Humanlike Behavior Generation in Urban Environment Based on Learning-Based Potentials With a Low-Cost Lane Graph

It is crucial to understand the surrounding cars with respect to the road context and interact with them harmoniously for the success of autonomous cars used in the mixed urban traffic. In this paper, a vision-based approach is proposed to implement the humanlike autonomous driving function along a predefined lane-level route in the complex urban environment with daily traffic. At first, the surrounding cars are located in the lane level by a deep neural network based detector with a low-cost lane graph. Subsequently, a Bayesian network is employed to classify the detected cars into six categories based on their states of operation, i.e., leader car, parked car, tail-end car, exiting car, merging car, and other car. Finally, a hybrid potential map, consisting of a trajectory-induction potential and a risk-prevention potential, is constructed for each of the cars according to their categories, which will be combined to be used for generating an appropriate behavior. Particularly, both of the trajectory-induction potentials and the risk-prevention potentials are learned from naturalistic driving data of the same situations in the daily urban traffic to encode the human driving skills and experiences. Therefore, the behavior generated based on the proposed learning-based potentials is close to a humanlike performance, which is important to achieve harmony in the mixed traffic. Experimental results in various typical but challenging urban traffic scenes have substantiated the effectiveness of the proposed system.