Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Varied Realistic Autonomous Vehicle Collision Scenario Generation

Priisalu, Maria LU ; Paduraru, Ciprian and Smichisescu, Cristian (2023) 23nd Scandinavian Conference on Image Analysis, SCIA 2023 In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 13886 LNCS. p.354-372
Abstract

Recently there has been an increase in the number of available autonomous vehicle (AV) models. To evaluate and compare the safety of the various models the AVs need to be tested in several diverse safety-critical scenarios. We propose the Adversarial Test Case Generator (ATCG) that differently from previous test case generators allows for the generation of realistic collision scenarios with varied AV and pedestrian behaviour models, on varied scenes and with varied traffic density. Given a top-view image and the semantic segmentation of a traffic scene, the ATCG learns to place multiple AVs and goal-reaching pedestrians in the scene such that collisions occur. Pedestrians in previous multi-agent traffic scenario generation works are... (More)

Recently there has been an increase in the number of available autonomous vehicle (AV) models. To evaluate and compare the safety of the various models the AVs need to be tested in several diverse safety-critical scenarios. We propose the Adversarial Test Case Generator (ATCG) that differently from previous test case generators allows for the generation of realistic collision scenarios with varied AV and pedestrian behaviour models, on varied scenes and with varied traffic density. Given a top-view image and the semantic segmentation of a traffic scene, the ATCG learns to place multiple AVs and goal-reaching pedestrians in the scene such that collisions occur. Pedestrians in previous multi-agent traffic scenario generation works are confined to unrealistic behaviours such as seeking collisions with the AV or ignoring the AV. Although such scenarios with multiple suicidal pedestrians are collision prone it is unlikely in reality that all pedestrians act abnormally. In realistic collision scenarios the generated pedestrians’ behaviours must resemble real pedestrians. The ATCG is a team of Reinforcement Learning (RL) agents and can be easily extended with additional RL agents to produce more complex scenes allowing for advanced AVs to be tested.

(Less)
Please use this url to cite or link to this publication:
author
; and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
Autonomous Vehicle, AV Testing, Multi Agent Reinforcement Learning
host publication
Image Analysis - 23rd Scandinavian Conference, SCIA 2023, Proceedings
series title
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
editor
Gade, Rikke ; Felsberg, Michael and Kämäräinen, Joni-Kristian
volume
13886 LNCS
pages
19 pages
publisher
Springer Science and Business Media B.V.
conference name
23nd Scandinavian Conference on Image Analysis, SCIA 2023
conference location
Lapland, Finland
conference dates
2023-04-18 - 2023-04-21
external identifiers
  • scopus:85161378395
ISSN
1611-3349
0302-9743
ISBN
9783031314377
DOI
10.1007/978-3-031-31438-4_24
project
Modelling Pedestrians in Autonomous Vehicle Testing
language
English
LU publication?
yes
id
4e1dea78-b8f6-4d86-853b-ceb5e1479d8b
date added to LUP
2023-09-15 14:47:19
date last changed
2024-04-19 00:59:30
@inproceedings{4e1dea78-b8f6-4d86-853b-ceb5e1479d8b,
  abstract     = {{<p>Recently there has been an increase in the number of available autonomous vehicle (AV) models. To evaluate and compare the safety of the various models the AVs need to be tested in several diverse safety-critical scenarios. We propose the Adversarial Test Case Generator (ATCG) that differently from previous test case generators allows for the generation of realistic collision scenarios with varied AV and pedestrian behaviour models, on varied scenes and with varied traffic density. Given a top-view image and the semantic segmentation of a traffic scene, the ATCG learns to place multiple AVs and goal-reaching pedestrians in the scene such that collisions occur. Pedestrians in previous multi-agent traffic scenario generation works are confined to unrealistic behaviours such as seeking collisions with the AV or ignoring the AV. Although such scenarios with multiple suicidal pedestrians are collision prone it is unlikely in reality that all pedestrians act abnormally. In realistic collision scenarios the generated pedestrians’ behaviours must resemble real pedestrians. The ATCG is a team of Reinforcement Learning (RL) agents and can be easily extended with additional RL agents to produce more complex scenes allowing for advanced AVs to be tested.</p>}},
  author       = {{Priisalu, Maria and Paduraru, Ciprian and Smichisescu, Cristian}},
  booktitle    = {{Image Analysis - 23rd Scandinavian Conference, SCIA 2023, Proceedings}},
  editor       = {{Gade, Rikke and Felsberg, Michael and Kämäräinen, Joni-Kristian}},
  isbn         = {{9783031314377}},
  issn         = {{1611-3349}},
  keywords     = {{Autonomous Vehicle; AV Testing; Multi Agent Reinforcement Learning}},
  language     = {{eng}},
  pages        = {{354--372}},
  publisher    = {{Springer Science and Business Media B.V.}},
  series       = {{Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)}},
  title        = {{Varied Realistic Autonomous Vehicle Collision Scenario Generation}},
  url          = {{http://dx.doi.org/10.1007/978-3-031-31438-4_24}},
  doi          = {{10.1007/978-3-031-31438-4_24}},
  volume       = {{13886 LNCS}},
  year         = {{2023}},
}