Development and Implementation of a Concept for the Meta Description

of Highway Driving Scenarios with Focus on Interactions of Road Users

Raphael Pfeffer, Jingyu He and Sax Eric

Institute for Information Processing Technologies, Karlsruhe Institute of Technology, Engesserstr. 5, Karlsruhe, Germany

Keywords:

Automated Driving, Scenario based Testing, Meta Model, Simulation Methods, Scenario Classiﬁcation.

Abstract:

Nowadays reducing the individual risk for advanced driver assistant systems (ADAS) and automated driving

while guaranteeing the overall safety on the highway remains a big challenge. The identiﬁcation of corner test

cases and driving scenarios is key in the development process but is still not entirely solved. In the past, many

contributed to a uniﬁed scenario deﬁnition but often with different application focus. In this paper, we develop

a new scenario meta model based on existing deﬁnitions serving a development and test process where the

test data is captured in real (test) drives and its contained scenarios are derived. We present the novelty of our

scenario model describing the behaviour of dynamic objects in highway situations and show a ﬁrst application

of our model and results calculating the uniqueness of scenarios using auto-encoders.

1 INTRODUCTION

With the rising public demand on the intelligent and

safe driving functionality, more and more key innova-

tions in the automotive domain include increasingly

autonomous features appeared in the recent decades,

especially under the conditions of driving on the high-

way, like Adaptive Cruise Control (ACC) in 1997

(Winner et al., 2012) and the ﬁrst Lane Keeping Sup-

port (LKS) system in a Nissan production car in 2001

(Kawazoe et al., 2001). Bearing this in mind, many

questions arise, such as how to verify upcoming au-

tomated driving functions and how to guarantee sufﬁ-

ciency of the veriﬁcation and validation process, if a

signiﬁcant test coverage must be achieved (Langner

et al., 2018). To test the automated vehicles, the

widely used approach is driving millions of test kilo-

metres in the real world. For example, Waymo an-

nounced in 2018 that its autonomous vehicles have

driven more than 16 million kilometres on public

roads in the United States (Alva, 2018). However,

for highly automated driving functions, this approach

is not a economically feasible method for a full vali-

dation as billions of kilometres are needed (Wachen-

feld and Winner, 2016). Contrary to this statisti-

cal approach of physical test coverage, with purely

simulation-based methods, a completeness of the re-

quired test coverage cannot be guaranteed.

Nevertheless, simulation methods offer advan-

tages in terms of conﬁgurability of test cases and ef-

ﬁciency in test execution. In this context, scenario-

based testing has become a central concept for

simulation-based XiL methods and beyond. Espe-

cially for the testing of driver assistance systems and

highly automated driving functions, deﬁned scenarios

often serve as the basis for deriving relevant test cases

for the system-under-test (Otten et al., 2018).

However, even scenario-based testing does not an-

swer the question of the completeness of the test cov-

erage, since the prior requirements cannot be derived

in the laboratory for highly automated driving func-

tions. Therefore, in our research work a concept for a

new scenario meta-description is developed. It is ap-

plicable to use recorded data from series or ﬂeet ve-

hicles or databases from previous recordings (Figure

1).

Figure 1: Process for real data based validation using simu-

lation methods according to (Pfeffer et al., 2019).

This contribution focuses on a description model

for the interaction with other road users and shows the

440

Pfeffer, R., He, J. and Eric, S.

Development and Implementation of a Concept for the Meta Description of Highway Driving Scenarios with Focus on Interactions of Road Users.

DOI: 10.5220/0009341804400447

In Proceedings of the 6th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2020), pages 440-447

ISBN: 978-989-758-419-0

application of machine learning techniques using this

model data with a ﬁrst target to discover the similarity

between random scenarios.

The paper is organized as follows. Section II

presents the previous works on deﬁning the testing

scenarios based on different application cases. Sec-

tion III illustrates the architecture of the matrices,

which can describe the scenario with the focus on the

interaction with other road users on a highway. Sec-

tions IV presents the implementation of the scenario

model in a simulation environment and the statistical

results of the generated scenario matrices. Section V

shows the ﬁrst approach on how the matrices can be

used to classify highway scenarios.

2 RELATED WORK

To test functionality, reliability and maturity of ad-

vanced driver assistance system and automated driv-

ing functions many achievements have been made in

establishing scenario-based testing throughout the de-

velopment process in the recent years.

2.1 Scenario based Testing

In 2014 Ulbrich et al. discuss different approaches of

context modelling for automated driving scenarios in

urban environments and analyze various context en-

tities and layers in this approach. In this work scene

is deﬁned as a relationship between the context enti-

ties in the system architecture of automated vehicles.

A scene can be taken as a link between vehicle guid-

ance module and perception module whereas a situ-

ation refers to a central interface in vehicle guidance

(Ulbrich et al., 2014).

Geyer et al. propose a metaphor-based terminol-

ogy that consists of scenario, situation, scenery and

scene of vehicles (Geyer et al., 2014). A driving

scenario of a vehicle is metaphorized as a theatre.

Scenery, as a structured collection of static elements,

allows the director to create an environment suitable

for the scene through different combinations, it con-

sists of the basic geometry with predeﬁned road types

(intersection, country road, a motorway with exit and

access, etc). Adding the dynamic elements, ego vehi-

cle (considered system) and instructions they become

a situation. A scenario consists of a series of situa-

tions (Figure 2).

Based on these works Ulbrich summarizes the

deﬁnitions and substantiations of the terms scenario,

situation and scene in the product life-cycle of au-

tomated vehicles and further illustrates the term sce-

nario in the context of the V-model and the ISO26262

Figure 2: Metaphorical scenario concept according to

(Geyer et al., 2014).

standard development process (Ulbrich et al., 2015).

To be speciﬁc, this is intended by use-case and

development-phase dependent levels of abstractions.

Figure 3: A scenario (dashed blue) as a temporal sequence

of actions/events (edges) and scenes (nodes) (Ulbrich et al.,

2015).

In this consideration, the term scenario can be un-

derstood as some kind of storyline, which can consist

of at least one scene, action (or events) and goals (or

values). The authors suggest the following deﬁnition

for a scenario:

”A scenario describes the temporal devel-

opment between several scenes in a sequence

of scenes. Every scenario starts with an initial

scene. Actions/events as well as goals/values

may be speciﬁed to characterize this tempo-

ral development in a scenario. Other than a

scene, a scenario spans a certain amount of

time.”(Ulbrich et al., 2015)

The principle of a scenario representation accord-

ing to this deﬁnition is illustrated in Figure 3.

Another description of the scenario concept is pro-

posed by Schuldt (Schuldt, 2017) in 2017 based on

(Schuldt et al., 2015) and (Schuldt et al., 2018). Based

Development and Implementation of a Concept for the Meta Description of Highway Driving Scenarios with Focus on Interactions of Road

Users

441

on the previous developments to the scenario term,

Schuldt introduces a test method, which is called a

scenario-based testing approach. In this work a 4

layer model for a testing scenario is developed which

contains different description layers for road geome-

try and topology, situation-speciﬁc adaptation of the

road, environment and actors in which the previous

scenario deﬁnitions are integrated. Compared with

(Ulbrich et al., 2015), actions and goals are speciﬁed

in detail and summarized together with the dynamic

elements as the maneuvers.

Taking into account different requirements for the

representation of scenarios in various process steps

and aiming for a human-understandable scenario no-

tation, in the context of the PEGASUS-project Men-

zel et al. discuss a detailed scenario description to

design, development and testing automated vehicles

in 2018 (Menzel et al., 2018). As evidenced by their

self-reported survey results, the term ”scenario” has

not been deﬁned uniformly, which makes it difﬁcult

to achieve a consistent understanding regarding the

role of scenarios in the development process. There-

fore, three abstract levels of the scenario are proposed,

namely, functional level, logical level and concrete

level (Menzel et al., 2019). While functional scenar-

ios represent a level at which relevant properties are

documented and described in language, logical sce-

narios deﬁne parameter areas in the state space for

these properties in the form of entities and scenario

relationships. Concrete scenarios are actual charac-

teristics of a logical scenario, meaning that they rep-

resent a ﬁxed parameter set. Logical scenarios can

therefore be converted into at least one concrete sce-

nario.

In the same year Steimle et al. propose a structure

for a scenario terminology in the context of scenario-

based testing using UML diagrams (Steimle et al.,

2018b). In (Steimle et al., 2018a), their theory is

further extended with three abstract layers combined

with the scenario described in (Menzel et al., 2018).

At this point the common general deﬁnition of sce-

nario and its elements have stabilized so far although

there is still no unifying standard.

2.2 Scenario Meta Formats

Scenarios and generated test cases must be struc-

tured in a generic way to be able to describe scenario

content in different environments, such as highways,

country roads or city streets, and to perform them with

different test tools. Ideally, the structure can support

the standardized usage of scenarios to seamlessly per-

form the scenarios throughout a variety of test tools,

such as simulation. Therefore, it must allow existing

test cases to be adapted to new or modiﬁed test ob-

jects. How can scenarios be represented technically

to fulﬁl the requirements on various layers of abstrac-

tion with a standard format?

The OpenScenario project tries to answer this

question. OpenScenario is an open ﬁle format for the

description of dynamic contents in driving simulation

applications developed by ASAM (Association for

Standardization of Automation and Measuring Sys-

tems)

. As a tool independent open ﬁle format, Open-

Scenario speciﬁes the time-variant behaviour of en-

tities during one simulation. State changes of traf-

ﬁc participants, infrastructure and other dynamic ob-

jects are described as actions triggered by conditions

(PEGASUS-project, 2018). Due to the powerful lan-

guage range connected with the OpenDrive (VIRES

Simulationstechnologie GmbH, 2015), complex traf-

ﬁc situations from the highway to urban areas can be

described. OpenScenario project hat so far not been

completed, but as an open description format it is cur-

rently widely used in research and industry.

For the description of dynamic components in

a scenario different further approaches of semantic

models exist. Exemplary, Wang et al. propose a

new description for trafﬁc element semantic features

(Wang et al., 2018). In this work the global position

of ego vehicle and trafﬁc elements are combined and

transformed to the relative distances and angles and

are described as a matrix of trafﬁc elements in each

scene. Kohlhaas et al. present a further concept for

the structured mapping of relations between the ego

vehicle and other road users as well as the environ-

ment (Kohlhaas et al., 2014). The concept also tries

to take into account the dynamics of the connections

and formalizes rules for the transitions of the dynamic

objects between the different states, which lead to new

possible conﬁgurations or exclude invalid conﬁgura-

tions. Petrich et al. extend this concept and design

a compact semantic tensor representation to describe

the relationships of all road users of a scenario in pairs

(Petrich et al., 2018).

3 NEW META MODEL FOR

HIGHWAY APPLICATIONS

Inspired by these approaches and following the re-

quirement that the data must be derivable by a (sen-

sor) record (see Figure 1), a model for the state and

scene representation of the ego vehicle in relation to

the surrounding trafﬁc objects is suggested in this

work. The spatial relations result in a possible dis-

https://www.asam.net/standards/detail/openscenario/

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

442

cretization for scenes on a highway to three-state pos-

sibilities each in longitudinal and lateral direction.

Besides, we propose an action model based on

the driving maneuvers for the transitions between the

scenes for all dynamic objects. This allows a scenario

in the sense of the previously presented deﬁnitions to

be modeled. The main elements will be presented in

the following subsections.

3.1 Matrix Representation for the Scene

Description

At ﬁrst, the relevant highway section depending on

the region of interest is divided into blocks using a

simple grid model.

The positions of all dynamic objects are captured

and longitudinally and laterally encoded relative to

the ego vehicle along the road surface. The scene

is divided laterally according to the lanes. Longitu-

dinally we assume a maximum observation radius of

200m and deﬁne the state ”same level” at a maximum

relative distance of 10m to the ego vehicle (see Figure

4).

Figure 4: Matrix for the scene description.

In doing so, each of the dynamic objects, that ap-

pears surrounding the ego vehicle are taken into ac-

count of the scene description. We use an 1*6 matrix

where the ﬁrst three digits represent the count of ob-

ject vehicles that longitudinally appear in the front,

same level and behind the ego vehicle. The last three

digits indicate the number of object vehicles that lat-

erally are located in the left lane, the same lane and

the right lane compared to the ego vehicle.

For example, if two object vehicles are longitudi-

nally in the behind position to the ego vehicle, then

the ﬁrst three digits of the 1*6 matrix should be 0, 0

and 2. At the same time, one of the two object ve-

hicles is on the left lane and the other is on the right

lane, as a result, the last three digits of the 1*6 ma-

trix should be 1, 0 and 1. So the whole 1*6 matrix

description should be [0;0;2;1;0;1].

3.2 Scenario Meta-description Model

Sequences between two scenes (as quasi static snap-

shots) are determined by the included actions and ma-

neuvers (see Figure 3). In our work we determined

so-called basic maneuvers (e.g. lane change left for

all dynamic objects which allow clear transitions be-

tween the scenes respectively the depicted scenes ab-

straction. To describe the characteristics precisely,

additional parameters are recorded. For longitudinal

maneuvers, these are, for example, the initial and ﬁnal

velocity, the duration and the distance. Lateral ma-

neuvers are described by the initial position and ﬁnal

lateral position as well as the duration of the maneu-

ver. The maneuvers and their parameters are captured

in a matrix representation analogous to the presented

scene representation.

To merge the information for the dynamic pro-

gression of a scenario we use a tensor with three di-

mensions: time, objects, and meta information. Meta

information contains the abstractions for the scenes

and actions previously presented. These are recorded

for all objects along a (recorded) scenario sequence.

Along the time vector, time stamps of scene changes

and action/maneuver changes are stored (Figure5).

Figure 5: Principle of scenario meta information tensor for

dynamic objects.

3.3 Scenario Selection

The scenario description model presented allows the

storage of records of any length according to the pro-

cess described in Figure 1. The question arises, how

long should the scenarios actually be? As the cur-

rent state of the art cannot provide a clear answer to

this question, the application possibilities in our meta

model are also kept ﬂexible by its structure. Thus,

scenarios of any length can be displayed with the ten-

sor model (or be extracted from it). For example, a

”cut-in vehicle” scenario could be represented by two

scenes, while a scenario ”overtaking” is possibly de-

scribed by 3-4 consecutive scenes (see also Figure 8.

Another advantage is that even scenarios of different

Development and Implementation of a Concept for the Meta Description of Highway Driving Scenarios with Focus on Interactions of Road

Users

443

lengths of time can be compared if they belong to the

same scenario class (e.g. ”overtaking”) since they still

have the same tensor representation length. Figure 6

shows an example for two consecutive scenes (as part

of an e.g. ”overtaking” scenario) and its scene matrix

representation. The 1*12 scenario matrix means the

chronological connection of two 1*6 scene matrices

and represents the transformation of different object

vehicle distributions in a scenario with two scenes.

Figure 6: Matrix description for a scenario with two scenes.

4 IMPLEMENTATION IN

SIMULATION

To identify the technical feasibility and adopt an engi-

neering assessment, the scenario tensors in the devel-

opment process can be extracted from real highway

driving. In this stage of our work the data is ﬁrst gen-

erated in a virtual driving simulation. As a replace-

ment for a real test drive, the simulation software Car-

Maker is used. IPG CarMaker is a test platform which

allows to create real-world test scenarios in a vir-

tual environment, simulating every type of road and

trafﬁc, and performing realistic execution through an

event and maneuver-based testing method (IPG Auto-

motive GmbH, 2019). CarMaker provides models for

vehicle with its sub-components, sensors, driver, road

and many surrounding elements. All models and ele-

ments can be changed to generate a different behavior.

Various of virtual sensors can be mounted on the ego

vehicle model representing a vehicle with ADAS or

automated driving functions (Figure 7).

The relevant information can be extracted using

those sensors. For example, distance, relative speed

and relative accelerations to all surrounding objects

within the sensor range can be calculated. These data

can also be used to calculate the grid positions accord-

ing to our presented concept. Consequently, the over-

all scenario tensor is created as a result of each simu-

Figure 7: Vehicle model equipped with sensors in CarMaker

and virtual highway environment.

Table 1: Test series in this work.

Tests Trafﬁc density Distance

Test series 1 17.5 vehicles per km 743.1 km

Test series 2 7.5 vehicles per km 181.9 km

lation run in a post-processing step. In this work, we

created two highway test series with different high-

way cases (see Table 1). One simulates driving sce-

narios with crowded object vehicles, which represent

slow moving trafﬁc situations and jam. In this case we

generated 128 autonomously driving objects on a 7.3

km two-lane highway, which means the trafﬁc ﬂow

density is 17.5 vehicles per kilometer (test series 1).

The other test series represents a lower trafﬁc density

situation of 7.5 vehicles per kilometer, which has 56

vehicles on the 7.3 km highway (test series 2). Based

on these two basic tests we created random variations

by manipulating parameters like cruising speed of ego

vehicle, driver aggressiveness or the lane change be-

haviour of the dynamic objects. Finally, we run all

variations and saved the ego vehicle’s (sensor) data

with a sample rate of 10 samples per second, from

which the scenario meta information tensors were cal-

culated. In total for the application presented here, we

generated an overall length of 925 kilometers of high-

way driving data in the simulation.

5 APPLICATION: SCENARIO

SIMILARITY

5.1 Scenario Composition by Highway

Maneuvers

The scenario meta-description has a wide range of

applications. The precise description of a feature

from a whole scenario is among the ﬁrst. Maneuvers

are a certain pattern of the whole scenario (Schuldt,

2017). As maneuvers are deﬁned as the transitions

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

444

between two or more unique scenes we can use scene

sequences to identify speciﬁc scenarios. Finally, our

concept can be used to identify more complex ma-

neuvers (than the basic maneuvers captured), maneu-

ver sequences, or even speciﬁc scenarios in the meta-

description. For this purpose we simply use the scene

encodings in the following application. In this con-

text, the concept can be used to classify and extract

speciﬁc scenarios out of a time series which can be

used for a particular unit-under-test. Figure 8 shows

possible examples with its scenario meta-descriptions

for four common scenarios on a highway.

Figure 8: Scenario descriptions using scene matrix.

5.2 Estimating the Uniqueness of

Scenario Matrices

In today’s scenario-based testing for automated func-

tions, a common approach is the manual scenario

speciﬁcation (Ulbrich et al., 2015) which can be ex-

tremely time-consuming if parameter boundaries are

not considered properly. Some dedicate attention

to critical situations and corner cases (Eckstein and

Zlocki, ) but where a full test coverage cannot be

achieved if the test cases generation happens without

the consideration of real data.

To overcome these challenges we try to apply our

scenario meta model to a test process where the data

is derived from real test drives and the scenarios can

be extracted automatically. Doing so, it is a necessity

to be able to distinguish between new and so far un-

seen scenarios and known scenarios from e.g. former

recordings. For an existing database of an arbitrary

quantity of scenarios in terms of the interaction be-

tween vehicles, all new scenarios are supposed to be

added to the database. For this reason the uniqueness

of new scenarios has to be estimated. As a result, new

(unique) scenario matrices provides no redundancy to

the database. To achieve a higher test coverage for the

function under test, test scenarios with new features

and new content are added to the existing test scenario

data pool. To solve this task we choose to apply the

concept of auto-encoders as a machine learning tech-

nique for the detection of anomalies (here new sce-

narios) in a ﬁrst implementation which is presented

as follows:

Auto-encoders are artiﬁcial neural networks with

the ability to learn a compressed representation (en-

coding) for a set of data and thus also to extract es-

sential features. Therefore, the similarity of exter-

nal data to a training data set can be derived using

the trained auto-encoder. While training the auto-

encoder, the trained model is validated with the val-

idation data in each training epoch. To get the model

with the best performance, the training hyperparame-

ters, in this case batch size and learning rate, are ad-

justed and the average training error is compared after

each training process. The Mean Square Error (MSE)

is used as the reconstruction error.

Figure 9: Calculation for uniqueness of scenario matrix us-

ing auto-encoders.

To ensure the uniqueness of the added data, sce-

nario matrices that have a certain similarity degree to

the existing scenarios are ﬁltered and rejected. Figure

9 depicts the process to calculate the uniqueness of

the test series 2 data and to ﬁlter the data that are sim-

ilar to test series 1. The scenario matrices from test

series 1 in CarMaker are set as the training data and

the data from the test series 2 are used to be detected

and ﬁltered. After the ﬁnished training, the models

are applied on the scenario data of test series 2. The

MSE as the reconstruction error is calculated for each

sample, the results are then visualised to provide an

intuitive impression.

A total of four auto-encoders is trained here to rep-

resent scenario descriptions consisting of one single

scene up to four consecutive scenes. For the training

data for each auto-encoder, the data set is split into

training data and validation data with a ratio of 9:1.

In the training data from test series 1, scenario with

one unique scene has 1439 samples, scenario with two

unique scenes has 6312 samples, scenario with three

unique scenes has 13302 samples and scenario with

four unique scenes has 17666 samples.

Ultimately, a threshold for the reconstruction er-

ror for each type of scenario is subjectively deter-

mined and according to this threshold the scenarios

in test series 2 are classiﬁed into two classes. Those

with reconstruction errors below the threshold are re-

garded as known scenarios, while the others with re-

construction error above the threshold are regarded as

unknown scenarios (anomaly scenarios). The classiﬁ-

Development and Implementation of a Concept for the Meta Description of Highway Driving Scenarios with Focus on Interactions of Road

Users

445

Figure 10: First results evaluation using confusion matrix.

cation model for unknown data detection is then eval-

uated by a confusion matrix.

Figure 10 describes the confusion matrix for sce-

narios with one unique scene in contrast to scenarios

with four unique scenes. The results are evaluated in a

confusion matrix in which each row of the matrix rep-

resents the scenarios in a predicted class while each

column represents the scenarios in the ground truth.

The visualization for the performance of the two clas-

siﬁcations shows the marked difference in precision

rate. With respect to the rate of positive predicted test

series 2 data, the scenario matrices with one unique

scene work better than scenario matrices with four

unique scenes. The implementation achieves an accu-

racy at least for one unique scene scenarios where no

unknown scenarios are falsely classiﬁed as known and

thus rejected. This shows auto-encoders can be con-

sidered as an applicable method to classify unknown

scenarios according to our meta model.

6 FUTURE WORKS

There are lots of additional challenges to be addressed

with regards to the superior testing process we refer

to in the introduction. Some ideas that would need

further investigation are the following. A concept for

the driving scenario meta-description is proposed in

this contribution with its advantages as mentioned be-

fore. However, this concept in the current stage only

considers the behaviour of the dynamic elements of

a scenario. Many other features that can appear in a

scenario e.g. trafﬁc signs, further object properties,

weather conditions and even the context of the total

scenario are additional aspects which could be taken

into account. In our future study, we will extend our

concept and integrate more and more features into our

meta model.

With regards to the auto-encoder, it has been

shown that the auto-encoder is an efﬁcient algorithm

to reconstruct the features of a scenario. In this work

auto-encoders with two hidden layers only have been

used so far which do not seem deep enough to learn

more features and ﬁt larger data sets. In the future

study we will extend the structure of auto-encoder so

that it can reconstruct new scenarios more efﬁciently

and achieve a better performance for classifying the

known and unknown data from the existing scenario

database.

The ultimate expectation for this research is to de-

velop an integrated approach for efﬁcient testing of

automated vehicles. To validate the concept and its

results so far it is needed to collect data from real tests

instead of simulation and apply the presented methods

in the future research.

7 CONCLUSIONS

In this paper, the state of the art for scenario deﬁnition

is reviewed. We show a new approach for a scenario

meta model with the focus on dynamic scenario ele-

ments which is suitable to be applied to a test process

where the data is derived by physical test drives. An

essential part of the scenario meta model is the rep-

resentation of scenes which tries to solve the problem

with time series data, where each sample has its du-

ration by splitting the time series data into segments

with the same state. Each scene segment is described

in the form of a matrix of the same size independent

of the actual time length so that the same type of sce-

nario is encoded by the same size. By this approach

scenarios of different recordings become comparable

by the use of machine learning algorithms. We pre-

sented a ﬁrst use case applied on our scenario meta

model calculating the uniqueness of new scenarios

compared with an existing data base with the help of

auto-encoders.

Another advantage of this new scenario meta

model is that the concept has a high versatility and

can be extended easily. It is applicable to all features

in a driving situation which can be recorded. For ex-

ample, further properties like the weather conditions

or trafﬁc signs could also be integrated into the scene

descriptions and therefore also be part of the scenario

analysis. Last but not least, the concept has the po-

tential to extend driving scenario descriptions to any

period of time. At the end it should be possible to ex-

tract any sequences of scenes (scenarios) even out of

recordings made over several hours.

REFERENCES

Alva, M. (2018). Waymo Self-Driving Cars Have Driven 16

Million Kilometres. [Online; accessed 01-July-2019].

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

446

Eckstein, L. and Zlocki, A. Safety potential of adas – com-

bined methods for an effective evaluation. In The 23rd

International Technical Conference on the Enhanced

Safety of Vehicles (ESV) Seoul, Republic of Korea,

May 27-30.

Geyer, S., Baltzer, M., Franz, B., Hakuli, S., Kauer, M.,

Kienle, M., Meier, S., Weißgerber, T., Bengler, K.,

Bruder, R., Flemisch, F., and Winner, H. (2014). Con-

cept and development of a uniﬁed ontology for gen-

erating test and use-case catalogues for assisted and

automated vehicle guidance. In IET Intelligent Trans-

port Systems, volume 8, pages 183–189.

IPG Automotive GmbH (2019). CarMaker User’s Guide

(Version 8.0).

Kawazoe, H., Murakami, T., Sadano, O., Suda, K., and

Ono, H. (2001). Development of a lane-keeping sup-

port system. In SAE Technical Paper. SAE Interna-

tional.

Kohlhaas, R., Bittner, T., Schamm, T., and Z

ollner, J. M.

(2014). Semantic state space for high-level maneuver

planning in structured trafﬁc scenes. In 17th Interna-

tional IEEE Conference on Intelligent Transportation

Systems (ITSC), pages 1060–1065.

Langner, J., Bach, J., Ries, L., Otten, S., Holz, M., and Sax,

E. (2018). Estimating the Uniqueness of Test Scenar-

ios derived from Recorded Real-World-Driving-Data

using Autoencoders. In 2018 IEEE Intelligent Vehi-

cles Symposium (IV), number Iv, pages 1860–1866,

Changshu, Suzhou, China.

Menzel, T., Bagschik, G., Isensee, L., Schomburg, A., and

Maurer, M. (2019). From functional to logical sce-

narios: Detailing a keyword-based scenario descrip-

tion for execution in a simulation environment. In

2019 IEEE Intelligent Vehicles Symposium (IV), pages

2383–2390.

Menzel, T., Bagschik, G., Maurer, and Markus (2018). Sce-

narios for Development, Test and Validation of Auto-

mated Vehicles. In IEEE Intelligent Vehicles Sympo-

sium, Proceedings, volume 2018-June, pages 1821–

1827.

Otten, S., Bach, J., Wohlfahrt, C., King, C., Lier, J.,

Schmid, H., Schmerler, S., and Sax, E. (2018). Auto-

mated assessment and evaluation of digital test drives.

In Zach

aus, C., M

uller, B., and Meyer, G., editors,

Advanced Microsystems for Automotive Applications

2017, pages 189–199, Cham. Springer International

Publishing.

PEGASUS-project (2018). Basics for Testing – Stand 11:

SCENARIO FORMATS. [Online; accessed 23-April-

2019].

Petrich, D., Azarfar, D., Kuhnt, F., and Zollner, J. M.

(2018). The ﬁngerprint of a trafﬁc situation: A seman-

tic relationship tensor for situation description and

awareness. In 21st IEEE International Conference on

Intelligent Transportation Systems, ITSC 2018; Maui;

United States; 4 November 2018 through 7 November

2018, pages 429–435. IEEE, Picataway (NJ).

Pfeffer, R., Schmidt, S., and Sax, E. (2019). Real data based

validation of highly automated driving functions us-

ing simulation methods. ATZelectronics worldwide,

14(11):24–29.

Schuldt, F. (2017). Ein Beitrag f

ur den methodischen

Test von automatisierten Fahrfunktionen mit Hilfe von

virtuellen Umgebungen. PhD thesis.

Schuldt, F., Maurer, M., and Menzel, T. (2015). Eine Meth-

ode fur die Zuordnung von Testf

allen f

ur automa-

tisierte Fahrfunktionen auf X-in-the-Loop Verfahren

im modularen virtuellen Testbaukasten. In 10. Work-

shop Fahrerassistenzsysteme., pages 171–182, Walt-

ing, Deutschland.

Schuldt, F., Reschka, A., and Maurer, M. (2018). A Method

for an Efﬁcient, Systematic Test Case Generation for

Advanced Driver Assistance Systems in Virtual Envi-

ronments, pages 147–175. Springer International Pub-

lishing, Cham.

Steimle, M., Bagschik, G., Menzel, T., Maurer, M., and

Wendler, J. T. (2018a). Anwendung eines Grund-

vokabulars f

ur den szenarienbasierten Testansatz au-

tomatisierter Fahrfunktionen anhand eines Beispiels.

In AAET - Automatisiertes und vernetztes Fahren.,

Braunschweig, Deutschland.

Steimle, M., Bagschik, G., Menzel, T., Wendler, J. T., and

Maurer, M. (2018b). Ein Beitrag zur Terminologie

ur den szenarienbasierten Testansatz automatisierter

Fahrfunktionen. In AAET 2018 // AAET - Automa-

tisiertes und Vernetztes Fahren, number September,

pages 11–32.

Ulbrich, S., Menzel, T., Reschka, A., Schuldt, F., and Mau-

rer, M. (2015). Deﬁning and Substantiating the Terms

Scene , Situation , and Scenario for Automated Driv-

ing. In 2015 IEEE 18th International Conference on

Intelligent Transportation Systems, pages 982–988.

Ulbrich, S., Nothdurft, T., Maurer, M., and Hecker, P.

(2014). Graph-based context representation, environ-

ment modeling and information aggregation for auto-

mated driving. In IEEE Intelligent Vehicles Sympo-

sium, Proceedings, pages 541–547.

VIRES Simulationstechnologie GmbH (2015).

OpenDRIVE

 Format Speciﬁcation, Rev. 1.4.

[Online; accessed 15-Feb-2020].

Wachenfeld, W. and Winner, H. (2016). The Release of Au-

tonomous Vehicles, pages 425–449. Springer Berlin

Heidelberg, Berlin, Heidelberg.

Wang, J., Zhang, C., Liu, Y., and Zhang, Q. (2018). Trafﬁc

Sensory Data Classiﬁcation by Quantifying Scenario

Complexity. In IEEE Intelligent Vehicles Symposium,

Proceedings, volume 2018-June, pages 1543–1548.

Winner, H., Danner, B., and Steinle, J. (2012). Adap-

tive Cruise Control, pages 478–521. Vieweg+Teubner

Verlag, Wiesbaden.

Development and Implementation of a Concept for the Meta Description of Highway Driving Scenarios with Focus on Interactions of Road

Users

447