Intelligent Assessment and Training of Hazard Perception for Human Drivers and Autonomous Vehicles Using Deep Reinforcement Learning and Hierarchical Clustering

Hazard perception (HP) is one of the few driving skills proven to reduce traffic accidents through training and testing. The driver licensing process in some countries includes assessing individuals through videos demonstrating hazards and measuring their reactions. Testing HP using a driving simulator has been shown to better approximate real-world driving and, therefore, be more effective. However, current methods need improvement for large-scale assessment and evaluation of autonomous vehicle HP abilities. This paper proposes a novel four-step procedure for training and assessing HP using deep reinforcement learning (DRL) and hierarchical clustering (HC). In the first step, an DRL agent is trained in an environment with a high collision probability and achieves superhuman performance compared with the subjects. The trained agent achieved 8.7% higher quickness and 3.5% longer traversal compared with the best human violation-free trial. Then, we segmented the scores into four levels using the HC method. The relationship between reward and three scenario-independent HP features—disparity, comfort, and haste—are identified through regression with a coefficient of determination of 0.94. Reviewing the subjects’ learning data averages, we determined that a learning period of 20 min is sufficient for the basic skills of HP in a driving simulator. The outcome is an intelligent, fair, accurate, and comprehensive system for jointly assessing and training HP for humans and autonomous vehicles without the drawbacks of conventional methods. It can be deployed at scale as a replacement for current approaches.