A Fleet Learning Architecture for Enhanced Behavior Predictions during Challenging External Conditions [original]

A Fleet Learning Architecture for Enhanced Behavior Predictions

during Challenging External Conditions

Florian Wirthm¨

uller , Marvin Klimke , Julian Schlechtriemen , Jochen Hipp and Manfred Reichert

Abstract— Already today, driver assistance systems help to

make daily traffic more comfortable and safer. However, there

are still situations that are quite rare but are hard to handle

at the same time. In order to cope with these situations and

to bridge the gap towards fully automated driving, it becomes

necessary to not only collect enormous amounts of data but

rather the right ones. This data can be used to develop and

validate the systems through machine learning and simulation

pipelines. Along this line this paper presents a fleet learning-

based architecture that enables continuous improvements of

systems predicting the movement of surrounding traffic par-

ticipants. Moreover, the presented architecture is applied to a

testing vehicle in order to prove the fundamental feasibility

of the system. Finally, it is shown that the system collects

meaningful data which are helpful to improve the underlying

prediction systems.

I. INTRODUCTION

Driver assistance systems are on the rise and help to

prevent accidents and to support drivers in various ways

more and more frequently. Thereby, modules predicting

future motions of surrounding traffic participants constitute

a central piece of such system’s intelligence. As shown in

[1], it is beneficial to integrate external information such

as knowledge about weather or traffic conditions into these

prediction modules. Therefore, the systems are enabled to

deal with rarely occuring and nevertheless challenging con-

ditions, resulting in increased system performances as well

as benefits for the drivers in general and especially during

challenging conditions. As a prerequisite for developing such

context-aware motion prediction modules huge amounts of

data need to be collected. But it is not only about gathering

the pure amount of data but rather about collecting the right

F. Wirthm¨

uller, M. Klimke, J. Schlechtriemen and J. Hipp are with

Mercedes-Benz AG, B¨

oblingen, Germany,

E-Mail: {first name.last name}@daimler.com

F. Wirthm¨

uller and M. Reichert are with the Institute of Databases and

Information Systems (DBIS), Ulm University, Ulm, Germany,

E-Mail: {first name.last name}@uni-ulm.de

J. Schlechtriemen is with the Institute of Realtime Learning Systems at

the University of Siegen, Siegen, Germany

ORCID (ordered as authors above):

https://orcid.org/0000-0002-9732-2561;

https://orcid.org/0000-0003-2647-9673;

https://orcid.org/0000-0002-9130-061X;

https://orcid.org/0000-0002-9037-9899;

https://orcid.org/0000-0003-2536-4153

IEEE must be obtained for all other uses, in any current or future media,

including reprinting/republishing this material for advertising or promotional

purposes, creating new collective works, for resale or redistribution to

servers or lists, or reuse of any copyrighted component of this work in

other works.

Fig. 1. Current software modules for predicting the movement of surround-

ing traffic participants are developed through an unidirectional process with

a single data collection campaign and model training. This is shown in the

upper part of the illustration. In this context a motion prediction module can

be thought of as any machine learning model, beeing illustrated as simple

neural network. The lower part of the image shows our idea of a continuous

fleet data collection in an illustrative way.

data. This means to collect data during conditions where

current systems face problems and data which facilitates

developers to improve their systems. As at least some of

these conditions occur rather rarely, it is crucial to collect

corresponding data with a large fleet of vehicles to ensure

a good coverage of all kinds of situations. Hence, this

work presents a data collection architecture enabling contin-

uously improving motion predictions over time. Besides, we

demonstrate the fundamental feasibility of the approach by

integrating it into a testing vehicle. Fig. 1 illustrates the idea

of such a fleet learning architecture indicating its potential

for system improvements.

The remainder of this work is structured as follows: Sec. II

gives an overview of related works. Sec. III introduces

the new fleet learning-based architecture concept. The con-

cept enables the detection of challenging conditions with

respect to behavior prediction. In particular, it allows re-

parametrizing onboard prediction modules in order to achieve

improved prediction performances. Afterwards, Sec. IV and

Sec. V describe the development of the desired prediction

watchdog and the prototypical realization of the needed

onboard components in a testing vehicle. Sec. VI summarizes

and concludes the article.

II. RELATED WORK

For the present study, works dealing with the development

of systems intended for data collection and management

(Sec. II-A) as well as such dealing with behavior prediction

(Sec. II-B) are of particular interest. After introducing and

categorizing characteristic approaches, Sec. II-C discusses

them and deduces the contribution of this article.

A. Data Collection and Management Systems

Works intending to collect, manage and preprocess data

to be used during the development, test and simulation of

algorithms for automated driving applications in general can

be divided into two main groups. The approaches of the

first group focus on data collection from an external point

of view. Therefore, camera systems or statically positioned

drones looking from a birds-eye view onto the scene are used.

Whereas the NGSIM data set [2] has been very popular,

most researchers recently switched over to the more exact

and larger highD and inD data sets [3], [4].

By contrast, the second group relies on dedicated mea-

surement vehicles [5]–[7]. Therefore, the data collection is

performed from a moving point of view within the scene.

Thus, challenges as occluded vehicles can occur and the

data collection mechanism itself can influence the collected

behaviors. In exchange, measurement durations for single ve-

hicles can be significantly increased compared to approaches

in the first group.

In addition to the data sets mentioned above, which

represent traffic scenes and their objects through numerical

descriptions of the traffic scenes and their objects, several

optical data sets exist. As examples the popular KITTI

[8] and Cityscapes [9] data sets or the upcoming Baidu

driving [10] and Audi autonomous driving [11] data sets

can be mentioned. While such optical data sets are more

suitable, for example, for the development of object detection

and semantic segmentation algorithms, numerical ones are

significantly more useful for object motion predictions.

Additionally, this work focuses on the development of a

fleet learning architecture. Consequently, it is desirable to

minimize the size of the data to be transmitted. Accord-

ingly, optical data sets are only of secondary interest for

the presented article, as images or even videos are much

larger compared to sparse numerical object representations.

This also applies to works focusing on the collection of

pedestrian motion data, such as the popular UCY, ETH, and

stanford drone data sets [12]–[14], as our research focuses

the prediction of vehicle motions.

B. Behavior Prediction Approaches

According to Lef`

evre [15], behavior prediction ap-

proaches can be divided into the three categories: physics-

based, maneuver-based, and interaction-aware prediction ap-

proaches. Physics-based approaches assume that future ve-

hicle motions solely depend on the laws of physics and can

be described with simple models such as constant velocity

or constant acceleration. A good overview on corresponding

approaches is provided in [16]. By contrast, maneuver-based

approaches (e. g. [1], [17]–[22]) try to infer the maneu-

ver a driver intends to perform. Finally, interaction-aware

approaches [23]–[26] provide the most advanced motion

models by predicting the motions of all vehicles in a given

situation simultaneously. In particular, these models consider

that all vehicles mutually influence each other.

[18] uses a categorization, which is more oriented towards

the representation of the prediction output and also allows

categorizing approaches that cannot be uniquely assigned

to one of the aforementioned classes. This categorization

distinguishes between approaches for maneuver prediction,

position prediction, and hybrid approaches. While maneuver

prediction approaches (e. g. [20], [24]) try to infer which

one of a fixed set of maneuvers a vehicle will perform,

position prediction approaches (e. g. [21], [23], [25]–[28]) try

to infer at which exact position a vehicle will be at a certain

future point in time, i. e. the latter approaches operate in a

continuous space. Finally, hybrid approaches (e. g. [1], [17]–

[19], [22]) integrate the outputs of maneuver and position

prediction approaches into a single or combined model.

C. Contribution

As the literature overview has revealed, a lot of research

has been spent on data collection for automated driving as

well as on motion prediction. However, the presented works

presume a setting, where a data set is collected once and

afterwards utilized to train and validate prediction models.

This procedure, obviously limits the variance as well as the

size of the data set. In general, it is not possible to collect a

data set covering all relevant corner cases through a single

data collection campaign.

As an exception, [29] uses growing hidden markov models

to learn trajectory prediction models for pedestrians and

vehicles at a fixed location. Essentially, the drivable space

is represented as a discretized graph with edges and nodes,

which is updated over runtime. Although this work is in line

with our research direction, it cannot be integrated into a

moving vehicle and a therefore changing surrounding.

In order to bridge the described research gap, this article

contributes in three respects:

C.1 A fleet learning-based architecture enabling enhanced

behavior predictions, especially during challenging ex-

ternal conditions, is presented.

C.2 As a key part of the described architecture, a prediction

watchdog is developed.

C.3 The necessary modules of the architecture are pro-

totypically implemented and integrated into a testing

vehicle, demonstrating the fundamental feasibility of the

approach.

III. ARCHITECTURE CONCEPT

We aim to develop an architecture concept enabling

continuous performance improvements to motion prediction

systems within a fleet of vehicles. The architecture needs to

fulfill the following requirements:

R.1 The prediction performance shall be equal over all

vehicles of the fleet in any situation.

R.2 Data transmission (e. g. via mobile communication) if

necessary shall be restricted to a minimum in order to

reduce communication costs.

Sensor Fusion

Communication

Module

Communication

Module

Trajectory

Planning

Actuators

Situation

Prediction

Comparison

Prediction Buffer Parameter

Storage

Parameter

Update Module

Situation

Database

{PE, PP, SI}

{PET0, PPT0, SIT0}

In-Vehicle Component (N) Backend Component (1)

PET0, PPT0, SIT0

SI: Sensor Information

PE: Predicted Environment

PT: Planned Trajectory

PP: Prediction Parameters

TR: Trigger

XT0: X Buffered

TD: Training Data

PU: Parameter Update

Communication

Channel

{PET0, PPT0, SIT0, SI}

AD Stack

Pred. Watchdog

Fig. 2. Overview of the proposed fleet learning architecture. The colored connections highlight the two main loops within the architecture:

blue: condition-adaptive parameter request loop; green: data collection and parameter update loop. As indicated by the 1-to-N relation, there are a lot of

vehicles with the described in-vehicle component communicating with a single backend component.

R.3 The prediction module shall produce reliable results

even if the system is offline (i. .e. if there is no mobile

communication signal).

R.4 The overall prediction performance shall increase over

lifetime.

R.5 All updates to be deployed to the vehicle fleet need to

go through a release process.

The developed architecture meeting these requirements

is depicted in Fig. 2. From a high-level perspective, the

architecture comprises of a communication channel as well

as an in-vehicle and a backend component. The in-vehicle

component, in turn, comprises seven modules:

•A sensor fusion module aggregating the raw informa-

tion from different sensors and providing a consistent

representation of the surrounding to other modules.

•A situation prediction module providing the trajectory

planning module with information about the evolution

of the current traffic situation. The module’s output can

be optimized through remote parametrization.

•A trajectory planning module planning trajectories for

the ego-vehicle based on the current sensor information

as well as the situation predictions. Good trajectories are

characterized by safety and comfort for the passengers.

To enable the planning module to generate such trajec-

tories, both inputs need to be as accurate as possible in

any situation.

•Actuators realizing planned trajectories.

•A prediction buffer storing the current prediction out-

put, the prediction parameters, and the current sensor

information until reaching the prediction time.

•A comparison module comparing the actual positions

of the surrounding vehicles with the predictions made

some moments ago. If a predicted position differs too

much from the actual one, this module triggers the

communication module to send a new package of train-

ing data containing the buffered prediction (output), the

buffered sensor information (input), and the actual po-

sition (desired output) to the backend. This mechanism

ensures that exactly those situations are detected and

used to increase the prediction performance, which are

currently handled sub-optimally during the predictions.

This contributes to meet requirement R.2.

•A communication module requesting condition-specific

prediction parameters from the backend and providing

them to the prediction module. This communication

module also transmits data backwards over the com-

munication channel.

The backend part on the other hand consists of only four

modules:

•A communication module receiving data from the ve-

hicle fleet and transmitting prediction parameters back-

wards.

•A condition-adaptive parameter storage, holding the

parameters currently used during different situations.

Due to the use of a single shared parameter storage

for all vehicles of the fleet, requirement R.1 is met.

•A situation database storing all data that are necessary to

(re-)train a machine learning-based prediction module.

In detail:

–All inputs necessary for the prediction module.

–The desired output of the prediction module.

–The external conditions of the measurement (e. g.

weather or speedlimit).

•A parameter update module using the measurements

stored in the situation database to calculate improved

parameter values. Before pushing an update to the

parameter storage, it is checked whether it increases the

performance in all known situations. Then the updated

parameters are released resulting in the fulfillment of

requirements R.4 and R.5.

As further shown in Fig. 2, essentially there are two

data loops. The loop emphasized in green collects data that

shall enable determining improved prediction parameters in

the backend. Within the blue-colored loop, vehicles request

condition-adaptive prediction parameters and use them to

ensure reliable predictions during all situations. To ensure

that the prediction also works in scenarios in which no com-

munication with the backend is possible, the system may fall

back to a basic parameter set (fulfilling requirement R.3). The

latter is initially used as well. To bridge short offline-phases,

it is advisable that the vehicles request parameters already

in advance if possible. This becomes possible, for example,

if the route ahead is known or if it is foreseeable that it will

start to rain soon. From the viewpoint of functional safety, it

might also be advantageous to rely on a fixed neural network

architecture and to solely adjust the weights during parameter

updates. This though is also transferable to other prediction

techniques.

Communication

Module

Backend

Communication

Module

Vehicle Communication Channel

Request Condition-

Adaptive Prediction

Parameters

Respond with

Prediction Parameters

New Trainingsdata

Fig. 3. Overview of the communication channel and the three transmitted

message types.

Fig. 3 shows the communication channel and the in-

vehicle- and backend-sided communication modules in more

detail. Basically, there are three types of messages to be

transmitted:

•New measurements collected by a vehicle that need to

be added to the situation database.

•A request for a parameter set fitting the external condi-

tions a vehicle is faced with.

•A reply to a parameter request.

IV. PREDICTION WATCHDOG

The prediction watchdog as depicted in Fig. 2 consists

of the prediction buffer and the comparison module. As

introduced briefly, it is able to memorize predicted positions

as well as the input features that led to the respective

model output. Moreover, it compares the actual prediction

outputs with the desired one, and triggers the transmission

of additional data to the backend. Sec. IV-A provides further

details on the concept of buffering predictions, whereas

Fig. 4. Illustration of the memory module update using ego-relative

coordinates. The striped green vehicle depicts the predicted position of the

green vehicle. The vehicle shown in blue is the ego-vehicle.

Sec. IV-B outlines the working principle of the comparison

module and the triggering.

A. Prediction Buffer

The prediction buffer’s function is to hold positions

[ˆxth,ˆyth]Tpredicted at the current point in time t=t0until

reaching the prediction horizon hat t=t0+th. In case of an

ideal prediction for any given point in time, the memorized

point lies on the continuously updated trajectory until the

vehicle arrives at that position.

Due a lacking world-fixed coordinate frame, the predicted

position has to be fixed to the current environment by

changing its vehicle-relative coordinates. The prediction is

memorized in lane or also called Frenet coordinates [30],

which enable a robust and reasonably precise way of updat-

ing the numeric values using Euler integration. The velocity

of the ego-vehicle is measured and split into longitudinal and

lateral components given the current driving lane, denoted by

~v = [vx, vy]T. For each memory entry, there is a countdown

variable tcthat is initialized with the prediction horizon th

when saving a new prediction. The memorized prediction is

updated according to Eq. 1 and Eq. 2:

ˆxth

ˆyth←ˆxth

ˆyth−vx

vy·∆t(1)

tc←tc−∆t(2)

∆tcorresponds to the time passed since the last model

update. Fig. 4 illustrates the process of updating the model.

Simply put, the predicted position declared relatively to

the ego-vehicle performing the prediction is adjusted with the

Fig. 5. Image showing the prediction and logging approach integrated into

a testing vehicle and connected to an AR visualization.

ego-vehicle’s movement in each step. As soon as tcreaches

zero, the memorized prediction is not updated anymore and

the comparison with the current position can be carried

out. Due to the limited model update frequency, the exact

moment when the countdown vanishes cannot be captured.

A significantly large movement of the vehicle can be caused

until the assessment is issued, as the longitudinal velocity

can be reasonably high. To deal with that effect a constant-

velocity correction in longitudinal direction is performed.

This prevents the slack due to finite update frequency to have

an effect on the accuracy evaluation. A perfect prediction

would feature a residual position of zero in the exact moment

the countdown reaches zero.

B. Comparison

The working principle of the comparison module is rather

simple. In order to send the appropriate data to enhance the

prediction modules in the backend, it is advisable to select

those predictions that are too far away from the desired

output, i. e. the actual position [xth, yth]T. Therefore, the

comparison module calculates the longitudinal ex,thand

lateral prediction errors ey,thaccording to Eq. 3 and Eq. 4:

ex,th=|xth−ˆxth|(3)

ey,th=|yth−ˆyth|(4)

A a new data package is sent to the backend when

exceeding one of the given thresholds Θxand Θyfor the

two directions.

V. APPLICATION IN A TESTING VEHICLE

In order to demonstrate the fundamental feasibility and

benefits of the presented approach, we implemented the

described modules in a testing vehicle. To do so, we relied

on the findings we published in [18] with regard to the

prediction component as well as on the prediction watchdog

presented in Sec. IV. Initially, we are restricting ourselves

to the prediction of the behavior of the ego-vehicle, as this

setting is easier and faster to realize. In general, the same

mechanisms can be transfered to surrounding vehicles as

well. Our investigations focus on the fundamental feasibility

of the data collection loop as well as the quality of the

collected data in order to enable model improvements. The

design of the communication channel and training adapted

prediction models are out of scope of this work. Instead,

the data, which would be transferred from the vehicle to

the backend in the final application is simply logged in

CSV files. This allows for a downstreamed inspection by

examining e. g. the histograms of the residuals.

The testing vehicle is equipped with a series-like sensor

setup consisting of automotive radars and cameras facing

the front and the back of the vehicle. Moreover, the testing

vehicle is equipped with an additional computing unit, where

a ROS environment [31] is deployed. Within the ROS envi-

ronment, the sensor signals published over the Flexray bus

are accessible, allowing for the easy implementation of new

functional blocks. The ROS environment is connected to an

integrated augmented reality display that allows visualizing

the prediction outputs in real time. Fig. 5 shows an example

of the AR visualization of the predictions. A video showing

the output for a short sequence can be found on Reserach-

Gate1. The visualization solution allows for additional visual

inspection of the system performance.

In our experiments, the denoted system frequency is 25 Hz

and the threshold for triggering an erroneous lateral predic-

tion logging is set to Θy= 0.2m2. Therefore, the logged

data have to show up significantly higher prediction errors

than the system shows during normal operation. According

to [18], the median lateral error should be around 0.11 m

for a prediction horizon of 3 s. These values cannot be

reached in the given experiment, as the prediction models

could probably be subject to transfer errors. This is because

a different, even though similar, vehicle and sensor setup

were used compared to the original work. However, as an

approximate estimate the values obtained are sufficient.

Fig. 6 depicts the distributions of the logged samples over

the longitudinal ax(upper part) and lateral ayacceleration

(lower part). This shows the samples collected during several

highway test drives with an overall measuring period of more

than one hour. According to the visualization, the employed

position prediction approach seems to produce more faulty

predictions when negative longitudinal or positive lateral

accelerations occur. Even this small example with a very

limited amount of collected data shows that the presented

collection strategy has great potential to enhance training

data sets for motion predictions with meaningful samples.

Although this example does not refer to external conditions,

the same effects could be observed for such, when collecting

more data e. g. through a fleet of several vehicles.

1https://www.researchgate.net/publication/343392786 Prediction and

Memory Module within AR display

2For the sake of simplicity we restricted ourselves to solely investigate

the lateral direction here.

< -1.2 -1.0 -0.5 0.0 0.5 1.0 > 1.2

ax[m

s2]

100

200

Logged Samples

< -1.2 -1.0 -0.5 0.0 0.5 1.0 > 1.2

ay[m

s2]

100

200

Logged Samples

Fig. 6. Histograms showing the distribution of the logged data’s longitu-

dinal and lateral acceleration.

VI. SUMMARY AND OUTLOOK

We presented a new fleet learning-based data collection

architecture that ensures continuous improvements of motion

predictions of surrounding traffic participants. Especially,

this is beneficial during challenging external conditions. In

addition, the relevant elements of the pipeline were proto-

typically applied to a testing vehicle. Empirical evaluations

conducted with the testing vehicle prove the fundamental

feasibility of the system. Besides, the investigations show

that meaningful samples, which can be used to improve the

motion predictions, can be collected.

As the next steps of our research, we will expand the

memory component and conduct investigations based on

vehicles other than the ego-vehicle. Additionally, we plan

a rollout of the data collection architecture on a larger fleet

of testing vehicles. Furthermore, we will study the actual im-

provements of the motion prediction modules being enabled

by the collected data with respect to external conditions, as

soon as a larger data basis becomes available.

REFERENCES

[1] F. Wirthm¨

uller, J. Schlechtriemen, J. Hipp, and M. Reichert,

“Towards incorporating contextual knowledge into the prediction

of driving behavior,” in 23th International Conference on Intel-

ligent Transportation Systems (ITSC), IEEE, 2020. available:

https://arxiv.org/abs/2006.08470.

[2] J. Colyar and J. Halkias, “US highway 101 dataset,” Federal Highway

Administration (FHWA), Tech. Rep. FHWA-HRT-07-030, 2007.

[3] R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highD

dataset: A drone dataset of naturalistic vehicle trajectories on german

highways for validation of highly automated driving systems,” in

21st International Conference on Intelligent Transportation Systems

(ITSC), pp. 2118–2125, IEEE, 2018.

[4] J. Bock, R. Krajewski, T. Moers, S. Runde, L. Vater, and L. Eckstein,

“The inD dataset: A drone dataset of naturalistic road user trajectories

at german intersections,” arXiv preprint arXiv:1911.07602, 2019.

[5] V. Ramanishka, Y.-T. Chen, T. Misu, and K. Saenko, “Toward driving

scene understanding: A dataset for learning driver behavior and

causal reasoning,” in Conference on Computer Vision and Pattern

Recognition (CVPR), pp. 7699–7707, IEEE, 2018.

[6] L. Klitzke, C. Koch, A. Haja, and F. K¨

oster, “Real-world test drive

vehicle data management system for validation of automated driving

systems,” in 5th International Conference on Vehicle Technology

and Intelligent Transport Systems (VEHITS), pp. 171–180, INSTICC.

SciTePress, 2019.

[7] M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett,

D. Wang, P. Carr, S. Lucey, D. Ramanan, et al., “Argoverse: 3D

tracking and forecasting with rich maps,” in Conference on Computer

Vision and Pattern Recognition (CVPR), pp. 8748–8757, IEEE, 2019.

[8] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics:

The KITTI dataset,” The International Journal of Robotics Research

(IJRR), vol. 32, no. 11, pp. 1231–1237, 2013.

[9] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be-

nenson, U. Franke, S. Roth, and B. Schiele, “The Cityscapes dataset

for semantic urban scene understanding,” in Conference on Computer

Vision and Pattern Recognition (CVPR), pp. 3213–3223, 2016.

[10] H. Yu, S. Yang, W. Gu, and S. Zhang, “Baidu driving dataset and end-

to-end reactive control model,” in 16th Intelligent Vehicles Symposium

(IV), pp. 341–346, IEEE, 2017.

[11] J. Geyer, Y. Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S.

Chung, L. Hauswald, V. H. Pham, M. M¨

uhlegg, S. Dorn, et al., “A2D2:

Audi autonomous driving dataset,” arXiv preprint arXiv:2004.06320,

2020.

[12] A. Lerner, Y. Chrysanthou, and D. Lischinski, “Crowds by example,”

in Computer Graphics forum, vol. 26, pp. 655–664, Wiley Online

Library, 2007.

[13] S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never

walk alone: Modeling social behavior for multi-target tracking,” in

12th International Conference on Computer Vision (ICCV), pp. 261–

268, IEEE, 2009.

[14] A. Robicquet, A. Sadeghian, A. Alahi, and S. Savarese, “Learning

social etiquette: Human trajectory understanding in crowded scenes,”

in European Conference on Computer Vision (ECCV), pp. 549–565,

Springer, 2016.

[15] S. Lef`

evre, D. Vasquez, and C. Laugier, “A survey on motion pre-

diction and risk assessment for intelligent vehicles,” ROBOMECH

journal, vol. 1, no. 1, p. 1. Nature Publishing Group, 2014.

[16] R. Schubert, E. Richter, and G. Wanielik, “Comparison and evaluation

of advanced motion models for vehicle tracking,” in 11th International

Conference on Information Fusion (FUSION), pp. 1–6, IEEE, 2008.

[17] J. Schlechtriemen, F. Wirthmueller, A. Wedel, G. Breuel, and K.-D.

Kuhnert, “When will it change the lane? A probabilistic regression

approach for rarely occurring events,” in 14th Intelligent Vehicles

Symposium (IV), pp. 1373–1379, IEEE, 2015.

[18] F. Wirthm¨

uller, J. Schlechtriemen, J. Hipp, and M. Reichert, “Teaching

vehicles to anticipate: A systematic study on probabilistic behavior

prediction using large data sets,” IEEE Transactions on Intelligent

Transportation Systems (T-ITS). IEEE, 2020.

[19] C. Wissing, T. Nattermann, K.-H. Glander, and T. Bertram, “Trajectory

prediction for safety critical maneuvers in automated highway driving,”

in 2018 IEEE International Conference on Intelligent Transportation

Systems (ITSC), pp. 131–136, IEEE, 2018.

[20] J. Schlechtriemen, A. Wedel, J. Hillenbrand, G. Breuel, and K.-D.

Kuhnert, “A lane change detection approach using feature ranking

with maximized predictive power,” in 2014 IEEE Intelligent Vehicles

Symposium Proceedings (IV), pp. 108–114, IEEE, 2014.

[21] H. Cui, V. Radosavljevic, F.-C. Chou, T.-H. Lin, T. Nguyen, T.-K.

Huang, J. Schneider, and N. Djuric, “Multimodal trajectory predic-

tions for autonomous driving using deep convolutional networks,” in

2019 International Conference on Robotics and Automation (ICRA),

pp. 2090–2096, IEEE, 2019.

[22] A. Benterki, M. Boukhnifer, V. Judalet, and C. Maaoui, “Artificial

intelligence for vehicle behavior anticipation: Hybrid approach based

on maneuver classification and trajectory prediction,” IEEE Access,

vol. 8, pp. 56992–57002. IEEE, 2020.

[23] D. Lenz, F. Diehl, M. T. Le, and A. Knoll, “Deep neural networks for

markovian interactive scene prediction in highway scenarios,” in 2017

IEEE Intelligent Vehicles Symposium (IV), pp. 685–692, IEEE, 2017.

[24] M. Bahram, C. Hubmann, A. Lawitzky, M. Aeberhard, and

D. Wollherr, “A combined model-and learning-based framework for

interaction-aware maneuver prediction,” Transactions on Intelligent

Transportation Systems (T-ITS), vol. 17, no. 6, pp. 1538–1550. IEEE,

2016.

[25] T. Zhao, Y. Xu, M. Monfort, W. Choi, C. Baker, Y. Zhao, Y. Wang,

and Y. N. Wu, “Multi-agent tensor fusion for contextual trajectory pre-

diction,” in Conference on Computer Vision and Pattern Recognition

(CVPR), pp. 12126–12134, IEEE, 2019.

[26] M. Khakzar, A. Rakotonirainy, A. Bond, and S. G. Dehkordi, “A dual

learning model for vehicle trajectory prediction,” IEEE Access, vol. 8,

pp. 21897–21908. IEEE, 2020.

[27] L. Fang, Q. Jiang, J. Shi, and B. Zhou, “Tpnet: Trajectory proposal

network for motion prediction,” arXiv preprint arXiv:2004.12255,

2020.

[28] F. Altch´

e and A. De La Fortelle, “An LSTM network for highway

trajectory prediction,” in 20th International Conference on Intelligent

Transportation Systems (ITSC), pp. 353–359, IEEE, 2017.

[29] D. Vasquez, T. Fraichard, and C. Laugier, “Growing hidden markov

models: An incremental tool for learning and predicting human and

vehicle motion,” The International Journal of Robotics Research

(IJRR), vol. 28, no. 11-12, pp. 1486–1506, 2009.

[30] A. Thorvaldsson and V. Bandi, Reference path estimation for lateral

vehicle control. Master thesis, Chalmers University of Technology,

2015.

[31] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,

R. Wheeler, and A. Y. Ng, “ROS: an open-source robot operating

system,” in ICRA workshop on open source software, vol. 3, p. 5,

Kobe, Japan, 2009.