Traffic Data Evaluation for Automated Driving Handover Scenarios

Eugenia Rykova

1,3,

, Juri Golanov

, Jonas Vogt

, Daniel Rau

and Horst Wieker

University of Applied Sciences TH Wildau, Wildau, Germany

ITS Research Group FGVT htw saar, Saarbrücken, Germany

University of Eastern Finland, Joensuu, Finland

Keywords: Connected and Automated Driving, Traffic Data Evaluation, Automated Driving Handover.

Abstract: At the current stage of automated vehicle development, the control handover from the system to a human

driver (and back) is inevitable. It is essential to distinguish between situations in which the handover is

possible and in which it could be dangerous and is therefore highly undesirable. We evaluated traffic situations

based on two modalities: own vehicle state and traffic objects. To assess the former, supervised machine

learning was applied, reaching an accuracy of 80.3% and specificity of 77.8% with Multilayer perceptron

Classification. Traffic objects data were subject to different clustering techniques. The final grouping was

done according to manually elaborated rules, resulting in a range of situation complexity scores. Improving

the discriminative power of vehicle state classification, including driver’s state and weather information, and

predicting situation complexity are to be addressed in future research.

1 INTRODUCTION

The future of mobility is automated. Researchers all

over the world are working towards fully automated

vehicles. Connected and cooperative automated

mobility (CCAM) is one important keystone to

accomplishing that goal. Vehicles by themselves can

only have a limited view as today the driver of a

vehicle. To enhance the safety and efficiency of road

traffic, cooperation and information exchange are

vital to see around the corner and to help traffic run

smoothly. The research goal of fully automated

condition-independent driving (SAE Level 5 (SAE

International, 2018)) is not yet on the horizon for

market introduction. Levels 3 and 4 serve as transition

steps from lower-level driver-support features. At the

current moment, Level 3 automated vehicles (AV)

just start being publicly available (Honda, 2020).

Level 3 implies occasional handover from

autonomous to human control, which is one of its

challenges. In this paper, we particularly focus on the

control handover from automated to manual driving.

There are many possible reasons for this handover.

Automated driving could only be allowed on some

roads or road classes, the driver could indicate the

willingness to drive because she wants the pleasure to

former member of FGVT htw saar

drive or feels more comfortable driving in a certain

situation. The handover, however, should not be a

surprise for the driver and therefore requires the

driver’s awareness and sufficient transition time

(Trimble et al., 2014). It can be planned, and the

transition can be done smoothly in situations, where

the traffic allows a switch of responsibility. There are,

however, more demanding and challenging

situations. What if an automated vehicle is not able to

steer through a certain situation because of technical

limitations or the fact that not enough sensor

information is available (e.g., due to weather

conditions or malfunction)? In some situations,

especially in urban environments, the handover is not

possible. For example, at a confusing construction

site or an intersection with many pedestrians, a human

driver needs time to adapt to the situation and gain an

overview.

In this paper, situations that could hinder a

handover process from the automated driving to the

driver are investigated. The focus is on determining

which situations in an urban environment are critical

and should not be used for handover scenarios.

The paper is structured as follows: first, the

related work is presented. Then, the methods of how

we handled the data are described in Section 3. In

Section 4, the results of the data modalities are

Rykova, E., Golanov, J., Vogt, J., Rau, D. and Wieker, H.

Trafﬁc Data Evaluation for Automated Driving Handover Scenarios.

DOI: 10.5220/0011599900003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 125-134

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

125

analyzed. The conclusion in Section 5 features the

significance of the results that were obtained and

highlights open research questions.

2 RELATED WORK

Most research related to the transition from

automated driving to manual driving is related to the

human factor of reaction to the takeover request

(TOR) (Eriksson and Stanton, 2017; Clark et al.,

2020). The awareness of the drive may be increased

by augmented reality (Schroeter and Steinberger,

2016). Some research sees the decision between the

two options: drivers are allowed to do other tasks or

drivers have to monitor the system at all times

(Borojeni et al., 2017). The traffic situation has an

important impact on the reaction time and the

probability of accidents (Radlmayr et al., 2014; Gold

et al., 2016).

To assess the possibility of a takeover, the time

necessary for the driver to react is the decisive factor

(Ayoub et al., 2022). Some research estimates the

readiness of the driver to take over the driving task

based on the complexity of the traffic situation, the

secondary task, and the gazes at the road (Braunagel

et al., 2017). For this estimation, only vehicle sensors

are used. To our knowledge, no research was

conducted using the combination of infrastructure

and vehicle sensor information to estimate factors for

takeover. More information can help to lower the

stress factor during and shortly after the handover

(Kerautret, 2023)

3 MATERIALS AND METHODS

3.1 Data Collection

Traffic situation information consisted of data

gathered from the following sources:

 Own vehicle state, consisting of sensor data

from the vehicle Controller Area Network

(CAN) bus (ISO, 2015) system.

 Information from message exchange between

other road users and infrastructure via Vehicle-

To-Everything (V2X) communication, such as:

o Detected road users in the immediate

vicinity by optical sensors (e.g., traffic

cameras) at intersections; distribution of this

information by Collective Perception

Messages (CPM) (ETSI, 2019a) of

Intelligent Transportation System (ITS)

Roadside Stations (IRS);

o Other V2X communication like warnings of

dangerous events via Cooperative

Awareness Messages (CAM) (ETSI, 2019c)

and Decentralized Environmental

Notification Messages (DENM) (ETSI,

2019d) from ITS Vehicle Stations (IVS’s)

and IRS’s (e.g., broken down vehicle

warning, pedestrian collision warning, etc.).

 Topology information of intersections: MAP

(ETSI, 2019b) information at traffic

intersections.

 Weather conditions: weather information

provider in the backend.

Since all the data were collected under similar

weather and lighting conditions (daylight, warm

temperature, zero precipitation), the weather data

were not included in the further analysis as a

discriminative factor.

To control the condition of the driver, driver-

monitoring functionality and corresponding

equipment must be included in an (automated)

vehicle. However, in the current research, data on the

driver’s condition are not available yet, so the

evaluation is based only on non-driver data.

Description of traffic situations included 1) data

recording, 2) aggregation and fusion of information

from several sources, and 3) storage in the database.

Data collection and pre-processing are described in

(Otte et al., 2021). Several test drives were performed

to generate the (training) data in the city of

Saarbrücken. It must be noted that our test vehicle

was not automated so the control handover was

explored hypothetically.

A traffic situation is a certain point in time

represented by the vehicle state and the detected

objects at the corresponding traffic intersection. A

test drive represents a chain of several successive

traffic situations. The time interval between two

situations is one second. Since situations are

snapshots, they were assessed individually, or

independently from the previous and following states.

During the execution of several test drives, a total of

7,854 traffic situations were recorded and stored in

the fusion database.

The goal of the evaluation is to determine the

degree of suitability of traffic situations for the

handover, in other words – the degree of situation

complexity. To keep the decision-making process

transparent and explainable, a cumulative multimodal

approach to situation evaluation was chosen. One

modality was sensor data of the test vehicle, and the

other modality was the information on detected traffic

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

126

objects. The final decision depended on the outcomes

of each modality evaluation.

3.2 Vehicle Data Modality

Own state of the vehicle has a direct straightforward

influence on the control handover. “Handover

possible” and “handover not possible” are binary

labels that could be assigned to situations by human

raters and are further used for supervised machine

learning (classification algorithms and logistic

regression) (Awad & Khana, 2015).

An evaluation basis was required for labeling the

handover as possible or not possible. For this purpose,

videos were recorded from the interior of the test

vehicle during the test drives, capturing the driver's

point of view. The videos were time-stamped so that

they could be matched to the data in the fusion

database. With the help of the self-generated MAP

messages and the GNNS (global navigation satellite

system) position information from the test vehicle, it

was possible to perform lane matching with the

corresponding traffic light phases. This provided an

information gain to the overall traffic situation.

Recorded traffic situations were assessed

empirically according to the four-eyes principle.

First, transition points, where the handover status

changed from possible to not possible and vice versa,

were determined. All time points in between counted

as having the same handover status. The time interval

between two situations was set to one second to keep

a certain degree of differentiation of adjacent

situations as well as to avoid generating too much

data. “Handover possible“ (1) and “handover not

possible” (0) labels were stored in the situation

database, which was linked to the vehicle database

and traffic objects database through equivalent

situation identification numbers.

It was assumed that the movement state of the

vehicle was the most decisive for the evaluation,

whether it was moving, stationary, accelerating, or

braking. Based on the mentioned features, a decision

was made on whether the vehicle state was suitable

for the handover to the driver in the current situation

(see Table 1).

Table 1: Presumed handover decision matrix.

Handover

ossible Handover not

ossible

Vehicle is stationary

(

even at red traffic li

)

Traffic light turns green,

vehicle starts movin

Vehicle moves at a

constant spee

Vehicle accelerates or

rakes

Vehicle is driving

strai

ht ahea

Vehicle is in a curve (or at

an intersection

)

This information was provided directly by the

vehicle sensors from the CAN bus interface. Vehicle

data dimensionality was reduced by filtering out the

features, whose values did not change during the test

drive. The remaining features are listed in Table 2.

Table 2: Vehicle data set.

Brake

actuation

Direction

of drivin

Current

Clutch switch

actuation

Door

osition

Hazard

warning

Lateral

acceleration

Longitudinal

acceleration

Pedal

force

Speed Steering

wheel angle

Steering

wheel angle

velocit

Turn

signal

level

Wiper

front

stem

Yaw rate

The classification process consisted of applying

different classification algorithms to the vehicle data

set and comparing the accuracy score. To achieve a

better understanding of how the model would perform

in practice, 10-fold cross-validation (70/30 split) was

applied, after which the mean accuracy score of each

algorithm was calculated. All utilized algorithms and

methods came from the free software machine

learning library scikit-learn (Pedregosa et al., 2011).

For selected algorithms, an attempt to increase the

accuracy was made by applying “GridSearchCV”

(CV = Cross Validation), which performed an

exhaustive search on parameter values for the best

estimator. Optimized models were evaluated not only

based on the accuracy but on the specificity of the

model (True Negatives / (True Negatives + False

Positives)), which was considered a more important

metric for the investigated scenarios (False positives

would be more dangerous errors than False

Negatives).

3.3 Traffic Objects Modality

Situation complexity/criticality level was considered

depending on the level of danger of each traffic object

present in the current situation so that the primary

goal was to find a way of assessing the danger level

of individual traffic objects.

Traffic objects data consisted of entries for 18,030

objects of two types: 13,384 (passenger) cars and

4,646 pedestrians. These objects corresponded to

1,868 situations. The number of objects in a situation

ranged from 1 to 33, mean (M) = 9.7, and standard

deviation (SD) = 6.9. Each data entry contained the

following features:

 situation id;

Trafﬁc Data Evaluation for Automated Driving Handover Scenarios

127

 object information: type of object; distance to

the test vehicle; speed; heading; longitude;

latitude; time to intersection (tti);

 test vehicle information in the corresponding

situation: heading; longitude; latitude; time to

intersection (tti_v).

Time-to-intersection calculations are explained in

(Jiménez et al., 2013). When the intersection point did

not lie ahead on the course of the object/test vehicle,

the time-to-intersection was set to -1 (e.g., when the

object and the test vehicle moved parallel). Including

a negative value, however, disrupted the continuity of

the feature: -1 was not less than non-negative values.

Since maximum positive values of time-to-

intersection features were great enough (ttimax=

83,542.98 s; tti_vmax = 43,955.02 s) to be seen as

irrelevant at the current time point, all the -1 values

were converted into the maximum positive value of

the corresponding feature.

To make the data points more comparable and

easier to visualize, they were brought into the same

2D-coordinate system where the test vehicle would be

at the origin (0,0) moving towards the geographical

North (compass bearing = 0°, which corresponds to

π/2 (90°) in the polar coordinate system). With this

we consider the world view in this paper to be flat,

because here the curvature of the earth is negligible.

First, the differences between the longitude (long)

and latitude (lat) of the object and the longitude and

the latitude of the vehicle, respectively, were

calculated. Then these differences were expressed in

meters:

𝑥=

∙



°∙(











)∙



; (1)

𝑦 =

∙



°∙



; (2)

𝑤ℎ𝑒𝑟𝑒 𝑟𝑎𝑑𝑖𝑢𝑠



= 6371000.8 𝑚.

Relative X and Relative Y were calculated, using

trigonometric formulae for an axis rotation (Becker et

al., 1999, p. 48) for clockwise rotation through an

angle of the test vehicle’s heading. Longitude and

latitude were mapped onto the abscissa and the

ordinate, ignoring the z-axis because the objects were

relatively close to each other (the maximum distance

from the test vehicle was 100 m).

The next step was to calculate the relative bearing

of the objects (IVAO, 2020). For still objects, the true

bearing was calculated first, using the coordinates of

an object and the test vehicle (Ellis, 2020). Then, the

relative bearing was calculated by subtracting the

vehicle’s heading (true heading) from the object’s

true bearing. For moving objects, their heading was

used as the true bearing. The values were normalized

to the [0, 360°) range.

These and the following calculations were

performed in JupyterLab environment, using various

Python libraries (Pedregosa et al., 2011; Van Rossum,

2020; Harris et al., 2020; Virtanen et al., 2020; Gillies

et al., 2007), unless otherwise specified. Data

visualizations were carried out in matplotlib (Hunter,

2007). Illustrations of traffic objects in the coordinate

system were obtained with the help of

TeachingDemos package (Snow, 2020) in R (R Core

Team, 2020).

3.4 Clustering

Situation complexity is multifaceted. First, the

number of participants and their type varied. On the

other hand, each object was described with several

features. The situation’s complexity itself was

difficult to define in terms of the limited number of

labels. It seemed more reasonable to explore the data

and find the tendencies to group the objects, which

were addressed via unsupervised machine learning, or

clustering (Awad & Khana, 2015).

Several clustering options were explored to group

the objects. The evaluation of clusters was empirical.

Thus, a sample (<=80 samples) of observations from

each cluster was visualized and analyzed, whether

most of the objects in one cluster could be described

as having the same danger level. The algorithms were

first applied to the subset of data for pedestrians

because it was smaller and required less

computational space and time. When the approach

was not considered suitable for the data (i.e., the

results of clustering were not interpretable in terms of

intuitive understanding of the danger level), it was not

further applied to the subset of data for cars.

3.4.1 Position and Speed Features

Position features included relative X and Y

coordinates and relative bearing. To be treated as a

circular variable in centroid-based clustering, the

relative bearing was transformed into two features:

sine and cosine of the angular value. Since both sine

and cosine functions take values from -1 to 1, the

other features were scaled by their maximum absolute

value. An attempt to treat relative bearing linearly

proved to be inappropriate for circular variables.

Two centroid-based clustering algorithms were

applied to five features (scaled speed, scaled relative

X, scaled relative Y, sine of relative bearing, and

cosine of relative bearing), namely, Mean shift and K-

Means. For Mean shift clustering, bandwidth was

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

128

estimated with a given quantile = 0.05. For K-Means

clustering, the Elbow Method was used to select the

optimal number of clusters.

Hierarchical agglomerative clustering was

performed in R (Murtagh & Legendre, 2014). For the

dissimilarity matrix with a circular variable, two

options were explored: one implemented in dist.ktab

function (Pavoine et al., 2009) of the ade4 package

(Dray & Dufour, 2007), and the other –

implementation of Gower’s dissimilarity (Maechler

et al., 2019) proposed by Will (2016). Will discusses

both methods applied to The Cape Blanco dataset,

which contains hourly measured temperature, wind

speed, and wind direction. The optimal number of

clusters (tree cuts in R terminology) was estimated

visually from the corresponding dendrograms.

3.4.2 Time-to-Intersection Features

Time-to-intersection (tti) features had exceptionally

wide ranges of values. For both subsets of data, the

median was equal to the maximum, which means that

the data was highly skewed. Classical feature-

clipping to a fixed maximum value did not solve this

problem. However, since situations were snapshots,

large time-to-intersection values seemed to be out of

interest then. Therefore, an intuitive border of 20 s

was established, so that all the samples of data with

tti > 20 s or tti_v > 20 s were excluded from further

analysis. The total number of objects with “suitable”

time-to-intersection features was 755: 139

pedestrians and 616 cars. Since reduced times to

intersection had the same range of values for both

pedestrians and cars subsets, the analysis was carried

out in three variations: for each of the subsets and the

whole dataset.

Mean shift and K-Means were also applied to

time-to-intersection features, both on reduced

pedestrians and cars subsets separately and all the

reduced data. For Mean shift clustering, the

bandwidths were estimated with quantiles 0.15, 0.1,

and 0.07 for the pedestrians’ subset, cars’ subset, and

all the data, respectively. For K-Means clustering, the

Elbow Method was used to select the optimal number

of clusters.

Hierarchical agglomerative clustering with

Ward’s linkage was performed. First, a dendrogram

of hierarchical clustering was plotted, from which the

optimal number of clusters was estimated. Then, the

clustering with the selected number of clusters was

performed using AgglomerativeClustering.

3.4.3 Manual Evaluation

Based on both the insights from machine learning

clustering and the empiric account of time-to-

intersection features, a manually elaborated scheme

for grouping the data points was proposed. Generally,

it was considered critical if the time-to-intersection of

the object and the test vehicle had similar values.

Besides, lower values of time-to-intersection features

were more dangerous than the higher ones. The

evaluation scheme is presented in Figure 1.

Figure 1: Manually elaborated scheme for grouping the data

points according to the values of time to intersection

features: Level 1 – the most critical (dangerous), Level 5 –

the least dangerous.

Since the objects were treated as points while they

had certain widths and lengths, and possible

imprecisions by data acquisition, it was decided to

add one more manually evaluated group of dangerous

objects. This group consisted of all objects located in

the 90° range zone in front of the test vehicle (45° to

the right and the left from the vehicle trajectory, or

octants 2 and 3 in the 2D coordinate system, in which

the test vehicle is at the origin and is moving towards

the geographical North) at the distance closer than

7.5 m.

3.4.4 Situation Evaluation

Eventually, the objects were divided into six groups

according to their danger level. Each object then

received a score from zero to five: zero – irrelevant in

the current situation, and five – representing a critical

level of danger in the current situation. To increase

the importance of critical objects, the situation score

was calculated as the sum of the squared scores of all

the objects.

Trafﬁc Data Evaluation for Automated Driving Handover Scenarios

129

4 RESULTS

4.1 Vehicle Data Modality

The results of classification with 10-fold cross-

validation are shown in Figure 2.

Figure 2: Classifier comparison - Mean accuracy.

4.1.1 Random Forest Classifier

Random Forest Classifier performed with the highest

accuracy. Implementation of GridSearchCV

algorithm achieved an accuracy score of 81.83%.

However, the specificity of this model (66.95%, True

Negatives = 640, False Positives = 316) was

considered unacceptably low to predict a safety-

critical handover.

Nevertheless, Random Forest Classifier provided

useful insights into the impurity-based feature

importance, or how much a single feature of the

vehicle data set affected the result (see Figure 3). The

presumed handover decision matrix (Table 1) could

be confirmed with the obtained feature ranking: the

values influencing the motion state of the vehicle

were the most decisive for the accuracy of prediction.

Figure 3: Feature importance revealed by Random Forest

Classifier.

From the bar chart in Figure 3, one can see that

the feature importance values varied strongly.

Furthermore, it is noticeable that the value

direction_of_driving had an unusually high variance

compared to the other features. Such variance

resulted from the fact that the direction_of_driving

feature in 1,609 cases was assigned as unknown,

which made those values incorrect. The position of

the three last-placed values can be explained by the

fact that during the test drives not enough data were

collected where these values were activated.

4.1.2 Multilayer Perceptron (MLP)

Classifier

Since the specificity of the Random Forest Classifier

was regarded as insufficient, it was necessary to

consider another classification algorithm. MLP

Classifier had a lower mean accuracy score than

Gradient Boost Classifier in the overall ranking, but a

lower standard deviation signaled higher robustness

of the former. After applying GridSearchCV method

for two different combinations of parameter grids and

increasing the accuracy by 0.25% compared to the

usage of the standard parameters, a final accuracy

value of 80.31% was achieved. The specificity of this

model reached 77.82% (see Figure 4), which was

14.64%, greater than that obtained with Random

Forest Classifier (see Figure 2).

Figure 4: MLP Classifier confusion matrix.

4.2 Traffic Objects Data Modality

4.2.1 Position and Speed Features

As mentioned above, the relative bearing is a circular

variable and should not be treated linearly. Such

attempts lead to losing the meaning of the values.

Objects from the same cluster corresponded to a wide

range of relative bearing values and could not be

meaningfully interpreted.

Sample visualizations of clusters obtained from

Mean shift (nine clusters), K-Means (five clusters),

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

130

and hierarchical clustering with a circular variable (six

clusters with both methods) applied on speed and

position features suggested that the clusters could be

described as having similar relative bearing features

but not as having certain tendencies in terms of danger

level so that the approach was not applied further.

4.2.2 Time-to-Intersection Features

According to visual analysis of dendrograms, the

optimal number of clusters would be two. Such gross

division, however, did not seem reasonable,

considering the nature of the data. Thus, the number

of clusters was selected as where the dendrogram was

“cut” before going into finer clusters (the same

principle was applied to hierarchical clustering with a

circular variable).

Visualization of the clustering results can be seen

in Figure 5. Cluster evaluation from the empirical

point of view suggested that the proposed clusters

might lack the necessary granularity and symmetry.

Machine learning algorithms provided, however,

useful insights on how the data tended to be grouped

naturally.

Figure 5: Mean shift, K-Means, and hierarchical clustering applied on time to intersection features (data reduced to <= 20 s).

Figure 6: Examples of situations with different danger scores/complexity.

Trafﬁc Data Evaluation for Automated Driving Handover Scenarios

131

4.2.3 Manual Evaluation and Final Situation

Evaluation

Objects in the dangerous proximity zone were

selected first. There were 319 such objects: 27

pedestrians and 292 cars. Of 755 that were to undergo

dangerous level analysis based on time to intersection

features, 62 were excluded because they were already

in the dangerous proximity zone (Danger Level 1). In

total, 1,012 objects were considered relevant in terms

of danger. These objects corresponded to 552

situations. The number of participants per situation

ranged from 1 to 10, M = 1.83, SD = 1.52. Excluding

situations with zero danger, situation scores ranged

from 1 to 143, M = 23.13, SD = 18.76. In 1,316

situations (70.4% of all the obtained data) there were

only objects with a danger Level of 0. Symbolic

representations of situations with different danger

scores/complexity can be seen in Figure 6.

5 CONCLUSIONS

We addressed the problem of a handover from

automated to human control through the multimodal

description and analysis of traffic situations. We have

focused on two modalities, namely own vehicle state

and detected traffic objects.

Own vehicle state can be directly mapped on one

of the two labels, “handover possible” or “handover

not possible”. These binary labels assigned by a

human rater have been used as ground truth for

training a range of classification models. The mean

accuracy of algorithms ranges from 65.7% to 81.4%,

the highest accuracy obtained with Random Forest

Classifier. Optimization of the algorithm allows to

improve the accuracy to 81.8%, However, False

negatives (algorithm attributes the handover as “not

possible” while in truth it is possible) do not seem to

have the same impact as False positives so the trade-

off should be made towards higher specificity. The

specificity of the Random Forest Classifier model is

67%, which is unacceptably low. With a slightly

lower overall accuracy of 80.3%, MLP Classifier

provides a significantly higher specificity of 77.8%,

which is nevertheless still unacceptable for robust

prediction of the handover. Balancing the dataset in

terms of outcome labels and feature weighting is seen

as the way to improve the performance of machine

learning models.

Traffic objects can be described in terms of their

danger level, which cumulatively corresponds to the

traffic situation complexity/danger score. Different

combinations of features have been explored with the

help of centroid-based and hierarchical clustering.

When features include a circular variable (relative

bearing), this feature seems to become dominant,

while the others are not interpreted by algorithms in a

way that could be explained in real life. Clustering on

time-to-intersection features with machine learning

algorithms does not result in desired granularity and

cluster symmetry, therefore a manual approach was

selected for grouping the data points. Almost 30% of

the investigated situations have been evaluated as

having a complexity/danger score higher than 0, with

maximum complexity of 143.

The possibility of the control handover is

determined via a cascade approach. First, own state

of the vehicle is assessed. In the current work, vehicle

data are available at any time, whereas the data of

detected objects (traffic situation data) are only

available at certain intersections. When the own state

of the vehicle allows the handover, the complexity of

the situation based on traffic objects is evaluated. The

higher the complexity, the more critical the

requirement of sufficient transition time, making an

immediate control handover not possible. Predicting

situation complexity in time, setting thresholds for

handover, and issuing corresponding warnings are the

topics for further research.

Limitations of the current study include the

absence of information on the driver’s state and

attention, and uniformity of weather conditions, as

well as the lack of perceptive capability of the

onboard vehicle sensors for obstacles on the road.

Adding these modalities and exploring the decision-

making with fusion at different levels (e.g., feature

fusion and modality fusion) are planned as the

following steps. Bringing data analysis into real-time,

while the vehicle is performing test routes, and

exploring delays in data processing and the ways of

minimizing them are also seen as one of the future

research directions.

ACKNOWLEDGMENTS

This work was funded by the German Federal

Ministry of Transport and Digital Infrastructure

within the project kantSaar (grant number

16AVF2129). The project consortium consists of the

University of Saarland and the University of Applied

Sciences Saarland – htw saar. The project approach

and the partial project outcomes result from the

collaborative work of the entire kantSaar project team

and other FGVT htw saar members.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

132

REFERENCES

Ayoub, J., Du, N., Yang, X. J. and Zhou F., Predicting

Driver Takeover Time in Conditionally Automated

Driving. (2022) in IEEE Transactions on Intelligent

Transportation Systems, vol. 23, no. 7, pp. 9580-9589,

DOI:10.1109/TITS.2022.3154329.

Awad, M., and Khana, R. (2015). Efficient learning

machines. Apress, Berkeley, CA, USA. DOI:

10.1007/978-1-4302-5990-9

Becker, F.-M., Boortz, G., Dietrich, V., Engelmann. L.,

Ernst, C., Fanghängel, G., … Höhne, H. (1999).

Formeln und Tabellen für die Sekundarstufen I und II

[Formulae and Tables for Secondary Education Levels

I and II]. Edition 7, paetec Gesellschaft für Bildung und

Technik mbH, Berlin.

Borojeni, S., Meschtscherjakov, A., Mirnig, A., Boll, S.

Naujoks, F., Politis, I., and Alverez, I. (2017). Control

Transition Workshop: Handover and Takeover

Procedures in Highly Automated Driving. In

Proceedings of the 9th International Conference on

Automotive User Interfaces and Interactive Vehicular

Applications Adjunct (AutomotiveUI '17). Association

for Computing Machinery, New York, NY, USA, 39–

46. DOI:10.1145/3131726.3131732

Braunagel, C., Rosenstiel, W., and Kasneci, E., Ready for

Take-Over? A New Driver Assistance System for an

Automated Classification of Driver Take-Over

Readiness. (2017). IEEE Intelligent Transportation

Systems Magazine, vol. 9, no. 4, pp. 10-22, DOI:

10.1109/MITS.2017.2743165.

Clark, J.R., Stanton, N.A. and Revell, K.M.A. Automated

Vehicle Handover Interface Design: Focus Groups

with Learner, Intermediate and Advanced Drivers.

Automot. Innov. 3, 14–29 (2020). DOI:10.

1007/s42154-019-00085-x

Dray, S., and Dufour, A.-B. (2007). The ade4 Package:

Implementing the Duality Diagram for Ecologists.

Journal of Statistical Software, 22(4), 1-20.

DOI:10.18637/jss.v022.i04

Ellis, D. (May 19, 2020). Calculating the bearing between

two geospatial coordinates. Available from https://

towardsdatascience.com/calculating-the-bearing-betwe

en-two-geospatial-coordinates-66203f57e4b4

Eriksson, A., and Stanton, N. A. (2017). Takeover Time in

Highly Automated Vehicles: Noncritical Transitions to

and From Manual Control. Human Factors, 59(4),

689–705. https://doi.org/10.1177/0018720816685832

ETSI – European Telecommunications Standards Institute.

(2019a). Intelligent Transport Systems (ITS); Vehicular

Communications; Basic Set of Applications; Analysis of

the Collective Perception Service (CPS) (ETSI TR 103

562). Release 2, Sophia Antipolis Cedex, France.

ETSI – European Telecommunications Standards Institute.

(2019b). Intelligent Transport Systems (ITS); Vehicular

Communications; Basic Set of Applications; Facilities

layer protocols and communication requirements for

infrastructure services (ETSI TS 103 301). Sophia

Antipolis Cedex, France.

ETSI – European Telecommunications Standards Institute.

(2019c). Intelligent Transport Systems (ITS); Vehicular

Communications; Basic Set of Applications; Part 2:

Specification of Cooperative Awareness Basic Service

(ETSI EN 302 637-2). Sophia Antipolis Cedex, France.

ETSI – European Telecommunications Standards Institute.

(2019d). Intelligent Transport Systems (ITS); Vehicular

Communications; Basic Set of Applications; Part 3:

Specification of Decentralized Environmental

Notification Basic Service (ETSI EN 302 637-3).

Sophia Antipolis Cedex, France.

Gillies, S., Bierbaum, A., Lautaportti, K., and

Tonnhofer, O. (2007). Shapely: manipulation and

analysis of geometric objects. Available from

https://github.com/Toblerity/Shapely

Gold, C., Körber, M., Lechner, D., and Bengler, K. (2016).

Taking Over Control From Highly Automated Vehicles

in Complex Traffic Situations: The Role of Traffic

Density. Human Factors, 58(4), 642–652.

DOI:10.1177/0018720816634226

Harris, C. R., Millman, K. J., van der Walt, S. J.,

Gommers, R., Virtanen, P., Cournapeau, D., …

Oliphant, T. E. (2020). Array programming with

NumPy. Nature 585, 357–362. DOI:10.1038/s41586-

020-2649-2

Honda. (2020, November 11). Honda Receives Type

Designation for Level 3 Automated Driving in Japan.

Available from https://global.honda/newsroom/

news/2020/4201111eng.html

Hunter, J. D. (2007). Matplotlib: A 2D Graphics

Environment. Computing in Science & Engineering, 9,

90-95. DOI:10.1109/MCSE.2007.55

ISO – International Organization for Standardization.

(2015). Road vehicles — Controller area network

(CAN) — Part 1: Data link layer and physical signaling

(ISO 11898-1:2015). Edition 2, Geneva, Switzerland.

IVAO – International Virtual Aviation Organisation.

(2020). Introduction to navigation. Available from

https://mediawiki.ivao.aero/index.php?title=Introducti

on_to_navigation

Jiménez, F., Naranjo, J. E., and García, F. (2013). An

Improved Method to Calculate the Time-to-Collision of

Two Vehicles. International Journal of Intelligent

Transportation Systems Research, 11(1), 34-42.

Kerautret, L., Dabic, S., and Navarro J., (2023).

Exploration of driver stress when resuming control

from highly automated driving in an emergency

situation, Transportation Research Part F: Traffic

Psychology and Behaviour, Volume 93, Pages 222-234,

ISSN 1369-8478, DOI:10.1016/j.trf.2023.01.016.

Maechler, M., Rousseeuw, P., Struyf, A., Hubert, M., and

Hornik, K. (2019). cluster: Cluster Analysis Basics and

Extensions. R package version 2.1.0.

Murtagh, F., & Legendre, P. (2014). Ward's hierarchical

agglomerative clustering method: which algorithms

implement Ward's criterion? Journal of Classification,

31, 274–295. DOI:10.1007/s00357-014-9161-z.

Otte, A., Staub, J., Vogt, J., and Wieker, H. (2021). Cloud-

based traffic data fusion for situation evaluation of

handover scenarios. ArXiv abs/2101.10912.

Trafﬁc Data Evaluation for Automated Driving Handover Scenarios

133

Pavoine, S., Vallet, J., Dufour, A.-B., Gachet, S., and

Daniel, H. (2009). On the challenge of treating various

types of variables: Application for improving the

measurement of functional diversity. Oikos, 118, 391–

402. DOI: 10.1111/j.1600-0706.2008.16668.x

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thition, B., Grisel, O., … Duchesnay, E. (2011). Scikit-

learn: Machine Learning in Python. Journal of Machine

Learning Research, 12, 2825-2830.

R Core Team. (2020). R: A language and environment for

statistical computing. R Foundation for Statistical

Computing, Vienna, Austria. Available from

https://www.R-project.org/.

Radlmayr, J., Gold, C., Lorenz, L., Farid, M., and Bengler,

K. (2014). How Traffic Situations and Non-Driving

Related Tasks Affect the Take-Over Quality in Highly

Automated Driving. Proceedings of the Human Factors

and Ergonomics Society Annual Meeting, 58(1), 2063–

2067. DOI:10.1177/1541931214581434

SAE International. (2018). SAE J3016 Levels of Driving

Automation, Standard.

Schroeter, R., and Steinberger, F. (2016). Pokémon DRIVE:

towards increased situational awareness in semi-

automated driving. In Proceedings of the 28th

Australian Conference on Computer-Human

Interaction (OzCHI '16). Association for Computing

Machinery, New York, NY, USA, 25–29.

DOI:10.1145/3010915.3010973

Snow, G. (2020). TeachingDemos: Demonstrations for

Teaching and Learning. R package version 2.12.

Available from https://CRAN.R-project.org/pack

age=TeachingDemos

Trimble, T., Bishop, R., Morgan, J. F., and Blanco, M.

(2014). Human factors evaluation of level 2 and level 3

automated driving concepts: Past research, state of

automation technology, and emerging system concepts.

(Report No. DOT HS 812 043). Washington, DC:

National Highway Traffic Safety Administration.

Van Rossum, G. (2020). The Python Library Reference,

release 3.8.2. Python Software Foundation.

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M.,

Reddy, T., Cournapeau, D., … SciPy 1.0 Contributors.

(2020). SciPy 1.0: Fundamental Algorithms for

Scientific Computing in Python. Nature Methods,

17(3), 261-272. DOI:10.1038/s41592-019-0686-2

Will, G. (2016). Visualizing and Clustering Data that

Includes Circular Variables. Master’s Thesis, Montana

State University, Bozeman, USA. Available from

https://math.montana.edu/grad_students/writing-

projects/2016/16will.pdf.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

134