A-Eye: Driving with the Eyes of AI for Corner Case Generation

Abstract

The overall goal of this work is to enrich training data for automated driving with so called corner cases. In road traffic, corner cases are critical, rare and unusual situations that challenge the perception by AI algorithms. For this purpose, we present the design of a test rig to generate synthetic corner cases using a human-in-the-loop approach. For the test rig, a real-time semantic segmentation network is trained and integrated into the driving simulation software CARLA in such a way that a human can drive on the network's prediction. In addition, a second person gets to see the same scene from the original CARLA output and is supposed to intervene with the help of a second control unit as soon as the semantic driver shows dangerous driving behavior. Interventions potentially indicate poor recognition of a critical scene by the segmentation network and then represents a corner case. In our experiments, we show that targeted enrichment of training data with corner cases leads to improvements in pedestrian detection in safety relevant episodes in road traffic.

Full text

A-Eye: Driving with the Eyes of AI for Corner Case Generation
Kamil Kowol1and Stefan Bracke2and Hanno Gottschalk1
Abstract— The overall goal of this work is to enrich training
data for automated driving with so called corner cases. In road
traffic, corner cases are critical, rare and unusual situations that
challenge the perception by AI algorithms. For this purpose, we
present the design of a test rig to generate synthetic corner cases
using a human-in-the-loop approach. For the test rig, a real-
time semantic segmentation network is trained and integrated
into the driving simulation software CARLA in such a way that
a human can drive on the network’s prediction. In addition,
a second person gets to see the same scene from the original
CARLA output and is supposed to intervene with the help of
a second control unit as soon as the semantic driver shows
dangerous driving behavior. Interventions potentially indicate
poor recognition of a critical scene by the segmentation network
and then represents a corner case. In our experiments, we show
that targeted enrichment of training data with corner cases
leads to improvements in pedestrian detection in safety relevant
episodes in road traffic.
I. INTRODUCTION
As digitization continues, more and more assistance sys-
tems are being developed for automated driving, using AI
systems to support them. Since AI systems need to be
extensively trained and validated, large amounts of data from
the real or virtual world are required. To increase robustness
and performance, this data should consist of a large number
of clean and diverse scenes3.
In this work, we would like to present an accelerated
testing strategy that uses a human-in-the-loop approach to
capture corner cases to achieve performance improvement
based on safety-critical scenes. In order to obtain many
safety-critical corner cases in a short time, we stop training at
an early stage so that the network is sufficiently well trained.
Nevertheless, the scenes generated in this way are still useful
to improve fully trained networks.
For this purpose, a semantic network is trained with
synthetic images from the open-source driving simulation
software CARLA[1]. In addition, a test rig consisting of 2
control units is connected to CARLA in such a way that the
ego vehicle can be controlled with both control units by a
human. In this process, the semantic segmentation network
is integrated into CARLA in such a way that first the original
CARLA image is sent through the network and the prediction
is displayed on the screen of one driver. The second driver,
1Kamil Kowol and Hanno Gottschalk are with School of
Mathematics and Natural Sciences, University of Wuppertal,
IZMD, Lise-Meitner-Straße 27-31, Wuppertal, Germany, {kowol,
hgottsch}@uni-wuppertal.de
2Stefan Bracke is with Chair of Reliability Engineering and Risk Analyt-
ics, University of Wuppertal, IZMD, Gaußstraße 20, Wuppertal, Germany,
{bracke}@uni-wuppertal.de
3Andrej Karpathy, keynote speaker at CVPR 2021 Workshop on Au-
tonomous Driving, online, URL: http://cvpr2021.wad.vision/
in turn, sees the real CARLA image and is supposed to
intervene as a safety driver only if he or she feels that a
situation is being wrongly assessed by the other driver. We
aim to consider situations in which the AI algorithms lead
to incorrect evaluations of the scene, which we refer to as
safety relevant corner cases, in order to improve performance
through targeted data enrichment. This is done by exchanging
images from the original dataset with the safety-critical
corner cases, thus keeping the total amount of data fixed. We
show that the semantic segmentation network that contains
safety-critical corner cases in the training data performs
better on similar critical situations than the network that does
not contain any safety critical situations.
Our approach somehow follows the idea of active learning,
where we get feedback on the quality of the prediction by
interactively querying the scene. However, unlike in standard
active learning we do not leave the query strategy to the
learning algorithm, but make use of the human’s fine tuned
sense of risk to query safety relevant scenes from a large
amount of street scenes, leading to enhanced performance in
safety critical situations.
The contributions of this work can be summarized as
follows:
•An experimental setup, that could also be implemented
in the real world, that permits testing the safety of the
AI perception separately from the full system safety
including the driving policy of an automated vehicle.
•A proof of concept for the retrieval of training data for
automated driving which is safety relevant.
•A proof that training on safety relevant situations is
beneficial for the recognition of street hazards.
Outline: Section II discusses related work on corner
cases, human-in-the-loop and accelerated testing. In Sec-
tion III we briefly describe the experiment setup used for
corner case generation and which data and network was used
for our experiments. In Section IV we explain our strategy
to generate corner cases. In Section V we demonstrate the
beneficial effects of training with corner cases for safety
critical situations in automated driving. Finally, we present
our conclusions and give an outlook to future directions of
research in Section VI.
II. RELATED WORKS
A. Corner Cases
Training data contains few, if any, critical, rare or unusual
scenes, so-called corner or edge cases. In the technical field,
the term corner case describes special situations that occur
outside the normal operating parameters [2].

Fig. 1. View of the semantic driver (top) and the safety driver (bottom). Fig. 2. Test rig including steering wheels, pedals, seats and screens.
According to [3], a corner case in the field of autonomous
driving describes a ”non-predictable relevant object/class in
relevant location”. Based on this definition, a corner case
detection framework was presented to calculate a corner
case score based on video sequences. The authors of [4]
subsequently developed a systematization of corner cases,
in which they divide corner cases into different levels and
according to the degree of complexity. In addition, examples
were given for each corner case level. This was also the
basis for a subsequent publication with additional examples
[5]. Since the approach in these references is camera-based,
a categorization of corner cases at sensor level was adapted
in [6], where RADAR and LiDAR sensors were also con-
sidered. Furthermore, this reference presents a toolchain for
data generation and processing for corner case detection.
Outside normal parameters also includes terms such as
anomalies, novelties, or outliers, which, according to [6],
correlate strongly with the term corner case. In road traffic,
the detection of new and unknown objects, anomalies or
obstacles, which must also be evaluated as ’outside the oper-
ating parameters’, is essential. To measure the performance
of methods for detecting such objects, the benchmark suite
”SegmentMeIfYouCan” was created [7], [8]. In addition, the
authors present two datasets for anomaly and obstacle seg-
mentation to help autonomous vehicles better assess safety-
critical situations.
In summary, the term ”corner case” can encompass rare
and unusual situations that may include anomalies, unknown
objects or outliers which are outside of operating parameters.
Outside the operating parameters, in the context of machine
learning, means that these situations or objects were not part
of the training data.
B. Human-In-The-Loop
In machine learning, human-in-the-loop approaches com-
bine human and machine intelligence to develop synergies
for solving a specific problem. Efficiency can be increased in
this area because machines learn from knowledge of a human
[9], [10]. This provide faster and/or more accurate results
[11]. As the interest in ”human-in-the-loop” and ”machine
learning” is increasing [12], different approaches with a
driving simulator and a test rig have been used and tested in
recent years, which will now be briefly summarized.
The authors of [13] propose a real-time Deep Rein-
forcement Learning (Hug-DRL) method based on human
guidance, where a person can intervene in driving situations
when the agent makes mistakes. These driving errors can
be fed directly back into the agents’s training procedure and
improve the training performance significantly.
In [14] a realistic test rig including a steering wheel and
pedals for data collection was developed. Therefore, thirteen
subjects were recruited to drive on different routes while
being distracted by static or dynamic objects or by answering
messages on their cell phones. By adding nonlinear human
behaviors and using realistic driving data, the authors have
been able to predict human driving behavior more accurately
in testing.
Another driving simulator was presented in [15] to develop
and evaluate safety and emergency systems in the first design
stage. In addition to a steering wheel, pedals and a gearshift
as well as monitors are connected to a computer. As software
they use a generic simulator for academic robotics which
uses the Modular Open Robots Simulation Engine MORSE
[16]. They used an experiment with four road users, one
human driver and three vehicles driving in pilot mode and
forcing two out of 36 collision situations (a lead vehicle
stopped and a vehicle changing lanes) defined by the Na-
tional Traffic Safety Administration (NHTSA). The impact
of a driver assistance system on the driver was one of the
factors studied.
In [17] a human-like driving and decision-making frame-
work is introduced, where the lane-changing process is
examined. Using the brain emotional learning circuit model
(BELCM), a human-like driving model is designed and
evaluated with human-in-the-loop experiments.
A decision-making corner case generation method for
connected and automated vehicles (CAVs) for test and eval-
uation purposes is proposed in [18]. For this, the behavioral
policy of background vehicles (BV) is learned through re-
inforcement learning and Markov’s decision process, which
leads to a more aggressive interaction with the CAV which
forces more corner cases under test conditions. The tests take

semantic driver
safety driver
true
CNN
Corner
Cases
intervention
CARLA
manual
control
computer output human decision
Fig. 3. Two human subjects can control the ego vehicle. The semantic driver moves the vehicle in compliance with traffic rules in the virtual world and
sees only the output of the semantic segmentation network. The safety driver, who sees only the original image, assumes the role of a driving instructor and
intervenes in the situation by braking or changing the steering angle as soon as a hazardous situation occurs. Intervening in the current situation indicates
poor situation recognition of the segmentation network and represents a corner case. Triggering a corner case ends the acquisition process and a new run
can be started.
place on the highway and include lane changes or rear-end
collisions.
C. Accelerated Testing
Accelerated testing strategies are intended to reduce work-
load and, therefore, costs [19], [20]. By imitating the real
process parameters, initial findings can be obtained in field
tests and, if necessary, optimizations can be made. From this,
predictions could be made about product life or performance
over time under more moderate conditions of use or design.
It is however necessary to carefully investigate the relevance
of the data acquired under more challenging conditions for
the actual use case [21].
III. EXP ERI MENTAL SE TUP
A. Driving Simulator
Targeted enrichment of training data with safety-critical
driving situations is essential to increase the performance
of AI algorithms. Since the generation of corner cases in
the real world is not an option for safety reasons, generation
remains in the synthetic world, where specific critical driving
situations can be simulated and recorded. For this purpose,
the autonomous driving simulator CARLA [1] is used. It is
an open-source software for data generation and/or testing
of AI algorithms. It provides various sensors to describe the
scenes such as camera, LiDAR and RADAR and delivers
ground truth data. CARLA is based on the Unreal Engine
game engine [22], which calculates and displays the behavior
of various road users while taking physics into account, thus
enabling realistic driving. Furthermore, the world of CARLA
can be modified and adapted to one’s own use case with the
help of a Python API.
For our work, we used the API to modify the script for
manual control from the CARLA repository. In doing so, we
added another sensor, the inference sensor, which evaluates
the CARLA RGB images in real-time and outputs the neural
network semantic prediction on the screen. An example is
shown in Figure 1. By connecting a control unit including
a steering wheel, pedals and a screen, to CARLA, we make
it possible to control a vehicle with ’the eyes of the AI’ in
the synthetic world of CARLA. We also connected a second
control unit with the same components to the simulator,
so that it is possible to control the same vehicle with 2
different control units, see Figure 2. The second control unit
is thus operated on the basis of CARLA clear image and can
intervene at any time. It always has priority and triggers that
the past 3 seconds of driving, which are buffered, are written
to the dataset on the hard disk. In order for the semantic
driver to follow the traffic rules in CARLA, the script had
to be modified additionally. The code has been modified to
display the current traffic light phase in the upper right corner
and the speed in the upper center.
B. Test Rig
The test rig consists of the following components: a
workstation with CPU, 3x GPU Quadro RTX 8000, 2 driving
seats, 2 control units (steering wheel with pedals), one
monitor for each control unit and two monitors for the control
center. The driving simulator software used is the open-
source software CARLA version 0.9.10.

C. Dataset for Initial Training and Testing
For training, a custom dataset was generated using
CARLA 0.9.10, consisting of 85 scenes ´
a 60 frames. In
addition, there is a validation dataset with 20 scenes. The
dataset was generated on seven maps with one fps and
contains the corresponding semantic segmentation image in
addition to the rendered synthetic image. The maps include
the five standard maps in CARLA and two additional maps
that offer a mix of city, highway and rural driving. Various
parameters can be set in CARLA, and we focused on the
number of Non-Player-Characters (NPC’s), including vehi-
cles and pedestrians, and on environment parameters such
as sun position, wind and clouds. Depending on the size
of the map, the number of NPC’s ranged from 50 to 150.
The clouds and wind parameters can be set in the range
between 0 and 100, with 100 being the highest value. The
wind parameter is responsible for the movement of tree limbs
and passing clouds and was in the range 0 and 50. The
cloud parameter describes the cloudiness, where 0 means that
there are no clouds at all, and 100 that the sky is complete
covered with clouds. We have chosen values between 0 and
30. The altitude describes the angle of the sun in relation
to the horizon of the CARLA world, with values between
−90 (midnight) and 90 (midday). Values between 20 and
90 were used for our purpose. The other environmental
parameters like rain, wetness, puddles or fog are set to zero.
The parameters are chosen so that the scenes reflect everyday
situations with a natural scattering of NPC’s and in similar
good weather. During data generation, the movement of all
NPC’s was controlled by CARLA. Furthermore, 21 corner
case scenes were used as test data, each containing 30 frames.
Another test dataset containing 21 standard scenes without
corner cases will serve as a comparison, each containing 30
frames.
D. Training
To drive on the predicted semantic mask, a real-time
capable network architecture is needed. For these purposes
the Fast Segmentation Convolutional Neural Network (Fast-
SCNN) model was used [23]. It uses two branches to
combine spatial details at high resolution and deep feature
extraction at lower resolution with an accuracy of 68.0%
mIoU at 123.5 fps on the Cityscapes dataset [24]. The
network was implemented in the python package PyTorch
[25] and training was done on a NVIDIA Quadro RTX
TABLE I
OVE RVI EW OF TH E DATASE TS USE D FOR TH E TRA ININ G.
no. comment meanpixels/scene
1 natural distribution 3583.6
2 pedestrian enriched 6101.1
3 corner case enriched 6215.7
The first dataset contains scenes with a natural spread of NPC’s based
on daily events. For the second dataset, scenes with a higher number of
pedestrians were generated to allow a fair comparison with dataset three,
where the training data was enriched with corner case scenes. All datasets
contain the same number of training images.
TABLE II
PER FORM ANCE M EASU REME NT ON TW O TEST D ATA SETS .
Traindata Safety Critical Testdata Natural Distributed Testdata
no. comment IoUpedestrians mI oU IoUpedestr ians mIoU
1 natural distribution 0.4600 0.6954 0.4937 0.761
2 pedestrian enriched 0.5399 0.6911 0.5586 0.7554
3 corner case enriched 0.5683 0.7173 0.5384 0.7517
The comparison shows that the addition of safety-critical scenes in training
also improves performance in testing with safety critical scenes.
8000 graphics card. Sixteen of the 23 classes available in
CARLA were used for training. Cross entropy was used as
the loss function and ADAM as the optimization algorithm.
A polynomial decay scheduler was also used to gradually
reduce the learning rate.
We intentionally stopped the training after 5 epochs to
increase the frequency of perception errors for the network.
The resulting network is sufficiently well trained to recognize
the road and all road users, although objects further away
are poorly recognized. An example is shown in in the top of
Figure 1.
IV. RETRIEVAL OF CORNER CASES
For the generation of corner cases we have considered the
following experimental setup. The scenes are recorded with
the help of two test operators in our specially constructed
test rig (see Figure 2), where one subject (safety driver) gets
to see the original virtual image and the other (semantic
driver) the output of the semantic segmentation network
(see Figure 1). The test rig is equipped with controls such
as steering wheels, pedals and car seats and connected to
CARLA to simulate realistic participation in road traffic.
The corner cases were generated as shown in Figure 3.
For this purpose, we use a real-time semantic segmenta-
tion network from Section III-D where visual perception
is limited. We note that autonomous vehicles according to
[26] 67619.81 km drive until an accident happens. Using a
poorly trained network as a part of our accelerated testing
strategy, we were able to generate corner cases after 3.34 km
in average between interventions of the safety driver. We note
however that the efficiency of the corner cases is evaluated
using a fully trained network.
If the safety driver triggers the recording of a corner
case the test operators label the corner case with one of
four options available (overlooking a walker or a vehicle,
disregarding traffic rules, intervening out of boredom) and
may leave a comment. Furthermore the kilometers driven
and duration of the ride are notated. The operators were told
to obey the traffic rules and not to drive faster than 50 km/h
during the test drives. After a certain familiarization period,
driving errors decreased and sudden braking by the semantic
driver was also reduced. The reason for this is that the
network partially represents areas as vehicles or pedestrians
with fewer pixels. Over time, a learning effect occurred in
the drivers to hide such situations because experience showed
that there was no object there due to the previous frames.

Fig. 4. Evaluation on corner case test data shows that the model using corner case data in training detects pedestrians better (top left) than the original
dataset (bottom left) and the dataset where more pedestrians were detected (top right). Bottom right shows the groundtruth image.
The rides are being tracked and by the intervention of
the safety driver the last 3 seconds of the scene are saved.
Subsequently, the scenes can be loaded and images saved
from the ego vehicle’s perspective using the camera and the
semantic segmentation sensor. We collect 50 corner cases
before retraining from scratch with a mixture of original and
corner case images. For each corner case, the last 3 seconds
are saved at 10 fps before the intervention by the safety
driver. In total, we get 1500 new frames. When using this
corner case data for retraining, we delete the same number
of frames from the original training dataset.
We selected 50 corner cases in connection with pedestri-
ans. Therefore, the inclusion of corner case scenes into the
training dataset significantly increases the average number
of pixels with the pedestrian class in the training data. To
establish a fair comparison of the efficiency of corner cases
as compared to a simple upsampling of the pedestrian class,
we created a third dataset that contains approximately the
same number of pixels per scene for the pedestrian class as
the dataset with the corner cases, see Table I.
V. EVALUATION AND RESULTS
All results in this section are averaged over 5 experiments
to obtain a better statistical validity. For testing purposes,
we generated 21 additional corner cases for validation. With
the same setup as before, we train the Fast-SCNN for
200 epochs on all three datasets and thereby obtain three
networks. Table II shows the evaluation of all three models
on the class pedestrian for the 21 safety critical test corner
cases. We see that adding corner cases to the training data
leads to an improvement in pedestrian detection in safety
critical situations, which can also be shown by an example in
Figure 4. There we see a situation with a pedestrian crossing
the road, with a slope directly behind him that seems to end
the road at the level of the horizon. Therefore, the networks
that did not have corner cases in the training data seem
to have problems with this situation, while the model with
corner cases detects the human much better. While training
the network using naive upsampling of pedestrians does not
have any positive effect on the classe’s IoU as compared
with the original training data, we achieve a gain in the IoU
by 2.19% when using the dataset containing corner cases.
In addition, the 3 models were tested on a dataset with a
natural distribution of pedestrians. Here it can be seen that
the model trained with corner cases does not perform as
well as the model with the same number of pedestrians. It
follows that the model performs better in critical situations,
while the models without corner cases perform less well. We
therefore demonstrated the benefits of our method to generate
corner cases, especially for safety critical situations. We were
also able to show that adding safety-critical corner cases
improves performance, so future datasets should include such
situations.
VI. CONCLUSION
This work presents an experimental setup for human
driving with the eyes of AI. We have designed a test rig
on which an ego-vehicle can be driven by 2 subjects in

the virtual world of CARLA. The semantic driver receives
the output of a semantic segmentation network in real-time,
based on which she or he is supposed to navigate in the
virtual world. The second driver takes the role of the driving
instructor and intervenes in dangerous driving situations
caused by misjudgements of the AI. We consider driver
interventions by the safety driver as safety-critical corner
cases which subsequently replaced part of the initial training
data. We were able to show that targeted data enrichment
with corner cases leads to improved pedestrian detection in
critical situations.
Future research projects include the use of some networks
of different quality, the world parameters such as the number
of pedestrians, vehicles, but also the weather will be changed
and accident scenarios are to be provoked so that the number
of corner cases can be artificially increased in test operation.
The intervention of the safety driver in the driving situation
will also be observed. To this end, criteria for measuring
human-machine interaction (HMI) will be developed to track,
for example, latency, attention, and intervention due to bore-
dom of the drivers.
ACKNOWLEDGMENT
The research leading to these results is funded by the
German Federal Ministry for Economic Affairs and Climate
Action within the project “KI Data Tooling – Methoden und
Werkzeuge f¨
ur das Generieren und Veredeln von Trainings-
, Validierungs- und Absicherungsdaten f ¨
ur KI-Funktionen
autonomer Fahrzeuge” under the grant number 19A20001O.
The authors thank the consortium for the successful cooper-
ation. We also thank Matthias Rottmann for his productive
support and Natalie Grabowsky and Ben Hamscher for
driving the streets of CARLA and capturing the corner cases.
REFERENCES
[1] A. Dosovitskiy et al., “CARLA: an open urban driving simulator,”
in 1st Annual Conference on Robot Learning, CoRL 2017, Mountain
View, California, USA, November 13-15, 2017, Proceedings, ser.
Proceedings of Machine Learning Research, vol. 78. PMLR,
2017, pp. 1–16. [Online]. Available: http://proceedings.mlr.press/v78/
dosovitskiy17a.html
[2] U. Chipengo, P. Krenz, and S. Carpenter, “From antenna design to high
fidelity, full physics automotive radar sensor corner case simulation,”
Modelling and Simulation in Engineering, vol. 2018, pp. 1–19, 12
2018.
[3] J.-A. Bolte et al., “Towards Corner Case Detection for Autonomous
Driving,” in 2019 IEEE Intelligent Vehicles Symposium, IV 2019,
Paris, France, June 9-12, 2019. IEEE, 2019, pp. 438–445. [Online].
Available: https://doi.org/10.1109/IVS.2019.8813817
[4] J. Breitenstein, J.-A. Term¨
ohlen, D. Lipinski, and Fingscheidt,
“Systematization of corner cases for visual perception in automated
driving,” in IEEE Intelligent Vehicles Symposium, IV 2020, Las
Vegas, NV, USA, October 19 - November 13, 2020. IEEE, 2020,
pp. 1257–1264. [Online]. Available: https://doi.org/10.1109/IV47402.
2020.9304789
[5] J. Breitenstein et al., “Corner cases for visual perception in
automated driving: Some guidance on detection approaches,”
CoRR, vol. abs/2102.05897, 2021. [Online]. Available: https:
//arxiv.org/abs/2102.05897
[6] F. Heidecker, J. Breitenstein, K. R¨
osch et al., “An Application-
Driven Conceptualization of Corner Cases for Perception in Highly
Automated Driving,” in 2021 IEEE Intelligent Vehicles Symposium
(IV), Nagoya, Japan, 2021.
[7] R. Chan et al., “SegmentMeIfYouCan: A Benchmark for Anomaly
Segmentation,” in Thirty-fifth Conference on Neural Information
Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2021.
[Online]. Available: http://arxiv.org/abs/2104.14812
[8] R. Chan, M. Rottmann, and H. Gottschalk, “Entropy maximization
and meta classification for out-of-distribution detection in semantic
segmentation,” in Proceedings of the IEEE/CVF International Confer-
ence on Computer Vision, 2021, pp. 5128–5137.
[9] F. M. Zanzotto, “Viewpoint: Human-in-the-loop artificial intelligence,”
in Discussion and Doctoral Consortium papers of AI*IA 2019 -
18th International Conference of the Italian Association for Artificial
Intelligence, Rende, Italy, November 19-22, 2019, ser. CEUR
Workshop Proceedings, vol. 2495. CEUR-WS.org, 2019, pp. 84–94.
[Online]. Available: http://ceur-ws.org/Vol-2495/paper10.pdf
[10] H. Thiruvengada, A. Tharanathan, and P. Derby, PerFECT:
An Automated Framework for Training on the Fly. London:
Springer London, 2011, pp. 221–238. [Online]. Available: https:
//doi.org/10.1007/978-0-85729-883-6 11
[11] M. Monarch, R. Munro, and R. Monarch, Human-in-the-Loop Ma-
chine Learning: Active Learning and Annotation for Human-centered
AI. Manning, 2021.
[12] X. Wu et al., “A survey of human-in-the-loop for machine
learning,” CoRR, vol. abs/2108.00941, 2021. [Online]. Available:
https://arxiv.org/abs/2108.00941
[13] J. Wu et al., “Human-in-the-loop deep reinforcement learning with
application to autonomous driving,” CoRR, vol. abs/2104.07246,
2021. [Online]. Available: https://arxiv.org/abs/2104.07246
[14] K. Driggs-Campbell, V. Shia, and R. Bajcsy, “Improved driver mod-
eling for human-in-the-loop vehicular control,” in 2015 IEEE Inter-
national Conference on Robotics and Automation (ICRA), 2015, pp.
1654–1661.
[15] A. E. G´
omez et al., “Driving simulator platform for development
and evaluation of safety and emergency systems,” CoRR, vol.
abs/1802.04104, 2018. [Online]. Available: http://arxiv.org/abs/1802.
04104
[16] G. Echeverria et al., “Modular open robots simulation engine: Morse,”
in 2011 IEEE International Conference on Robotics and Automation,
2011, pp. 46–51.
[17] P. Hang, Y. Zhang, and C. Lv, “Interacting with human drivers:
Human-like driving and decision making for autonomous vehicles,”
2022.
[18] H. Sun et al., “Corner case generation and analysis for safety
assessment of autonomous vehicles,” Transportation Research
Record, vol. 2675, no. 11, pp. 587–600, 2021. [Online]. Available:
https://doi.org/10.1177/03611981211018697
[19] B. Dodson and H. Schwab, Accelerated Testing: A Practitioner’s
Guide to Accelerated and Reliability Testing, ser. Knovel Library.
SAE International, 2006.
[20] W. Nelson, Accelerated Testing: Statistical Models, Test Plans, and
Data Analysis, ser. A Wiley-interscience publication. Wiley, 1990.
[21] W. Q. Meeker and L. A. Escobar, “A review of recent research and
current issues in accelerated testing,” International Statistical Review
/ Revue Internationale de Statistique, vol. 61, no. 1, pp. 147–168,
1993. [Online]. Available: http://www.jstor.org/stable/1403600
[22] Epic Games, “Unreal engine.” [Online]. Available: https://www.
unrealengine.com
[23] R. P. K. Poudel et al., “Fast-scnn: Fast semantic segmentation
network,” in 30th British Machine Vision Conference 2019, BMVC
2019, Cardiff, UK, September 9-12, 2019. BMVA Press, 2019, p.
289. [Online]. Available: https://bmvc2019.org/wp-content/uploads/
papers/0959-paper.pdf
[24] M. Cordts et al., “The cityscapes dataset for semantic urban scene
understanding,” in Proc. of the IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2016.
[25] A. Paszke et al., “Pytorch: An imperative style, high-performance
deep learning library,” in Advances in Neural Information
Processing Systems 32. Curran Associates, Inc., 2019,
pp. 8024–8035. [Online]. Available: http://papers.neurips.cc/paper/
9015-pytorch- an-imperative-style-high- performance-deep-learning- library.
pdf
[26] F. M. Favar`
oet al., “Examining accident reports involving autonomous
vehicles in california,” PLOS ONE, vol. 12, pp. 1–20, 09 2017.
[Online]. Available: https://doi.org/10.1371/journal.pone.0184952