Learning Driver-Specific Behavior for Overtaking: A Combined Learning Framework

Learning-based methods have gained increasing attention in the intelligent vehicle community for developing highly autonomous vehicles and advanced driving assistance systems (ADAS). However, traditional offline learning methods lack the ability to adapt to individual driving behavior. To overcome this limitation, a combined learning framework (CLF) based on the Natural Actor Critic (NAC) learning and general regression neural network (GRNN) is developed in this paper. GRNN can be trained offline based on the historical data, while NAC is carried out online. In this way, the general behavior learned by the offline module can be reused and adjusted by the online module to capture the driver-specific behavior. Driving data collected from human drivers through a driving simulator are used to test the proposed learning framework. The complex overtaking behavior is selected to formulate the learning problem and test scenarios. Experimental results show that the proposed system performs well on learning driver-specific behavior for overtaking, and compared with the Gaussian mixture model-maximum-a-posterior method, CLF shows a more flexible performance when newly-involved drivers are considered.