Prediction of sensor damage in automated vehicle involved collisions using parametric and non-parametric approaches

Abstract

Sensors are key components of autonomous vehicles (AVs) as they are used to detect other objects to avoid collisions. Thus, sensor damages should be avoided for AVs to operate properly. Therefore, this study explored factors associated with sensor damage in AV-involved crashes using Text Network Analysis (TNA) and Bayesian Networks (BNs) using 276 AV-involved crashes that occurred between 2017 and 2021 in California. It was found that 21% of crashes involved sensor damage whereby most crashes occurred during morning peak hours. TNA results portrayed patterns of keywords associated with sensor damage from crash narratives. On the other hand, BN results revealed that the side other than rear plays a great role in the likelihood of sensor damage. Further, weather condition, surface condition, and movements of both AV and conventional vehicles during the collision also play a significant role in the likelihood of sensor damage. In light of the study’s findings, suggestions for enhanced designs aimed at optimising performance under adverse weather conditions, particularly in wet or unclear environments, are put forth. Additionally, the study advocates for a strategic reconsideration of sensor placement to mitigate potential damage in the event of a crash.