TY - JOUR
T1 - Analyzing scenario criticality and rider's intervention behavior during high-level autonomous driving
T2 - A VR-enabled approach and empirical insights
AU - Xu, Zheng
AU - Zheng, Nan
AU - Lv, Yisheng
AU - Fang, Yihai
AU - Vu, Hai L.
N1 - Funding Information:
The authors sincerely express their deepest gratitude to each participant who has contributed to this study. Your generous investment of time, patience, encouragement, and support have been invaluable to the success of our research.
Publisher Copyright:
© 2023 The Author(s)
PY - 2024/1
Y1 - 2024/1
N2 - The safety of autonomous driving systems (ADS) has been attracting extensive attentions of research and practice worldwide. One major concern on the existing ADS is the failure to identify and react to certain situations, resulting in either critical fatalities or urgent intervention by riders. There is an increasing need to understand the characteristics of such situations and scenarios. This paper sheds light on this direction by defining and measuring the “criticality” of typical driving scenarios during high-level autonomous driving in a mixed traffic environment. Autonomous driving simulations and human-in-the-loop experiments are conducted using a virtual reality enabled simulation platform. We correlate ADS-involved traffic conflicts with environmental features via autonomous driving simulations from over 10,000 testing scenarios and statistically evaluate the criticality level. We collect people's behavioral data from 900 representative scenarios to observe the differences in reactions between ADS and human riders when facing the same situation. Our results demonstrate apparent differences in criticality perceptions between humans and ADS. The risk perception and action by ADS are mainly governed by on-road dynamics according to other road users’ behavior, whereas human riders are affected by both on-road dynamics and environmental features such as traffic, lighting, and weather conditions. To gain more insights, we analyze the intervention behavior performed by the riders. Interestingly, riders tend to intervene more under seemingly “uncritical” scenarios with low criticality values compared to scenarios with high criticality values. These are useful insights for rethinking and enhancing the intelligence of the current ADS. Our work advocates for the creation of a library of critical scenarios and emphasizes the importance of incorporating human perceptions into the system's decision-making process. Furthermore, our results suggest that the adoption of the ADS depends on the role of the riders. People tend to intervene less and thus show more trust while “feeling” as a passenger.
AB - The safety of autonomous driving systems (ADS) has been attracting extensive attentions of research and practice worldwide. One major concern on the existing ADS is the failure to identify and react to certain situations, resulting in either critical fatalities or urgent intervention by riders. There is an increasing need to understand the characteristics of such situations and scenarios. This paper sheds light on this direction by defining and measuring the “criticality” of typical driving scenarios during high-level autonomous driving in a mixed traffic environment. Autonomous driving simulations and human-in-the-loop experiments are conducted using a virtual reality enabled simulation platform. We correlate ADS-involved traffic conflicts with environmental features via autonomous driving simulations from over 10,000 testing scenarios and statistically evaluate the criticality level. We collect people's behavioral data from 900 representative scenarios to observe the differences in reactions between ADS and human riders when facing the same situation. Our results demonstrate apparent differences in criticality perceptions between humans and ADS. The risk perception and action by ADS are mainly governed by on-road dynamics according to other road users’ behavior, whereas human riders are affected by both on-road dynamics and environmental features such as traffic, lighting, and weather conditions. To gain more insights, we analyze the intervention behavior performed by the riders. Interestingly, riders tend to intervene more under seemingly “uncritical” scenarios with low criticality values compared to scenarios with high criticality values. These are useful insights for rethinking and enhancing the intelligence of the current ADS. Our work advocates for the creation of a library of critical scenarios and emphasizes the importance of incorporating human perceptions into the system's decision-making process. Furthermore, our results suggest that the adoption of the ADS depends on the role of the riders. People tend to intervene less and thus show more trust while “feeling” as a passenger.
KW - Autonomous driving system
KW - Behavioral adaptation
KW - Rider intervention
KW - Scenario criticality
KW - Virtual reality
UR - http://www.scopus.com/inward/record.url?scp=85183740078&partnerID=8YFLogxK
U2 - 10.1016/j.trc.2023.104451
DO - 10.1016/j.trc.2023.104451
M3 - Article
AN - SCOPUS:85183740078
SN - 0968-090X
VL - 158
JO - Transportation Research Part C: Emerging Technologies
JF - Transportation Research Part C: Emerging Technologies
M1 - 104451
ER -