A new methodology for collision risk assessment of autonomous vehicles

Risk assessment methods of autonomous vehicles (AVs) have recently begun to treat the motion of the vehicles as dependent on the context of the traffic scene that the vehicle resides in. In most of the cases, Dynamic Bayesian Network (DBN) models are employed for interaction aware motion models (i.e. models that take inter-vehicle dependencies into account). However, communications between vehicles are assumed and the developed models require a lot of parameters to be tuned. Even with these requirements, current approaches cannot cope with traffic scenarios of high complexity. To overcome these limitations, the current study proposes a new methodology that integrates real-time collision prediction as studied by traffic engineers with an interaction-aware motion model for autonomous vehicles real-time risk assessment. Results from a random forest classifier for real-time collision prediction are used as an example for the estimation of probabilities required for the DBN model. It is shown that a well-calibrated collision prediction classifier can provide a supplementary hint to already developed interaction-aware motion models and enhance real-time risk assessment for autonomous vehicles.