Existing approaches to smart parking: An overview

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/318100909 Existing Approaches to Smart Parking: An Overview Conference Paper · May 2017 DOI: 10.1007/978-3-319-59513-9_7 CITATIONS 8 READS 289 5 authors, including: Some of the authors of this publication are also working on these related projects: Automatic diatom identification for water quality assessment and monitoring View project Face and facial elements detection View project Fernando Enriquez Universidad de Sevilla 43 PUBLICATIONS297 CITATIONS SEE PROFILE Luis Miguel Soria Morillo Universidad de Sevilla 32 PUBLICATIONS382 CITATIONS SEE PROFILE Juan Antonio Alvarez-Garcia Universidad de Sevilla 75 PUBLICATIONS630 CITATIONS SEE PROFILE Francisco Velasco-Morente Universidad de Sevilla 129 PUBLICATIONS794 CITATIONS SEE PROFILE All content following this page was uploaded by Juan Antonio Alvarez-Garcia on 27 April 2021. The user has requested enhancement of the downloaded file. Existing approaches to smart parking: An overview Fernando Enr´ ıquez1, Luis Miguel Soria1, Juan Antonio ´ Alvarez-Garc´ ıa1, Francisco Velasco2, and Oscar D´ eniz3 1Computer Languages and Systems Department, University of Seville, 41012 Seville, Spain [email protected],[email protected],[email protected] 2Applied Economics I Department, University of Seville, 41018 Seville, Spain [email protected] 3VISILAB, E.T.S.I.I, University of Castilla-La Mancha, Ciudad Real, Spain [email protected] Abstract. After years of technological advances, parking is still a problem for many people, as a time-consuming task, they have to face on a day-by-day basis, and also for cities, that see how traffic and pollution increases. There have been multiple attempts to find a partial or global technological solution to this problem, ranging from using different types of sensors or cams for automatically detecting free spaces to collaborative apps that let users share related information. In this paper, we give an overview of the methods that have been developed so far, showing their main features, differences, pros, and cons, as well as other factors that may contribute to the success or failure of new proposals that will come in the future. Keywords: Smart city, Parking, Crowdsensing, Computer vision 1 Introduction According to the United Nations [1], 54% of the population lives in urban zones and this is expected to grow to 66% in 2050. This tendency makes urban mobility more difficult and makes finding a parking space one of the most repetitive problems which citizens in big cities have to face up. Furthermore, vehicles cruising for parking are responsible for at least 30% of traffic jams [2], with the average time for search being more than 20 minutes [3]. Besides the personal problems that this can generate, this is a problem in terms of fuel consumption, CO2 emissions and in general a waste of resources for the community. Beyond private spaces, there is no worldwide accepted solution for monitoring vehicles that enter or leave an area, though several efforts have been made to solve the problem in a specific context. Smart parking is the term used for a set of technologies and applications targeting issues related to parking in Smart Cities and this work aims at providing an up-to-date survey of the most interesting and relevant solutions. The rest of the paper is organized as follows. Section 2 analyses the state of the art in Smart Parking solutions. The main conclusions are drawn in Section 3. 2 2 Classification Parking solutions are offered to off-street or on-street. Off-street parking may refer to multistory car parks, while on-street refers to parking spaces along public roads and streets. This work considers both types of scenarios. Typically there is a focus on infrastructure, where several sensors are installed to monitor the places, crowdsensing where mobile phones’ drivers are the source of the information or a hybrid proposal. Here, vision solutions are analyzed as a specific category due to the increase of cameras and the potential of this type of sensor. 2.1 Infraestructure In order to get the occupancy status of parking places, fixed or mobile sensors are installed in on/off-street parking to detect vehicular events. Mobile sensors are not the most common solution but Parknet [4] in San Francisco is a most prominent study: taxi cars collected the occupancy status of the parking place when they passed beside it, gathering data from GPS receiver and ultrasonic sensors. Although every taxi can detect multiple spots, updating the information of the same spot can take 25 minutes with a fleet of 300 vehicles. Parking Spotter is based on the same idea, leveraging sonar and radar sensors of some Ford vehicles. There are also some LiDAR-based solutions [5] but only focused on surveying parking spaces with one equipped car. These works are hardly scalable since all the vehicles must have the same kind of sensor or else share the information through the same database and the number of mounted sensor-vehicles must be enough to update the information frequently. Fixed sensors are the more extended and popular option. Although there are a wide variety of sensors [6] (active infrared, ultrasound, acoustic, accelerometer, etc.), magnetometer is by far the most common fixed sensor. The magnetometer is accurate although normally solutions require at least one sensor per place, increasing the cost of the deployment. The sensor itself measures the current magnetic fields and detects the arrival of metal vehicles. Most municipal deployment projects or large shopping centers drill magnetometers on parking places, sharing this information through visual signals or mobile applications. Installation and maintenance processes involve access to the property and road surface so it is only possible for off-street or on-street with government permission. Table 1 refers to the most prominent smart city parking solutions. SmartSantander and San Francisco can be seen as pilot studies. SmartSantander was conceived as a Smart City Laboratory and San Francisco finished its pilot study at the end of 2013 due to the cost of sensor maintenance. Nice withdrew mobile application Nice Passport in 2016 after some organizational4and security problems5. M´ alaga, London (the only one based on RFID), Moscow or Los Angeles are success cases of Smart Parking. 4http://www.20minutes.fr/nice/1839579-20160504-nice-trois-ans-apres-installation- stationnement-intelligent-disparait 5http://www.lefigaro.fr/politique/le-scan/couacs/2014/06/06/25005-20140606 ARTFIG00112- securite-informatique-a-nice-pris-de-court-estrosi-interrompt-une-interview.php 3 Fixed sensor-based solutions research is focused on reducing the installation time and cost [7] using surface-mounted magnetometers that can be glued to the road and enlarging battery life of wireless sensors. Smart City #places Company URL Year San Francisco 68.2K Fybrtech http://www.fybr-tech.com 2011 Santander [8] 0.4K Libelium http://bit.ly/2mOd38r 2011 Nice 74.5K Urbiotica http://www.urbiotica.com 2012 Los Angeles [9] 6.3K StreetLine http://lat.ms/1BVDxpD 2012 London 83.4K Smart Parking http://bit.ly/2mTSTcm 2012 Moscow 950k WorldSensing http://bit.ly/2lCMAbU 2012 Malaga 10 2.2K Parkhelp http://bit.ly/2mTPNFc 2014 Table 1. Infrastructure parking projects 2.2 Vision Due to the increasing interest of the scientific community and users in general by the techniques of artificial intelligence based on images, in recent years there have proliferated solutions to the problem of parking control through systems based on vision. These systems, unlike previous infrastructures, are not yet established, but are mostly used in controlled and experimental environments. Although there are companies that base their systems on these techniques, most of the work is still under development. In general, the techniques of detection of parking spaces based on vision, suffer from several problems. The first one is the quality of the image. For certain works based on object recognition, the image must be of sufficient quality to be processed. During the day, this may not be an obstacle, but in conditions of insufficient lighting or adverse weather conditions, it becomes a real problem. Another problem related to obtaining the image are occlusions. Depending on the location of the image capture system, the vehicles themselves or surrounding elements (trees, buildings, street furniture, shadows) may obstruct the view of the parking area. This problem can be solved by changing the placement of the capture system, but this may not always be possible, or reduce the monitored surface due to the change in perspective. Finally, an inherent problem with this detection technique is the classification itself. Image-based classification systems have proliferated over the past decade, but they are still far from offering the assurance of systems based on structural elements. 6http://sfpark.org/resources/parking-sensor-technology-performance-evaluation 7http://www.urbiotica.com/en/inauguration-of-our-smart-parking-project-in-the-city-of-nice- fr/ 8https://www.westminster.gov.uk/parkright 9http://parking.mos.ru/en 10 http://www.eesc.europa.eu/?i=portal.en.events-and-activities-smart-cities-malaga 4 However, in favor of systems for detecting parking spaces using vision algorithms, there is the coverage, cost, and versatility of the systems. Since a single capture system can cover dozens (or hundreds) of parking spaces, the costs reduction compared to infrastructure-based systems is significant. In addition, the maintenance cost of these systems is negligible, except in those cases where drones or satellites are required to acquire the images. The latter are usually multi-functional, which leads to another strength: versatility. Because the images obtained can be applied not only to the parking control, systems implementing the vision-based solution are often used for other purposes at the same time. Examples of this functionality are surveillance systems, pedestrian control, maintenance tasks, and other scenarios that make the investment made in the implementation of the image capture systems to be quickly amortized. The parking lots detection solutions based on vision can use different systems to get the images of the zone under control. Among other, the most common methods for the image acquisition are external video cameras, vehicles equipped with vision systems, three-dimensional capture systems, and zenithal or aerial images obtained by satellites or drones. In the same way, as discussed above, detection systems are supported by a set of algorithms. These algorithms can be grouped by the treatment performed over the images. Most works in the state of the art base their development on some of the following approaches: –Appearance based approaches. Based on the comparison of the current appearance of a parking place with an original appearance of the vacant state. Many techniques are tailored and fine-tuned to specific contexts and scenarios. However, these techniques can not be easily generalized, and even the adaptation of one solution to a different parking lot is not straightforward. –Recognition based approaches. Approaches based on object recognition aim to detect and classify the vehicles occupying the parking space using machine learning algorithms. Complex approaches because of the large variety of the target objects. –Three dimensional image processing. –Combined techniques applying image processing in order to improve the quality and avoid light variation effects, and machine learning algorithms to classify image content. Finally, monitoring techniques can be divided into two types depending on the way in which lots are processed: –Estimating occupancy of an entire parking lot, for example, by counting incoming vehicles. –Checking for the presence of a vehicle in each cell. Most vision-based approaches require the presence of vehicles in individual parking lots. Below, a compilation of some relevant papers from the related bibliography are shown. Among them, there are different approaches as a sample of the heterogeneity of the existing processing, recognition, and image acquisition systems. In [10] is proposed a distributed and efficient system to solve the problem of parking through vision systems. For this purpose convolutional neural networks specifically 5 designed for smart cameras are used. Two visual datasets have been used to verify the accuracy of the proposed system: PKLot and CNRPark-EXT. This last dataset was created by the own authors. Thanks to convolutional neural networks applied in the classification process, the proposed solution is robust to images exposed to partial occlusion, shadows, and changes in light conditions. In addition, it has a good capacity of generalization. A reduced version of the AlexNet neural network were used in this paper. The new convolutional neural network is able to recognize only two classes: free or occupied parking space. Authors in [11] implements a parking detection system based on real-time image processing from video cameras. It divides the system into three sections: image acquisition module, image pre-processing module and image detection module. In the first one, the image is filtered. The detection module is based on the use of a reference image from an empty parking space, without any interference. From this image, converted to grayscale, a comparison system is established with successive images. The algorithm obtains the edges of the reference image and compares them with the last captured. Finally, a function is applied to decide whether the reference image and the compared one have similar characteristics. The accuracy of the system is fair (81%). The solution proposed in [12] uses individual images from a single surveillance camera previously installed in indoor parking lots. This work is also based on reference images from the empty parking lot. The process is slightly different from [11]. The model is generated using Principal Component Analysis, supporting lighting invariance to avoid lighting changes were recognized as potential objects. In addition to the system adaptation to lighting changes, textures are also used to detect objects and isolate them from the background. This makes the system much more robust under extreme lighting changes (large trees, street lamps or intense shadows). In areas with high visibility, the number of occlusions between vehicles is minimized and, hence, systems use to be more accurate. This work has 90% of accuracy, although it can decrease if occlusions occur in the image. In [13], the classification is done by a 3-layer Bayesian hierarchical detection Framework (BHDF). To avoid occlusion problems with other cars and surrounding objects, this paper uses the full image for processing, rather than determining the state of the parking spaces one by one. To analyze the whole parking space, the scene is divided into 3D cubes. Each cube corresponds to a row of car lots. Once the image is obtained and the row of parking spaces to be processed is defined, the BHDF framework is applied. This framework is composed of three pre-trained models, one for the local classification (observation layer), other with the adjacency model (labeling layer) and, the last, with the semantic layer. The local classification model (input model) can be pixel-based or texture-based. The work presented in [14] deals with the problem of the sunlight, the dark shadows during the day, and the low light intensity at nighttime. To do this, ParkLotD uses a classifier based on the fuzzy c-means clustering algorithm (FCM) and a hyper-parametric fit with particle swarm optimization (PSO). The algorithm has been tested during different hours of the day and under different weather conditions. However, the nighttime use with poor lighting conditions is not advised. At the beginning of the process, it is necessary to define the limits of each parking space, as well as the total number of lots. 6 In the next stage, PCA is applied from the feature vectors obtained from each image. Thanks to a user interface, the operator can correct the algorithm operation and show if any error occurs. Furthermore, the operator can retrain the system when needed. The evaluation process was carried out from 2000 images captured by a camera on a roof. Each image covered 27 parking spaces. In [15], parking states are determined by the combination of an adaptive background subtraction algorithm for moving objects detection (overcoming problems of light changes and shadow effects), with speeded up robust features (SURF) algorithm (robust to scale and rotation changes). Authors propose a solution based on the image appearance and the comparison with a reference image where the lots appear vacant. By using a homogenous transformation to change the point of view, facilitates the extraction of the parking model eliminating perspective distortion. For this process, the user must previously select at least four points that define the parking area. Once the transformation is completed, the parking lots are defined. For the adaptive background subtraction, the Mixture Gaussian Model algorithm is used. This requires a learning process for updating the distributions obtained by the model through a set of images. Once the background is extracted, the characteristics of the images to be classified are obtained. For this process, the algorithms of Speeded Up Robust Features Speeded (invariant to scale changes) and Histogram of Oriented Gradients (HOG) are proposed. The algorithms SVM or KNN can be combined during the classification stage. The authors, through their evaluation with the VI-RAT video database, consider that the best combination of proposed solutions is SURF +SVM, obtaining an average precision of 93%. However, the main problem of this method is the lack of robustness to partial occlusions. In [16], the aim is not focused on parking areas recognition, but on detecting cars and distances by means of vehicles equipped with a stereo vision system. The authors identify three methods to perform this detection: using radar, using active sensors (laser, radar, lidar) and monocular vision, and finally, using only vision. 2D-based systems may not be enough accurate in identification of the vehicles. To solve this drawback, this system process the 3D image identifying certain potentially characteristics belonging to a vehicle (vertical edges). The algorithm is divided into two main stages: the extraction of three-dimensional characteristics and the detection of the vehicle from them. Extraction of vertical characteristics allows isolating more precisely the obstacles (and vehicles) of the own highway. Two cameras are used for generates 3-D sparse map. As aforementioned, although this kind of algorithm is not specific for parking lots recognition, they can be applied in this context thanks to the capacities to recognize vehicles. Finally, table 2 shows a schematic representation of some of the most relevant works in parking slots recognition. In table are collected various parameters representative of the different solutions proposed. These parameters include hardware required, algorithm error rate, robustness to lighting changes (low , medium W, high ), automation level of the recognition and learning process (manual identification of parking lots for training process , semiautomatic system allowing manual changes in the training process W, fully-automated system ), robustness to perspective changes (low , medium W, high ), robustness to partial occlusions (low , medium W, high ), preferred environment for the system use (indoor, outdoor, both), deployment costs (low , medium 7 if cameras are already installed W, complex infrastructure could be required [satellites, drones, ultrasonic sensors] ), computational requirements (low [could be high during the training stage] , mediumW, high ), scalability (low , mediumW, high ) and, finally, the way in which the car parks are processed (individually using masks or collectively over the whole area). Author Year Error rate Hardware required Lighting Learning Perspective Occlusions Coverage Economic Computational Scalability Grouping G. Amato et al. [10] 2017 3% Smart-camera  W   Both  W  Masks G. Amato et al. [17] 2016 9% Smart-camera +RPi  W   Both  W  Masks I. Masmoudi et al. [15] 2016 7% Videocamera W W W  Outdoor W W W Masks J. Suhr et al. [18] 2014 1.90% AVM and ultrasound     Both    Vehicle J. Jermsurawong et al. [19] 2012 1% Videocamera  W   Outdoor W W  Full area H. Ichihashi et al. [14] 2009 7% Videocamera +server W W   Both W W  Masks Z. Bin et al. [11] 2009 19% Video camera +CPU     Both W  W Masks C. Huang et al. [13] 2008 6% IP camera +CPU W W   Indoor W W W Rows G. Toulminet et al. [16] 2006 N/D Stereo-vision system +CPU     Outdoor W W  Vehicle Q. Zhang et al. [20] 2006 20% Satellite images     Outdoor    Full area S. Funck et al. [12] 2004 10% CCTV-cameras W W   Indoor W  W Masks K. Yamada et al. [21] 2001 0.60% Videocamera W W   Outdoor W  W Masks X. Wang et al. [22] 1998 25% Aerial images  W   Outdoor    Full area Table 2. Vision-based parking lot detection algorithms 2.3 Social crowdsensing Using software applications to help drivers park their vehicles is not new, even with solutions that include the location, reservation, access and payment tasks all together as in [23]. Nevertheless, the growing popularity of smartphones in the last years, full of sensors and capable of registering geo-positional information in short periods of time, has created new ways of searching for a parking spot favoring the rise of multiple mobile and web applications. Table 3 shows a comparison between some of the apps for on-street or off-streets parking we can find on the Internet. The selection criteria used to filter the huge catalog that exists nowadays is based on the popularity, but also on the special features offered by some of them. Most of the apps are centered uniquely on off-street parking, maybe collecting data from installed sensors or cams. Parkopedia (available worldwide) or wesmartPark (working in Madrid and Barcelona), are different examples of these off-street parking solutions. The latter offers the installation of sensor technology and a management module for owners to easily monitor their business while they obtain data for their app users. While the basic operations in these apps remain the same, we find interesting proposals in the literature that could change the scene someday, i.e. bringing automatic price negotiation capabilities [24]. There are also apps that give information of on-street parking, recommending the best zones to search for unoccupied parking spots, or giving the opportunity to announce when a user is about to move a vehicle, leaving a free space, and in some cases, allowing reservations for this 8 new free spots. Each of these functionalities can use data obtained automatically by cell phones (crowdsensing) or introduced manually by the app users (crowdsourcing). This approach, automatic or manual, is particularly useful for on-street parking, where the scalability of other approaches is more difficult due to the size of the areas to be covered. While off-street parking information can primarily help users reduce the price to pay or select the closest space to their destination, on-street parking information usually has as its main goal the mitigation of the ”multiple-car-chasing-single-space” phenomenon [25] derived from the common ”blind search” strategy. On-street Off-street URL Coverage Mobile stats pay reserve location prices pay reserve www.parkme.com 4200 cities,worldwide 3 3 3 3 3 3 en.parkopedia.com 6308 cities, worldwide 3 3 3 3 3 parknav.com 70 cities, USA & Germany 3 3 www.waze.com N/A3 3 maps.google.com 25 cities, USA 3 3 3 www.wazypark.com big cities, Spain 3 3 3 www.aparcandgo.com airports/stations 2 cities, Spain 3 3 3 3 www.parkapp.com 16 cities, Spain 3 3 3 3 www.wesmartpark.com 2 cities, Spain 3 3 3 3 3 parclick.es 170 cities, Spain & EU 3 3 3 3 www.telpark.com +60 cities, Spain & Portugal 3 3 3 3 3 www.e-park.es 13 cities, Spain 3 3 www.parkwhiz.com +200 cities, USA 3 3 3 3 3 www.streetline.com N/A3 3 3 3 3 3 www.bestparking.com +100 cities & airports USA & Canada 3 3 3 3 3 www.spotoops.com worldwide 3 3 www.justpark.com +1000 cities, UK 3 3 3 3 3 parktag.mobi 7 cities, EU 3 3 *11 www.peertopark.com nationwide, Spain 3 3 3 3 www.aparcalia.com 3 cities, Spain +airports & stations 3 3 3 3 www.carpling.com nationwide, 13 countries 3 www.appyparking.com 11 cities, UK 3 3 3 3 3 3 Table 3. Smart parking apps When the information is retrieved in an automatic way, it is usually visualized as a statistical estimation, showing the street maps with different colors indicating whether it is easy, medium or difficult to find parking spots. This coarse grain classification tags can help drivers heading towards zones where it is more likely to be successful but it obviously does not guarantee an available parking space. This analysis is based on the processing of data (i.e. cell phone signals) in a massive way, generating mathematical models that represent global data, just like it is done to show the traffic density during different hours of the day. ParkMe (owned by INRIX), Parknav, Streetline and Google Maps, are examples of parking solutions that employ machine learning algorithms to process automatically retrieved parking data. Streetline is heavily focused on data analytics and offers different solutions, being ”Parker” the one that shows real time parking 11 Planned