Computing Machinery, 2013, 13–22. isbn:9781450317702.doi:10.1145/

2493432.2493448.url:https://doi.org/10.1145/2493432.2493448.

[148] Pengfei Zhou, Yuanqing Zheng, and Mo Li. “How Long to Wait? Predicting

Bus Arrival Time With Mobile Phone Based Participatory Sensing”. In: IEEE

Transactions on Mobile Computing 13.6 (2014), pp. 1228–1241. doi:10.

1109/TMC.2013.136.

[149] Hanwei Zhu and Sid Chi-Kin Chau. “Integrating IoT-Sensing and Crowdsensing

for Privacy-Preserving Parking Monitoring”. In: Proceedings of the 8th ACM

International Conference on Systems for Energy-Efficient Buildings, Cities, and

Transportation. BuildSys ’21. Coimbra, Portugal: Association for Computing

Machinery, 2021, 226–227. isbn:9781450391146.doi:10.1145/3486611.

3492229.url:https://doi.org/10.1145/3486611.3492229.

[150] Qiuxi Zhu et al. “Spatiotemporal Scheduling for Crowd Augmented Urban

Sensing”. In: IEEE INFOCOM 2018 - IEEE Conference on Computer Commu-

nications.2018,pp.1997–2005.doi:10.1109/INFOCOM.2018.8485869.