Lane-Changing Decisions Making for Autonomous Vehicles via Behavior Cloning and Decision Tree

The use of behavior cloning technology for autonomous driving decision-making has become a research hot spot at home and abroad. Existing research methods are limited to existing data sets. The data is single and scarce, and the robustness of the algorithm cannot be verified interactively. The accuracy and convergence speed of the algorithm are limited. In this paper, we divide the lane-changing decision-making process of autonomous driving vehicles into two layers. The upper layer uses the decision tree algorithm to select tasks, selects lane-changing task or lane keeping task according to road conditions, subdivides traffic scenes, and simplifies the data space for behavior cloning. The lower layer uses end-to-end behavior cloning and outputs the steering wheel angle and the throttle angle control amount. Firstly, build the lane changing scene on the CARLA simulator, collect and preprocess the driving data, then use attribute reduction to filter appropriate attributes, and use decision tree algorithm to extract driving rules, finally the lane-changing task and the lane-keeping task are learned through the behavior cloning algorithm respectively. The training results are analyzed and simulated in CARLA. The results show that the accuracy and convergence speed of the behavior cloning algorithm are improved after the division of decision tree algorithm, the lane changing task loss is reduced by 6.25%, and the lane keeping task loss is reduced by 9.38%.