Road traffic safety assessment in self-driving vehicles based on time-to-collision with motion orientation

Abstract

Traffic conflict analysis based on Surrogate Safety Measures (SSMs) helps to estimate the risk level of an ego-vehicle interacting with other road users. Nonetheless, risk assessment for autonomous vehicles (AVs) is still incipient, given that most of the AVs are currently prototypes and current SSMs do not directly apply to autonomous driving styles. Therefore, to assess and quantify the potential risk arising from AV interactions with other road users, this study introduces the TTCmo (Time-to-Collision with motion orientation), a metric that considers the yaw angle of conflicting objects. In fact, the yaw angle represents the orientation of the other road users and objects detected by the AV sensors, enabling a better identification of potential risk events from changes in the motion orientation and position through the geometric analysis of the boundaries for each detected object. Using the 3D object detection data annotations available from the publicly available AV datasets nuScenes and Lyft5 and the TTCmo metric, we find that at least 8% of the interactions with objects detected around the AV present some risk level. This is meaningful, since it is possible to reduce the proportion of data analyzed by up to 60% when replacing regular TTC by our improved TTC computation.