A Bayesian Network Based Approach Toward An Anticipatory Safety Reasoning System Autonomous Vehicle Copilot

Future ground vehicle transportation is expected to rely heavily on autonomous mobility. However, the technical progress required to ensure a completely safe autonomous vehicle for unlimited roadway use, and reliable ways to measure its safety, is behind expectations. It is believed that a research breakthrough is required to address this gap. This dissertation defines a novel method for addressing on-road autonomous vehicle safety, explicitly focusing on unsignalized intersections. A method is described to generate an anticipatory safety copilot to assist the autonomous system with motion decisions by combining data collected from global online sources and the local autonomous vehicle sensors. This anticipatory copilot reasons about the environment around the autonomous vehicle and projects the vehicle's real-time motion intent forward into a projected future version of the environment created via features from the combined local and global source information. Based on this processed information, the copilot anticipates the probabilistic success of the autonomous vehicle safely executing its intended action.