Varied Realistic Autonomous Vehicle Collision Scenario Generation
(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)
- author
- Priisalu, Maria LU ; Paduraru, Ciprian and Smichisescu, Cristian
- organization
- publishing date
- 2023
- 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}}, }