Multi-objective Optimization of Lane-changing Trajectories for Autonomous Vehicles Based on Deep Reinforcement Learning

Abstract

Autonomous driving technology has been largely due to its driving efficiency and driving safety. Optimizing the lane-changing trajectory planning accuracy is particularly important. To solve the low accuracy and insufficient multi-objective collaborative optimization, this study combines the temporal convolution network, spatial multi-head attention mechanism and Transformer encoder-decoder to build a trajectory prediction module, and integrates the driving risk field into the multi-objective reward function. A multi-objective optimized lane-changing trajectory planning method for autonomous vehicles based on the improved depth deterministic strategy gradient is proposed. The trajectory prediction module performed a performance comparison analysis with other methods. The prediction root mean square errors under going straight, turning left, and turning right were 1.59m, 1.19m, and 1.27m. Subsequently, the proposed method was compared with other methods. The results found that the obstacle avoidance success rate and target advancement success rate were 98.92% and 98.26%, respectively, which were both better than those of the comparison method. In addition, the analysis results in the lane-changing overtaking and lane-changing obstacle avoidance scenarios showed that the vehicle lane-changing correction was more timely, the vehicle speed fluctuation was smaller, and the front wheel angle control was smoother, all of which were better than those of the comparison method. The lane-changing trajectory planning method effectively improves the trajectory prediction accuracy and planning performance, and can provide a theoretical basis for related research fields.