Towards Incorporating Contextual Knowledge into the Prediction of Driving Behavior [original]

Towards Incorporating Contextual Knowledge into the Prediction of

Driving Behavior

Florian Wirthm¨

uller , Julian Schlechtriemen , Jochen Hipp and Manfred Reichert

Abstract— Predicting the behavior of surrounding traffic

participants is crucial for advanced driver assistance systems

and autonomous driving. Most researchers however do not

consider contextual knowledge when predicting vehicle motion.

Extending former studies, we investigate how predictions are

affected by external conditions. To do so, we categorize different

kinds of contextual information and provide a carefully chosen

definition as well as examples for external conditions. More

precisely, we investigate how a state-of-the-art approach for

lateral motion prediction is influenced by one selected external

condition, namely the traffic density. Our investigations demon-

strate that this kind of information is highly relevant in order to

improve the performance of prediction algorithms. Therefore,

this study constitutes the first step towards the integration

of such information into automated vehicles. Moreover, our

motion prediction approach is evaluated based on the public

highD data set showing a maneuver prediction performance

with areas under the ROC curve above 97 % and a median

lateral prediction error of only 0.18 m on a prediction horizon

of 5 s.

I. INTRODUCTION

When thinking about human driving behavior, it seems

to be obvious, that it is not only affected by the current

traffic situation, but by various external conditions as well.

For example, the weather situation, traffic density or day-

time can depict such conditions. Knowledge about external

conditions is also used by human drivers for improving

their motion predictions of other traffic participants. This

context-awareness is one important aspect distinguishing the

ability of humans in predicting other vehicles movements

from the one of current advanced driver assistance systems.

Therefore, our hypothesis is that an improvement of the

system’s performance towards a human-like one can be

achieved by taking contextual information and especially

F. Wirthm¨

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

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. Most current motion prediction systems are not relying on external

conditions as inputs (visualized in blue). Our hypothesis is that future

systems can benefit by the integration of such information with an increased

prediction performance. Therefore, the output (yellow) of a future system is

in this example visualized as less optimistic and closer to the actual positions

as desired output (green).

knowledge about external conditions into account when

developing motion prediction systems. Fig. 1 visualizes this

thought. In this paper, we are investigating the impacts of

external conditions on driving behavior as well as on the

performance of current motion prediction systems.

Sec. I-A defines our understanding of the terms contextual

information and discriminates between situation context and

external conditions. Accordingly, we deduce the problem

definition. Sec. I-B follows with the articles’ contribution.

Hence, Sec. I-C closes the section with the following articles’

structure.

A. Problem Definition

We aim to investigate to which extent current motion

prediction systems are affected by varying context situations

and, thus, would potentially benefit from contextual informa-

tion. Especially, we investigate the lateral motion behavior,

which is more challenging to predict, but at the same time

less investigated compared to longitudinal behavior. To do

so, we have to provide a proper definition of the notion

of context first, as definitions and understandings are very

heterogeneous [1]. If we look at research related to the term

context-aware motion prediction, most works use a context

definition that solely includes features directly describing the

traffic situation. Exemplary features include traffic rules [2],

[3], intentions [4], information about the objects and lane

markings within the scene [2], [5]–[8], or map information as

intersection distances or topologies [2], [6], [9]–[11]. In [2],

all information on top of the vehicle kinematic is considered

to be context. As this is a rather wide definition, including a

lot of different aspects, we subdivide context features for

motion prediction into two categories. The first category

contains all above mentioned aspects, which are directly

connected to the traffic situation and can be characterized

as highly dynamic and mostly continuous. In the following,

this type of context information is called situation context.

In contrast, we use the term external conditions for features

such as weather, daytime, traffic density or country, which

apply to all vehicles in the given situation and can be

characterized as quasi-static. Therefore, the problem we are

tackling is to explicitly investigate the influence of external

conditions on lateral driving behavior in contrast to other

research focusing on the traffic situation context.

B. Contribution

The contribution of this article is threefold:

1) We discuss and categorize contextual knowledge for

motion prediction into traffic situation context and

external conditions.

2) We adopt the prediction methods presented in [8] to a

public data set and evaluate them carefully to enhance

comparability.

3) We present examinations showing the impact of ex-

emplary external conditions on driving behavior and,

thus, on motion prediction. Particularly, we showcase

the need for a fully-context-aware motion prediction

approach.

C. Article Structure

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

discusses related work. Sec. III summarizes the properties of

the used data set and presents data preprocessing steps. Thus,

Sec. IV outlines our experiments and their results. Finally,

Sec. V summarizes the contribution and gives an outlook on

future work.

II. RELATED WORK

This section gives an overview of related research on

context-aware motion prediction. The term context is mostly

used for what we call traffic situation context. The considered

works include research focusing on maneuver estimation [8]–

[10] as well as works from the field of trajectory or position

prediction [2], [3], [7], [8], [11]. Besides these approaches

intended for vehicles, context-aware motion prediction is

studied in other domains as well, such as human (e.g. [4],

[5]) or cyclist (e.g. [6]) path prediction. Studies with regard

to external conditions can be found in [12], [13]. These

works investigate the impact of weather conditions on driving

behavior but are not directly connected to the field of motion

prediction.

In [9], [10], an approach for predicting maneuvers at in-

tersections using topological and geometrical characteristics

is presented. The maneuver class is inferred with a Bayesian

network incorporating uncertainties.

[11] presents a Hierarchical Mixture of Experts (HME)

architecture for predicting spatial probability distributions at

intersections. The HME methodology divides the input space

in a binary decision tree fashion. Thereby, in each sub-space

the motion modeling is done by a specific expert. Within

the division process contextual or categorical features can

be incorporated as split criteria. The experts are infered from

training data and are represented as conditional probability

density function, which models the relationship between past

and future motions. As a benefit of this approach the input

features of the individual experts can remain constant, even

if additional context features shall be added. The evaluation

shows, that incorporating contextual information can be ben-

eficial for the prediction performance and that a hierarchical

division of the input space is preferable compared to an

augmentation of the feature space. [2] presents an approach

for predicting positions at intersections based on a dynamic

Bayesian network, combining modules learned from data and

constructed by expert knowledge. [3] presents a probabilistic

position prediction approach for road intersections, presum-

ing a predefined maneuver estimation. The actual regression

problem to find a tempo-spatial distribution is transformed

by a discretization into a multiclass classification problem.

The classification problem is addressed by training a neural

net that, amongst others, contains map and traffic information

as features.

Besides these approaches we want to mention two works,

directly related to our studies. In [7], for the first time, the

highD data set ([14]), which is also used in this study, is

used to implement and evaluate long-term motion predictions

in highway situations. A relational recurrent neural network

based encoder-decoder architecture is used for the predic-

tions. The approach is able to predict lateral vehicle motions

over a time horizon of 5 s with a root mean squared error of

0.48 m. Moreover, [8] presents a systematic machine learning

workflow for the development of a system for maneuver

detection and probabilistic motion prediction. It compares

different classifiers and strategies, showing that a Mixture of

Experts architecture using a multilayer perceptron classifier

as gating node is beneficial compared to other combinations.

This combination reaches a median lateral prediction error

of 0.21 m on a prediction horizon of 5 s. To conduct our

context-related studies, we use this architecture as starting

point and will add further details in Sec. IV-A.

As aforementioned, there are only few works investigating

the influence of external conditions on road traffic [12], [13].

These works study the impact of weather conditions on the

driving behavior, but are not directly linked to motion pre-

diction. [12] investigates longitudinal driving behavior under

fog influence in a driving simulator study. The investigations

show decreased speeds and accelerations as well as increased

distances to preceding vehicles during foggy weather. [13]

studies both, the impact of weather conditions and road

geometries on longitudinal driving behavior when following

a lead vehicle in a driving simulator study and gives a

good literature survey in that area. Although [13] states that

the overall effect of weather conditions is smaller than the

one of challenging road geometries, the impact of weather

conditions becomes apparent. As opposed to [12] the study

shows that fog has only little impact on the driven speed,

but in compliance with [12] fog implicates higher distances.

In addition, lower speeds are observed when driving on wet

or icy roads.

Altogether, our literature review confirms, that context-

aware prediction of driving behavior is studied by various

works, most of them focusing on the traffic situation context,

whereas external conditions have been neglected so far (cf.

Sec. I-A). As additional limitations, most works focus on

urban driving and only investigate the behavior change on

maneuver layer. However, as shown in [12], [13], contextual

features play also an important role for the concrete trajectory

realization and not only for a maneuver decision. Interest-

ingly, most works are studying longitudinal rather than lateral

behavior.

III. DATA PREPARATION

To train and evaluate our algorithms, we rely on the

recently published highD data set [14]. It consists of ap-

proximately 45 000 km of highway driving. The data set was

recorded in 6 different highway sections in Germany using a

drone-mounted camera system. As opposed to the formerly

frequently used NGSIM data set [15], the highD data set

exhibits a considerably higher data quality and quantity as

well as a higher variety.

We prepare the highD data set for our algorithms and

context-related investigations, by performing several pre-

processing steps: First, we calculate a range of additional

features describing traffic density, lane changes and lane

marking types as well as relations to the eight surrounding

vehicles. The relations are described with a static environ-

ment model used by many researchers (see e. g. [16] for

an introduction). Besides, we transform the data in an equal

representation for both driving directions, containing the lane

number in ascending order from right to left. In addition,

we exclude data from the training and evaluation processes

if front- or backsight distances are lower than 80 m. This

ensures that we only train and perform predictions based on

a known system environment as expected in an in-vehicle

application. In order to reflect real-world sensor capacities,

we assume that vehicles are not able to perceive objects being

farther away than 150 m and put virtual sensor limitations by

setting default values. Moreover, we adopted the strategy to

label lane changes within a prediction horizon THof 5 s

presented in [8]. As opposed to [8] we added a fourth label,

representing samples with an undefined status NDEF due

to a too short observation time TO. Eq. 1 defines our labeling

strategy.

TLCL and TLCR reflect the times to the next lane change

to the left LCL and to the right LCR, respectively. The

FLW label is used for lane following behavior. TOdescribes

the remaining time meanwhile the object is in recording

range and can thus be observed.

After the labeling, we split all data with a defined maneu-

ver class into 6 folds, of which 5 are used for training and the

remaining one for evaluating the models. In our experiment,

we only work with data from passenger cars, as the motion

of cars and trucks deviates noticeable.











LCL, if TLCL ≤TH&

TLCL ≤TLCR &

TLCL ≤TO

LCR, if TLCR ≤TH&

TLCR < TLCL &

TLCR ≤TO

FLW, if TH< TLCL &

TH< TLCR &

TH≤TO

NDEF, otherwise

(1)

IV. EMPIRICAL STUDIES & RESULTS

This section first describes the experimental setup

(Sec. IV-A). Afterwards, Sec. IV-B presents empirical in-

vestigations with regard to the system’s ability to predict

upcoming maneuvers and future lateral positions. Sec. IV-

C presents investigations that show the impact of features

associated with external conditions on the driving behavior.

A. Experimental Setup

For performing our studies, we adopted the prediction

approach we introduced in [8]. As aforementioned, our

approach uses a multilayer perceptron to predict the upcom-

ing maneuver. For the sake of simplicity, we re-used the

hyper-parameters of the model and re-trained the maneuver

classifier with the highD data set:

•Step size: 0.02

•Structure: one hidden layer with 27 neurons; 3 output

neurons

•Feature set: The considered feature set is corresponding

to the one that had shown the best results in [8] except to

the yaw angle, as this feature is not available in highD.

Furthermore, a transformation to lane coordinates and

a differentiation between the features in the different

coordinate systems is not necessary in the given study,

as applied in [8] as the highD data set solely contains

straight road segments.

In accordance with [8], we used a random undersampling

strategy to ensure that the training data set is balanced over

time to the next lane change as well as over maneuver

classes.

Besides the multilayer perceptron, the approach uses three

Gaussian mixture models, modeling future lateral positions

depending on several input features. Thereby, each model

is intended as expert for one of the three maneuver classes

LCL,F LW and LCR. To train the models, we again used

the following hyper-parameters from [8] and retrained them

with the highD data set variationally:

•Maximum number of kernels: 50

•Type of the covariance matrix: full

•Input features: lateral velocity vy, distance to the center

of the current lane dcl

•Output dimensions: lateral position y, time t

CV Labels MLP_PW-Raw

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Ey,5 [m]

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0

Prediction Horizon t [s]

0.00

0.05

0.10

0.15

0.20

0.25

0.30



Ey,t [m]

CV Labels MLP_PW-Raw

Fig. 2. Distribution of the lateral error at t=5 s (left) and the median lateral error as function of the prediction time (right). The lateral error of the

used system (denoted as ’MLP PW-Raw’) is shown in comparison to a constant velocity (’CV’) estimation and a Mixture of Experts assuming a perfect

classifier (’Labels’).

The described individual components are used together

in a Mixture of Experts fashion to predict distributions of

future vehicle positions. A single position estimate can be

calculated out of the distribution as center of gravity. For a

more elaborate overview of the prediction technique and its

parametrization see [8].

B. Prediction Performance Evaluation

To evaluate the performance of the maneuver classifier, we

use the sixth data fold, which was left out during the models

training. As evaluation metrics, we rely on the same measures

as [8]: the balanced accuracy (BACC), the receiver operator

characteristic (ROC), the area under that curve (AUC) and

the time gain τc. Note, that τcmeasures the time between the

vehicle center crosses the centerline and the moment from

that the classifier is certain about its decision for a specific

maneuver class and does not change it till the end of the

situation. The results are presented in Fig. 3.

0.00 0.05 0.10 0.15 0.20

False Positive Rate

0.2

0.4

0.6

0.8

1.0

True Positive Rate

BACC = 0.908

τLCL

c=3.250±1.984 s

τLCR

c=1.916±2.056 s

FLW (AUC = 0.971)

LCL (AUC = 0.991)

LCR (AUC = 0.990)

WP for BACC

WP for τ

Fig. 3. ROC curve showing the performance of the used maneuver classifier

applied to the highD data set [14].

As Tab. I shows, the maneuver classification performance

is even superior to the results presented in [8]. This can

be explained with the fact that the highD data set does

not contain curved roads, as in [8] where an error-prone

and complex transformation to curvilinear coordinates had

to be performed. Solely the time gain for lane changes to

the right drops, although the overall maneuver classification

performance increases. This may be explained with the

removal of the yaw angle from the feature set, resulting in

a decreased classification stability.

TABLE I

MANEUVER CLASSIFICATION PERFORMANCE COMPARED TO THE

ORIGINAL STUDY [8]

AUC τc[s]

LCL F LW LCR LCL LCR

[8] 0.976 0.915 0.960 2.72 2.68

This study 0.991 0.971 0.990 3.250 1.916

Benefit +0.015 +0.066 +0.030 +0.530 -0.764

For evaluating the performance of the position prediction,

we use the distribution of the lateral prediction error Ey, 5

at a prediction time of 5 s and the median lateral error ˜

as a function of the prediction time tas metrics. The results

of our approach are visualized in comparison to a constant

velocity (CV) prediction and a Mixture of Experts assuming

a perfect maneuver classification (Labels) in Fig. 2.

As the evaluation shows, the maneuver classifier seems

to be that good, that the performance of the down-streamed

position prediction is nearly as good as when presuming a

perfect classifier. When looking at the errors as a function

of the prediction horizon, it becomes apparent that at some

prediction horizons (e. g. around 2.5 s) our approach is even

better than a perfect classifier. This can be explained with

the fact that the predicted position is calculated as a correctly

weighted mixture of different hypotheses. As opposed to the

original study [8], the median lateral position error ˜

Ey, 5at a

prediction horizon of 5 s is decreased from less than 0.21 m

to less than 0.18 m.

C. Context Influence Investigations

As discussed, our main goal is to investigate and substanti-

ate the impact of features associated with external conditions

on lateral driving behavior predictions. To accomplish this

goal, first, we need to discuss which conditions can be

investigated appropriately based on the available data set:

5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 27.5

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

Ey,5 [m]

FLW

LCL

LCR

Fig. 4. Median lateral prediction error on a prediction horizon of 5 s as

function of the traffic density for the different maneuver classes.

1) Daytime: Although the highD data set contains day-

time data, they are limited to a range from 9 am to 8 pm.

Thus, the nighttime, for which we would expect a significant

behavior change due to poor visibility conditions, is not

included. We further believe that other effects, such as

commuters, can be preferably represented with other features

such as traffic density.

2) Day of the Week: As the highD data set does not

contain recordings at weekends and the data density over

the other days of the week is very unbalanced, we do

not consider investigations concerning that property to be

fruitful.

3) Weather: Other researchers as e. g. [12], [13] already

showed that the impact of varying weather conditions (on the

road e. g. icy or wet roads, or above the road e. g. rain, fog

or snow) on driving behavior is remarkable. As the highD

data set, however, is recorded with optical drone-mounted

sensors, sufficient visibility as well as flight conditions have

to be satisfied. Consequently, the data set does not contain

such data points.

4) Location: Another external condition constitutes the

location, which has assorted characteristics as well. For

example, the context location may depict the current country,

prevalent speed limits, special road elements (e. g. bridges,

tunnels or on-/off-ramps), patterns describing the road ge-

ometry or only one particular geo-location. As most other

mentioned characteristics of the location context are hard to

represent and to investigate due to an insufficient data situa-

tion, we are very optimistic with respect to the investigation

concerning different countries. While the highD data set, as

of now, only contains data recorded on German highways,

we aim to gather additional data in the same format from

other countries.

5) Traffic Density: Another external condition whose im-

pact on driving behavior is apparently obvious constitutes

the traffic density. As opposed to other effects, however, the

traffic density offers two important benefits. In the first place,

we are able to calculate traffic density measures for the highD

6 8 10 12 14 16 18 20

8.0

8.2

8.4

8.6

8.8

9.0

∆TLC [s]

0.80

0.85

0.90

0.95

1.00

1.05

1.10

|vmax

y|[m

Fig. 5. Lane change duration (black) and maximum lateral speed (yellow)

during lane change as functions of the traffic density

data set in a very stable way due to the vertical view onto

the scene. In addition, we are able to represent also other

effects as previously mentioned in the discussion concerning

daytimes through the traffic density. Therefore, we believe

that the most promising research direction, is to investigate

the influence of the traffic density.

To work with that property, we calculate the density values

TD in the highD data set according to Eq. 2 as number of

vehicles NVper km and number of lanes NL(cf. [17]):

TD =NV

km ·NL(2)

Fig. 4 presents the median lateral prediction error ˜

Ey, 5on

a prediction horizon of 5 s as function of the traffic density

for the different maneuver classes. This investigation shows

contrary trends for lane change and lane following situations.

While the prediction error during lane following decreases

with increasing traffic density, it increases or at least remains

in the same magnitude during lane changes.

A possible explanation for this observation could be that

the increased traffic density decreases the average longitudi-

nal speed. On the one hand, the lateral oscillation, human

drivers normally conduct when following a highway lane,

has a smaller amplitude in the local area, as the frequency is

constant. On the other, the duration of lane changes increases

while the traffic becomes more dense. For example, that

effect is substantiated in [18] and confirmed by our own

investigations in Fig. 5.

Moreover, Fig. 5 shows that the maximum lateral speed

during lane changes tends to increase, while the traffic

density increases. This means, that one can imagine the

procedure of a lane change during dense traffic as a long

period of time where the driver slightly steers towards

the lane marking and announces his lane change intention.

Afterwards, the actual lane change is performed very abrupt

and with a high maximum lateral speed, if a proper gap

is reachable. In contrast, in light traffic, lane changes are

performed slower and smoothly. To illustrate this, Fig. 6

shows two lane change maneuvers from the highD data set.

50 100 150 200 250 300 350 400

x[m]

11.67

7.69

4.07

0.00

y[m]

P(t)

O(t)

TR(P)

vmax

y= 0.64m/s

TD = 9.75

∆TL C = 8.08s

t=TB

533

534

535

539

529

t=TE

533

535

539

50 100 150 200 250 300 350 400

x[m]

11.67

7.69

4.07

0.00

y[m]

P(t)

O(t)

TR(P)

vmax

y= 1.48m/s

TD = 12.69

∆TL C = 9.60s

633

635

636

637

t=TB

639

640

641

643

633

636

637

t=TE

639 641

643

644

Fig. 6. Visualization of two lane change situations out of the highD data set (recording 35). Whereas, the lane change in the upper example is performed

under an increased traffic density and abruptly with a high maximum lateral velocity, the one in the lower example is performed very smoothly. The

visualization shows for both examples the positions of the prediction target P(t)and of the relevant surrounding objects O(t)at the begining TBand the

end TEof the lane change maneuver as well as the whole trajectory of the prediction target T R(P).

As result of the described effects, it becomes presumably

more complex to predict lane changes in more dense traffic

as the variance increases.

To investigate this, we defined the duration of lane changes

∆TLC according to Eq. 3 as time difference between begin

TBand end TEof the lane change:

∆TLC =TE−TB(3)

TBand TEare defined as the first points in time before

and after the lane change satisfying the conditions C(TB)

and C(TE)according to Eq. 4 and Eq. 5:

C(TB) = |dcl

y| ≥ 1m

∨|vy| ≥ 0.1m

∨|ay| ≥ 0.1m

∨|jy| ≥ 0.1m

(4)

C(TE) = C(TB)(5)

dcl

ydescribes the lateral distance to the lane center, vythe

lateral velocity, aythe lateral acceleration and jythe lateral

jerk. We determined the actual threshold values through

visual inspection of all lane change situations.

V. SUMMARY AND OUTLOOK

Other works have already shown the impact of external

conditions such as weather ([12], [13]) on driving behavior.

We are complementing this with investigations with respect

to the traffic density. Our study goes beyond others, not

only examining the impact on driving behavior itself, but

instead the direct impact on the prediction performance.

The results substantiate that motion prediction systems for

automated vehicles could significantly benefit from explicitly

considering what we call external conditions.

Besides, we investigate the prediction performance of

the approach originally presented in [8] with the publicly

available highD data set. The examinations show a maneuver

classification with areas under the curve above 97 % and

a median lateral prediction error of less than 0.18 m for a

prediction horizon of 5 s.

As extension of the presented study, a full adaption of the

approach presented in [8], including the entire featureselec-

tion and hyper-parameter optimization pipeline, could further

improve the results. In addition, we are preparing a follow-up

study, presenting a fleet-learning-based architectural concept.

Thereby, we will show how relevant data for context-aware

motion prediction applications can be collected and used to

ensure a continuous improvement of a vehicle fleets pre-

diction capabilities under varying contextual influences. To

enable more context-related investigations it is also desirable

to conduct measurements during various contexts as night,

rain, snow or in different countries.

REFERENCES

[1] M. Baldauf, S. Dustdar, and F. Rosenberg, “A survey on context-aware

systems,” International Journal of Ad Hoc and Ubiquitous Computing

(IJAHUC), vol. 2, no. 4, pp. 263–277. Inderscience Publishers, 2007.

[2] T. Gindele, S. Brechtel, and R. Dillmann, “Learning context sensitive

behavior models from observations for predicting traffic situations,”

in 16th International IEEE Conference on Intelligent Transportation

Systems (ITSC), pp. 1764–1771, IEEE, 2013.

[3] S. Klingelschmitt and J. Eggert, “Using context information and

probabilistic classification for making extended long-term trajectory

predictions,” in 18th International Conference on Intelligent Trans-

portation Systems (ITSC), pp. 705–711, IEEE, 2015.

[4] F. Schneemann and P. Heinemann, “Context-based detection of pedes-

trian crossing intention for autonomous driving in urban environ-

ments,” in IEEE/RSJ International Conference on Intelligent Robots

and Systems (IROS), pp. 2243–2248, IEEE, 2016.

[5] F. Bartoli, G. Lisanti, L. Ballan, and A. Del Bimbo, “Context-

aware trajectory prediction,” in International Conference on Pattern

Recognition (ICPR), pp. 1941–1946, IEEE, 2018.

[6] E. A. Pool, J. F. Kooij, and D. M. Gavrila, “Context-based cyclist

path prediction using recurrent neural networks,” in 18th Intelligent

Vehicles Symposium (IV), pp. 824–830, IEEE, 2019.

[7] K. Messaoud, I. Yahiaoui, A. Verroust-Blondet, and F. Nashashibi,

“Relational recurrent neural networks for vehicle trajectory predic-

tion,” in 22nd Intelligent Transportation Systems Conference (ITSC),

pp. 1813–1818, IEEE, 2019.

[8] 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,” Transactions on Intelligent Trans-

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

available: arXiv preprint arXiv:1910.07772.

[9] S. Lef`

evre, J. Iba˜

nez-Guzm´

an, and C. Laugier, “Context-based es-

timation of driver intent at road intersections,” in Symposium on

Computational Intelligence in Vehicles and Transportation Systems

(CIVTS), pp. 67–72, IEEE, 2011.

[10] S. Lef`

evre, C. Laugier, and J. Iba˜

nez-Guzm´

an, “Exploiting map

information for driver intention estimation at road intersections,” in

10th Intelligent Vehicles Symposium (IV), pp. 583–588, IEEE, 2011.

[11] J. Wiest, F. Kunz, U. Kreßel, and K. Dietmayer, “Incorporating cate-

gorical information for enhanced probabilistic trajectory prediction,” in

12th International Conference on Machine Learning and Applications

(ICMLA), vol. 1, pp. 402–407, IEEE, 2013.

[12] R. G. Hoogendoorn, G. Tamminga, S. P. Hoogendoorn, and W. Daa-

men, “Longitudinal driving behavior under adverse weather condi-

tions: Adaptation effects, model performance and freeway capacity in

case of fog,” in 13th International IEEE Conference on Intelligent

Transportation Systems (ITSC), pp. 450–455, IEEE, 2010.

[13] S. H. Hamdar, L. Qin, and A. Talebpour, “Weather and road geometry

impact on longitudinal driving behavior: Exploratory analysis using an

empirically supported acceleration modeling framework,” Transporta-

tion Research Part C: Emerging Technologies, vol. 67, pp. 193–213.

Elsevier, 2016.

[14] 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.

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

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

[16] 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.

[17] Forschungsgesellschaft f¨

ur Straßen-und Verkehrswesen, “Begriffsbes-

timmungen - Teil: Verkehrsplanung, Straßenentwurf und Straßenbe-

trieb,” 2000.

[18] T. Toledo and D. Zohar, “Modeling duration of lane changes,” Trans-

portation Research Record, vol. 1999, no. 1, pp. 71–78. SAGE

Publications Sage CA, 2007.