Bayesian Network-Based Probabilistic Constraints for Safe Autonomous Driving in Occlusion Environments

This paper considers how to realize safe autonomous driving against hazards that lurk in occlusion, such as pedestrians darting out. Occlusion is a potential risk factor that cannot be directly detected by sensors and is one of the critical issues that need to be addressed to realize safe autonomous driving. However, if all dangerous scenarios resulting from occlusion are considered, the system may become overly conservative. One of the promising control schemes to guarantee the safety of traffic participants who dart out from the occlusion is based on a probabilistic framework that always maintains the safety of the vehicle with a certain probability. Although the control including probabilistic constraints achieves the control that is not too conservative, it has not been fully discussed how to derive reasonable probabilistic constraints for unknown occlusion environments. In addition, the computational complexity to calculate the probability must be reduced in order to be exploited for real-time control. Therefore, this paper proposes a method to calculate probability constraints for occlusion hazards based on a Bayesian network. In the proposed Bayesian network, prior information including the driving environment and features of occlusion are considered naturally. This property will not only drastically simplify the analysis of failure behaviors but also enable transparency in the exposed risks. The proposed method is confirmed to be computationally less expensive than the Monte Carlo method, and the usefulness of the control using the proposed probability constraints is verified by simulation.