Identifying Scenarios in Field Data to Enable Validation of Highly

Automated Driving Systems

Christian Reichenb

acher

1,2 a

, Maximilian Rasch

1,3 b

, Zafer Kayatas

1,3 c

, Florian Wirthm

uller

1,4 d

Jochen Hipp

1 e

, Thao Dang

5 f

and Oliver Bringmann

2 g

Mercedes-Benz AG, Sindelﬁngen, Germany

Department of Computer Science, University of T

ubingen, T

ubingen, Germany

Institute of Technical Mechanics and Vehicle Dynamics, Brandenburg University of Technology, Cottbus, Germany

Institute of Databases and Information Systems (DBIS), Ulm University, Ulm, Germany

Faculty Computer Science and Engineering, Esslingen University of Applied Sciences, Esslingen, Germany

Keywords:

Autonomous Vehicles and Automated Driving, Analytics for Intelligent Transportation, Trafﬁc and Vehicle

Data Collection and Processing, Vehicle Environment Perception, Pattern Recognition for Vehicles.

Abstract:

Scenario-based approaches for the validation of highly automated driving functions are based on the search for

safety-critical characteristics of driving scenarios using software-in-the-loop simulations. This search requires

information about the shape and probability of scenarios in real-world trafﬁc. The scope of this work is to

develop a method that identiﬁes predeﬁned logical driving scenarios in ﬁeld data, so that this information can

be derived subsequently. More precisely, a suitable approach is developed, implemented and validated using

a trafﬁc scenario as an example. The presented methodology is based on qualitative modelling of scenarios,

which can be detected in abstracted ﬁeld data. The abstraction is achieved by using universal elements of an

ontology represented by a domain model. Already published approaches for such an abstraction are discussed

and concretised with regard to the given application. By examining a ﬁrst set of test data, it is shown that the

developed method is a suitable approach for the identiﬁcation of further driving scenarios.

1 INTRODUCTION

The development of highly automated driving is gain-

ing more and more momentum and is about to ﬁnd

its way into production vehicles soon. However, this

technology raises new challenges for the system val-

idation required for series approval. The increased

interconnection of driving functions and the replace-

ment of human drivers as a control instance expands

the range of requirements and test cases to be vali-

dated and veriﬁed. Distance-based methods proving

adequate system safety that have been used so far, are

https://orcid.org/0000-0002-0907-3287

https://orcid.org/0000-0002-7554-2619

https://orcid.org/0000-0003-0880-0304

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

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

https://orcid.org/0000-0001-5505-8953

https://orcid.org/0000-0002-1615-507X

Figure 1: Trajectory and simulation parameters for the sce-

nario cut-in.

too costly from an economic point of view. Maurer et

al. (2015, p. 458) conﬁrm this fact by determining the

test kilometers to be completed for automated driving

on the highway in order to prove that the system is

at least twice as safe as a human driver, with at least

134

Reichenbächer, C., Rasch, M., Kayatas, Z., Wirthmüller, F., Hipp, J., Dang, T. and Bringmann, O.

Identifying Scenarios in Field Data to Enable Validation of Highly Automated Driving Systems.

DOI: 10.5220/0011081500003191

In Proceedings of the 8th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2022), pages 134-142

ISBN: 978-989-758-573-9; ISSN: 2184-495X

6.62 billion kilometers.

A new and uniform validation concept was there-

fore developed as part of the joint project PEGA-

SUS (see pegasusprojekt.de/en). The project was con-

ducted in cooperation with automotive manufactur-

ers, suppliers, research institutions and the German

Federal Ministry for Economic Affairs and Energy.

The developed concept enables to proof the safety of

highly automated driving systems with an economi-

cally justiﬁable effort. The methodology derived from

this concept for the highway domain in (Rasch et al.,

2019) consists of a multi-stage process. In this pro-

cess, the space of test cases is searched for driving

scenarios such as a close cut-in, the end of a traf-

ﬁc jam or other scenarios that are particularly safety-

critical. The search for such scenarios is carried out

with the help of software-in-the-loop simulations. Ini-

tially, common scenarios in the highway domain are

logically described on the basis of parameters. These

scenarios are among others: following a car in the

same lane, a cut-in or a cut-out in front of a vehicle.

Logical scenarios qualitatively describe the basic

behaviour of vehicles (King et al., 2017). The latter

can occur in different varieties in reality, for exam-

ple at different speeds or distances between the road

users involved. Correspondingly quantitative descrip-

tions are called concrete scenarios. A logical scenario

thus comprises concrete scenarios with different char-

acteristics of the same basic behaviour of road users.

Figure 1 shows the logical scenario cut-in and vi-

sualises some of the simulation parameters. The log-

ical scenario is deﬁned as the description of all sce-

narios in which a vehicle, here called cut-in vehicle,

changes from a lane next to the ego vehicle to the lat-

ter’s lane. The ego vehicle, shown in red in Fig. 1,

is the test object with the highly automated driving

functions to be validated. Since its system behaviour

is to be examined, its position and state of movement

are not speciﬁed or parameterised.

By varying the parameters in individual simu-

lation runs, the parameter space of test cases can

be searched for critical scenarios in which the sys-

tem fails. However, even the low-cost software-in-

the-loop simulation of all possible variants compared

to hardware-in-the-loop simulation, test site or ﬁeld

drives does not make sense from an economic point of

view. Thus, those scenario parameterisations should

preferably be simulated, which have a sufﬁciently

high probability of occurrence in real-world trafﬁc.

In order to determine the probabilities of various sce-

narios, information about realistic parameter distribu-

tions are required.

In this paper a procedure to identify parameterised

scenarios in ﬁeld data is developed. It enables to ex-

tract parameter distributions in future. In the course

of this, an approach is developed that allows to reduce

the ﬁeld data, describing driving condition and vehi-

cle environment, recorded by the vehicle sensors in

an object-oriented manner, to relevant variables. This

approach is implemented and validated using the driv-

ing scenario Cut-in as an example. In the long term,

the concept of this procedure will enable the identiﬁ-

cation of an entire selection of highway scenarios in

ﬁeld data.

The remainder of this work is structured as fol-

lows: In Section 2, the current state of research on the

identiﬁcation of driving scenarios is presented. In par-

ticular, existing approaches for the abstraction of sce-

narios and measurement data by means of metamod-

elling are discussed. Section 3 explains the approach

chosen for the identiﬁcation of scenarios in the con-

text of this work, namely metamodelling with pattern

recognition. In detail, deﬁnitions are speciﬁed and

an ontology represented by a domain model is intro-

duced for the metamodelling. Subsequently, the logic

of the developed procedure for the abstraction of mea-

surement data and the identiﬁcation of driving sce-

narios is described. Section 4 presents the procedure

for validating the implemented method and the results

obtained. In addition, the results and the methodology

of the identiﬁcation procedure are evaluated. Section

5 summarises the core results of the work, highlights

their signiﬁcance and provides a brief outlook on fur-

ther research.

2 RELATED WORK

The identiﬁcation of logical driving scenarios in ﬁeld

data has already been subject to various research.

King et al. (2017) present an approach for the iden-

tiﬁcation of individual driving maneuvers and logi-

cal scenarios in a - in contrast to this work - virtual

test drive. According to the authors, driving maneu-

vers or complex interactions between road users can

only be read with difﬁculty from just the speed and

position information of the individual vehicles. For

this reason, they consider a reduction and abstraction

of the recorded data for interpretation to be manda-

tory. In this context, they present their approach of

using knowledge-based metamodels of scenarios for

pattern recognition. For the modelling of the sce-

narios, they propose to use concept of Bach, Otten,

and Sax (2016), which enables an abstraction of real-

world driving scenarios down to the logical level.

The concept is based on a domain model for map-

ping the relevant classes of a driving scenario. Bach,

Otten, and Sax (2016) use terminology from the ﬁeld

Identifying Scenarios in Field Data to Enable Validation of Highly Automated Driving Systems

135

of theater and ﬁlm and refer to Geyer et al. (2014).

Bach, Otten, and Sax (2016) consider a scenario as

a speciﬁc timespan with a postulated sequence of

events. According to the authors, the principle of

metamodelling lies in the abstraction of all relevant

elements within a logical scenario. The term partici-

pant is used to describe all dynamic elements within

a scenario that can interact with one another. In order

to specify the state and action of the participants on

an abstract level, Bach, Otten, and Sax (2016) refer

to a selection of maneuvers. By assigning a speciﬁc

sequence of maneuvers, the behaviour of the partic-

ipants can be modeled within a scenario. A tempo-

ral abstraction occurs through so-called acts. Individ-

ual acts are distinguished from each other by the ma-

neuvers of the road users. Accordingly, a scenario is

deﬁned as a sequence of speciﬁc acts that differ pre-

cisely in one maneuver by the participants.

For identifying logical scenarios, the StreetWise

methodology relies on the use of knowledge-based

metamodels for pattern recognition as well (Elrofai

et al., 2018). With the StreetWise methodology the

Dutch Organization for Applied Scientiﬁc Research

aims to provide a data-based method for the realistic

generation of test cases in order to enable the valida-

tion and development of automated driving functions.

For the metamodelling of scenarios, Elrofai et al.

(2018) revised ontologies from Geyer et al. (2014),

Ulbrich et al. (2015), and Hala Elrofai, Worm, and Op

den Camp (2016), further speciﬁed them and created

a domain model that is supposed to meet the practical

requirements of scenario identiﬁcation. Elrofai et al.

(2018) speak of activities instead of manoeuvres and

concretises their deﬁnition as the temporal change of

status variables such as speed and orientation in order

to describe, for example, a lane change or braking to

a standstill. Jan-Pieter Paardekooper et al. (2019) dis-

tinguish between possible longitudinal activities with

accelerating, cruising and decelerating as well as lat-

eral activities with lane keeping, lane change to the

left and lane change to the right. If at least two of

the six activities (longitudinal and lateral) are always

assigned to every road user, it should be possible to

describe every possible movement trajectory.

Elrofai et al. (2018) claim the StreetWise pattern

recognition algorithm is supposed to use artiﬁcial in-

telligence methods such as machine learning. Jan-

Pieter Paardekooper et al. (2019) adds that a method

of ”template matching on a graphical network” is

used.

For the description of logical scenarios, de Gelder

et al. (2020) introduce so-called scenario categories

in combination with element categories for activities,

static and dynamic environment. To specify a sce-

nario category, the authors suggest to reference the el-

ements that the scenario contains, for example, actors,

activities, static environment and events. By doing so,

these elements can be used for various scenarios. An

activity is described as the temporal change in one or

more status variables of an actor between two events.

Like Jan-Pieter Paardekooper et al. (2019), de Gelder

et al. (2020) suggest a distinction between longitudi-

nal and lateral activities. The latter deﬁnes an event

as a point in time at which either a deﬁned threshold

value is reached or a mode change takes place through

which the change in one or more status variables is

described by a different behaviour respectively a dif-

ferent activity.

In summary, a method for pattern recognition and

its application to real-world data has not yet been de-

scribed in detail in the research work presented. In

this aspect, our work differs from the previous ones.

3 METHODS

The approach chosen in this work for the identiﬁca-

tion of driving scenarios builds on knowledge-based

metamodelling of scenarios in connection with pat-

tern recognition as suggested by King et al. (2017),

Elrofai et al. (2018) and Jan-Pieter Paardekooper et al.

(2019) (see Section 2). In order to be able to clearly

identify and differentiate between logical driving sce-

narios, an equally clear deﬁnition of the latter is nec-

essary. Knowledge-based metamodels are a suitable

means for such deﬁnitions.

The explicit modelling of the scenarios sought has

the advantage that their recognition in ﬁeld data can

be derived directly and is also fully traceable com-

pared to the use of, for example, neural networks.

Furthermore, in contrast to machine learning based

approaches, large amounts of training data are not

necessary for the implementation.

3.1 Nomenclature and Deﬁnitions for

Driving Scenarios

The term scenario is used in a wide variety of areas,

which is why there is no generally valid, but a large

number of different deﬁnitions for this term. The def-

inition of a logical scenario and related terms in this

work largely follow the ones presented in Section 2,

but are further speciﬁed or simpliﬁed with regard to

the present application. This Section summarises the

concepts and terms scenario, ego vehicle, actor, act,

activity and event used in this work:

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

136

Scenario. In general, a scenario is deﬁned as a pos-

tulated sequence of events with certain activities and

properties of the ego vehicle and its static and dy-

namic environment.

Ego Vehicle. The ego vehicle relates to the perspec-

tive from which the world is seen. More precisely, the

ego vehicle refers to the vehicle that perceives the en-

vironment through the built-in sensors. This includes

the driver and optional automated driving functions.

Actor. Actors are all dynamic elements of a sce-

nario that are able to interact with each other. The ego

vehicle is also assigned to the category actor, as it has

the same characteristics as other vehicles. Here, the

term actor does not presuppose an actual movement or

interaction of the elements, but merely requires their

ability to do so in principle.

Act. The deﬁnition of acts enables a temporal ab-

straction of the driving events. In this way, the course

of a scenario can be divided into several acts. The in-

dividual acts represent time periods of the scenario in

which the actors involved perform speciﬁed activities.

For the concept applied in this work, further condi-

tions, for example on distance and relative speed of

the actors, can optionally be included in the deﬁnition

of an act.

Activity. An activity describes the change of state

of an actor in a qualitative manner. E. g., for the iden-

tiﬁcation of the scenario cut-in a purely lateral dis-

tinction of the activities is sufﬁcient: lane change to

the left, lane change to the right and lane following.

This is the case as the longitudinal behaviour of the

vehicles plays no role in the logical description of the

scenario.

Event. An event marks a point in time when one or

more threshold values are reached, exceeded or un-

dershot. Accordingly, a set of one or more conditions

is speciﬁed for each event. As soon as such a set of

conditions is fulﬁlled, the corresponding event is said

to have occurred.

3.2 Ontology for Driving Scenarios

In this Section the developed ontology for the identi-

ﬁcation of driving scenarios in ﬁeld data is presented.

As in (de Gelder et al., 2020), the ontology intro-

duced in this work is formally represented by a do-

main model, which is brieﬂy explained in the fol-

lowing. Afterwards, it is exemplarily shown how the

Figure 2: UML class diagram of the ontology for scenarios.

driving scenario cut-in is deﬁned using this domain

model.

Domain Model. The classes of the domain model

that are used to deﬁne a logical scenario are shown

in Fig. 2. The representation is based on the Uniﬁed

Modeling Language. Each of the blue blocks repre-

sents a class. The relationships between these differ-

ent classes are described by the arrows. An arrow with

a diamond symbol can be described as “is part of”. A

“N” at the beginning of the arrow means that zero,

one or more objects are part of an object of the class

at the end of the arrow.

The ego vehicle and its dynamic environment are

qualitatively described by objects of the classes activ-

ity, actor and event. One attribute of an object of the

class scenario is the chronological list of acts. The

acts describe which actor carries out which activity as

well as which event conditions apply. It is possible

that an actor carries out different activities and that an

activity is carried out by different actors at the same

time.

Scenario Cut-in and Its Attributes. In the follow-

ing, the beneﬁts of the ontology are illustrated using

the domain model for the deﬁnition of the cut-in sce-

nario. To describe the scenario, it is necessary to de-

ﬁne two events. The ﬁrst one is the event follow-

ing. The latter’s conditions are met as soon as the

ego vehicle follows a vehicle in the same lane within

a speciﬁed maximum distance and within a speciﬁed

maximum relative speed. The threshold value for the

distance can be speciﬁed as a function of the driving

speed of the ego vehicle. In general, a higher driv-

ing speed also requires interaction with vehicles that

are further away. In addition, a threshold value that

is independent of the driving speed can be deﬁned, at

which the distance condition of the subsequent driv-

ing event is met in every case. This ensures that the

following event can be detected even when trafﬁc is

Identifying Scenarios in Field Data to Enable Validation of Highly Automated Driving Systems

137

Figure 3: Objects for deﬁning the scenario cut-in.

moving slowly. The second required event is called

driving parallel. Its conditions are met as soon as a

selected vehicle drives in one of the adjacent lanes in

the detection area in front of or next to the ego vehi-

cle.

To describe the cut-in scenario according to the

presented domain model, objects are instantiated

from the classes shown in Fig. 2. Figure 3 shows the

objects for the qualitative description of the scenario.

The ﬁrst line contains the object name in italics and

after the colon the name of the class from which the

object was instantiated.

3.3 Identiﬁcation of Driving Scenarios

Objects in the measurement data are classiﬁed accord-

ing to their relative position to the ego vehicle. A dis-

tinction is made between ﬁrst-left, ﬁrst-right, second-

left, second-right, ﬁrst-ego and second-ego. Left in

the designation means that the object is driving in the

lane that is adjacent to the one of the ego vehicle on

the left (equivalent for right). Ego means that the ve-

hicle drives in the same lane as the ego vehicle. First

means that the vehicle is the ﬁrst object detected in

the respective lane in the direction of travel, Second

is the vehicle that drives in front of the First.

For the identiﬁcation of driving scenarios, an ab-

stract description of ﬁeld data is derived. The abstrac-

tion is based on the ontology explained in Section 3.2

with actors, their activities and events. By automat-

ically comparing the measurement drives abstracted

in this form with the scenarios previously described

using the same elements, any concurrences can be

recognised and the scenarios sought can thus be iden-

tiﬁed.

Abstraction of Measurement Data. The identiﬁ-

cation of different events forms the basis for deriv-

ing the abstract description of a measurement drive

from the ﬁeld data. A distinction is made between

two different types of events. On the one hand, there

are the events according to the deﬁnition in Section

3.1. On the other hand, there are the events that mark

the points in time at which a status change takes place.

The latter are only necessary for the practical deriva-

tion of the abstract description of a test drive and are

not part of the deﬁnition of scenarios.

The events that mark a threshold value violation

include the events following and driving parallel. The

basis for the identiﬁcation of the lane change activi-

ties of the ego vehicle and the other vehicles in its dy-

namic environment are formed by the lead change and

ego lane change events. The latter mark times when

a status change takes place. The lead change event

marks the point in time at which the signal for the ob-

ject ID of the ﬁrst-ego vehicle changes. Accordingly,

this also includes the times at which a vehicle driving

directly ahead of the ego vehicle leaves the detection

area or changes lanes to one of the adjacent lanes and

there is no new driver ahead in the lane of the ego ve-

hicle. The ego lane change event marks the point in

time at which the ego lane assignment changes. Both

events - lead change and ego lane change - occur in

the case of a lane change when the corresponding ve-

hicle is in the middle between two lanes.

Amongst others, possible activities to be identi-

ﬁed for the derivation of an abstract description of

the measurement drive are lane changes of the ego

vehicle, lane keeping of the ego vehicle, cut-in lane

change activities in front of the ego vehicle, lane

keeping of a vehicle that later on cuts into the lane of

the ego vehicle and lane keeping of a driver directly

in front in the lane of the ego vehicle.

In principle, the following variants are conceiv-

able for all lane change activities: the vehicle changes

to one of the adjacent lanes, the vehicle changes

over more than one lane or the vehicle only brieﬂy

switches to one of the adjacent lanes and then

switches back again. In the given context, when iden-

tifying the time spans in which the relevant actors

perform a lane change activity, a distinction is made

between valid and invalid lane changes according to

the conformity for the scenarios to be identiﬁed. If

a vehicle changes from a common lateral position

of one lane to a common lane position of another

lane, a valid lane change activity is recognised for the

time span in which the vehicle moves between these

two positions. The change over more than one lane

is also included, since a common lateral position on

one of the originally adjacent lanes is also necessar-

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

138

Figure 4: Scheme for pattern recognition.

ily crossed. In the context of this work, a common

lateral lane position is given when the center of a ve-

hicle is more than 35 % of the lane width away from

the lane markings. If a vehicle only slightly overlaps

one of the adjacent lane markings and consequently

does not reach a common lateral lane position before

it changes back to the original lane, an invalid lane

change activity is recognised.

The activity lane keeping is recognised for all time

spans in which the relevant actors do not carry out any

valid or invalid lane change activity. In general, activ-

ities of the actors in the dynamic environment of the

ego vehicle are only registered while the ego vehicle

is performing the activity lane keeping. This is sufﬁ-

cient, since the continuous lane keeping activity of the

ego vehicle is required for all scenarios to be identi-

ﬁed.

Pattern Recognition. For the identiﬁcation of the

scenarios sought, their descriptions according to the

domain model are compared with the abstract descrip-

tions derived from the ﬁeld data. This comparison can

always be carried out using the same procedure, re-

gardless of the scenario (c. f. Fig. 4).

For identiﬁcation, the derived activities and events

are checked chronologically for each point in time

of the measurement drive. The default status of the

check is free driving - this means that no potential sce-

nario has been identiﬁed. If the activities and events

speciﬁed for the ﬁrst act of a scenario are active at

a point in time during this check, a potential sce-

nario with the status of the ﬁrst act is recognized. If

these conditions for events and activities are violated

in the further temporal iteration, the potentially rec-

ognized scenario is discarded if the changed activities

and events do not correspond to those of the second

act. In the latter case, the status is set to the second

act. The same approach is used for the other acts of

the scenario sought.

If a potentially recognized scenario is discarded

during the iterations because the conditions for the

current status and the subsequent act of a potentially

recognized scenario are violated, the status free driv-

ing is set again. In the subsequent time steps, the

conditions for events and activities of the ﬁrst act are

checked again.

If a potentially recognized scenario has the status

of its last deﬁned act (in Fig. 4 act N), the scenario is

noted as identiﬁed after a period of time speciﬁed for

this act from the beginning of the ﬁrst act to this point

in time. If a maximum period of time has also been

speciﬁed for the ﬁrst act, the beginning of the scenario

is set accordingly after the point in time at which the

conditions of the ﬁrst act were met for the ﬁrst time.

4 RESULTS AND DISCUSSION

The exemplary implementation of the developed pro-

cedure for identifying driving scenarios was car-

ried out in the object-oriented programming language

Python. For its validation, a selection of measurement

ﬁles, spanning one minute each, from test drives on

highways was available together with video record-

ings. The aim of the validation was to check all rele-

vant sub-functions and aspects for the identiﬁcation of

each potential scenario element for the scenario cut-

in. For this purpose, a test catalogue was prepared,

which speciﬁes corresponding test cases for the iden-

tiﬁcation of each event and activity. Suitable mea-

surement ﬁles were assigned to these test cases by re-

viewing the video material. To carry out the valida-

tion, the measurement drive sections speciﬁed for the

test cases were evaluated one after another using the

implemented procedure. Afterwards, the evaluation

was manually checked for false-positives and false-

negatives using the video material according to the

test speciﬁcation.

All tests were performed for straight road sections.

Since ﬁrst trials have shown that vehicles overtaking

on the outside of curves are sometimes mistakenly de-

tected as ﬁrst-ego and thus lead change events are de-

tected by mistake. This is caused as the sensor sys-

tem of the ego vehicle in our setting does not use the

course of the road to classify the detected objects, but

only their lateral offset to the ego vehicle. Similarly,

it is not possible to determine the position of the de-

tected objects in their lanes, since this type of infor-

mation is not required for the function of the produc-

Identifying Scenarios in Field Data to Enable Validation of Highly Automated Driving Systems

139

tion vehicles.

Apart from such problems, which can be traced

back to insufﬁcient coverage of the vehicle environ-

ment, these inital investigations delivered a promising

result. They show, a suitable approach for the iden-

tiﬁcation of driving scenarios in ﬁeld data could be

elaborated, implemented and validated. The devel-

oped methodology with an universal ontology allows

to reduce a large amount of measurement data to the

relevant variables for scenario identiﬁcation.

4.1 Event Identiﬁcation

The identiﬁcation of events worked without errors for

the tests carried out on straight sections. The valida-

tion included both the events that mark points in time

when one or more thresholds are violated, such as

driving parallel and following, but also the events that

mark times when a status change takes place, such as

lead change and ego lane change.

The partially incorrect detection of lead change

events in curves is not an error in the methodology.

The requirement for the overall system is that the role

assignment of the sensor system for the detected ve-

hicles is correct. Thus, an object declared as ﬁrst-ego

should actually drive ahead of the ego vehicle in the

same lane. It is therefore obvious to directly address

the cause in the sensor technology for the problem of

false event identiﬁcation. In addition to the relative

position to the ego vehicle, the object detection layer

should also take into account the position of the de-

tected objects in relation to the lane markings. How-

ever, an optimisation to this effect has not been carried

out at the time of publication of this work. If it should

turn out that an adaptation is not possible, further con-

siderations should be made how this problem can be

circumvented and how the object classiﬁcation can be

corrected.

4.2 Activity Identiﬁcation

During the identiﬁcation tests, it has been shown that

both valid and invalid lane changes of the ego vehicle

are detected correctly. However, a lane change that

is not carried out into the lane centre area in front of

the ego vehicle is incorrectly recognised as valid. The

reason for this is that lane-related position informa-

tion, for a procedure as described in Section 3.3, was

only available for the ego vehicle in the used measure-

ment data.

Vehicles driving ahead of the ego vehicle already

have a high lateral offset to the ego vehicle when the

road is slightly curved. This effect increases with

increasing distance between the vehicles. Informa-

tion on the course of the road and lane markings with

which this effect could be calculated was not available

for this work.

Instead, a simpliﬁed identiﬁcation for lane change

activities was implemented. Here, the time spans of

all lane changes, independent of the actual duration,

are determined by means of constant time intervals to

the lead change event. Consequently, a lane change

without reaching the lane centre area followed by a

change back is falsely identiﬁed as valid. Similarly,

the identiﬁed time span of the lane change does not

always match the actual duration. Speciﬁcally, the

described investigations were carried out with a du-

ration of one second before and after the correspond-

ing lane change event. The evaluation of video ma-

terial showed that this duration mostly satisfactorily

describes the real driving event for the application.

Despite this problem, the correct identiﬁcation of

ego vehicle lane changes shows that the developed

method for identifying lane changes works. In prin-

ciple, the implementation for the ego vehicle can be

transferred to other vehicles equivalently, provided

that required position signals are available in the fu-

ture. In terms of the requirements of this work, this is

a satisfactory result.

4.3 Scenario Identiﬁcation

In the examinations carried out, the correct identiﬁca-

tion of driving scenarios was conﬁrmed by the manual

comparison with the identiﬁed activities and events.

Speciﬁcally, the timing and overlap of these activities

and events were checked against the pattern described

for the scenario.

4.4 Signiﬁcance of the Results

The achieved results provide a proof of concept for

the functionality of the developed procedure for the

identiﬁcation of driving scenarios and the deﬁned ac-

tivities and events. However, the test cases only cover

the driving situations that could be derived from the

limited amount of available measurement data at the

time of publication.

A further validation is recommended on the ba-

sis of a larger amount of measurement data for which

the time periods of the scenarios contained are al-

ready marked correctly. Not least in order to be able

to exclude possible errors in the identiﬁcation due to

special cases not taken into account with regard to

the course of the road or the behaviour of other road

users. In order to obtain a statistically reliable result

for the application in system validation, the existing

problems should be eliminated in advance by opti-

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

140

Figure 5: Objects for deﬁning the scenario cut-through.

mising the environment perception of the measuring

vehicles.

4.5 Discussion of the Methodology

The presented method for metamodelling scenarios

enables clear communication about their deﬁnition.

The employed class structure also allows a simple

and easy to understand transfer into an object-oriented

software for the identiﬁcation of scenarios in ﬁeld

data. Besides, it offers added value for the imple-

mentation of new scenarios to be identiﬁed. A new

scenario instance can be created straight forward by

referencing existing activities and events. The logic

for identifying a scenario instantiated this way can

be implemented universally and does not have to be

adapted for new scenarios.

Another relevant scenario for the highway domain

is the so-called cut-through scenario. In the latter,

a vehicle initially driving in a lane adjacent to the

ego vehicle’s, changes lanes ahead and into the lane

of the ego vehicle and holds this lane for an unde-

ﬁned period of time. Afterwards, the vehicle changes

lanes again in the same direction to the other adjacent

lane and keeps it. The scenario can be instantiated as

shown in Fig. 5 analogous to the scenario presented

in Section 3.2.

However, there are scenarios in the highway do-

main that require more detailed metamodelling, such

as the lost-cargo or crossing-object scenarios. In these

cases, the modelling must be extended to include ob-

jects lying on or moving across the street. Further-

more, corresponding signals must be used to derive

them. Nevertheless, it should be possible to retain the

presented concept for metamodelling with a domain

model for such scenarios.

5 CONCLUSIONS

In this work, an approach enabling the identiﬁcation

of highway scenarios in ﬁeld data was presented, ex-

emplarily implemented and validated. The approach

is based on qualitative scenario modelling. The mod-

elled scenarios are detected in ﬁeld data, which are

abstracted beforehand. To the best of our knowledge,

a detailed description of the actual process of pat-

tern recognition and its application to real-world data

has not yet been published. The results obtained in

ﬁrst experiments are promising and provide a proof

of concept for the approach’s functionality. Neverthe-

less, further investigations based on a larger amount

of measurement data will be carried out after solving

shortcomings in the measurement technology.

In addition to an even more detailed modelling of

the scenarios, the identiﬁcation algorithm could be

extended with machine learning methods in the fu-

ture. Such methods would require a large amount of

learning data in which the scenarios sought are clearly

identiﬁed.

REFERENCES

Bach, Johannes, Stefan Otten, and Eric Sax (2016). “Model

based scenario speciﬁcation for development and test of

automated driving functions.” In: 2016 IEEE Intelligent

Vehicles Symposium (IV). IEEE, pp. 1149–1155.

de Gelder, Erwin et al. (2020). “Ontology for scenarios

for the assessment of automated vehicles.” In: arXiv

preprint arXiv:2001.11507.

Elrofai, Hala, Dani

el Worm, and Olaf Op den Camp (2016).

“Scenario identiﬁcation for validation of automated

driving functions.” In: Advanced Microsystems for Au-

tomotive Applications 2016. Springer, pp. 153–163.

Elrofai, H et al. (2018). StreetWise: scenario-based safety

validation of connected automated driving. Accessed:

2022-09-03. URL: http : / / publications . tno . nl /

publication/34626550/AyT8Zc/TNO-2018-streetwise.

pdf.

Geyer, Sebastian et al. (2014). “Concept and develop-

ment of a uniﬁed ontology for generating test and

use-case catalogues for assisted and automated vehi-

cle guidance.” In: IET Intelligent Transport Systems 8.3,

pp. 183–189.

King, Christian et al. (2017). “Identiﬁkation von

Fahrszenarien w

ahrend einer virtuellen Testfahrt.”

In: INFORMATIK 2017.

Identifying Scenarios in Field Data to Enable Validation of Highly Automated Driving Systems

141

Maurer, Markus et al. (2015). Autonomes Fahren: technis-

che, rechtliche und gesellschaftliche Aspekte. Springer

Nature.

Paardekooper, Jan-Pieter et al. (2019). “Automatic identi-

ﬁcation of critical scenarios in a public dataset of 6000

km of public-road driving.” In: 26th International Tech-

nical Conference on the Enhanced Safety of Vehicles

(ESV). Mira Smart, pp. 1–8.

Rasch, Maximilian et al. (2019). “Safety Assessment and

Uncertainty Quantiﬁcation of Automated Driver Assis-

tance Systems using Stochastic Analysis Methods.” In:

NAFEMS World Congress.

Ulbrich, Simon et al. (2015). “Deﬁning and substantiating

the terms scene, situation, and scenario for automated

driving.” In: 2015 IEEE 18th International Conference

on Intelligent Transportation Systems. IEEE, pp. 982–

988.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

142