基于驾驶特征学习的自动驾驶车辆换道轨迹规划

Lane Changing Trajectory Planning of Autonomous Vehicle Based on Driving Characteristic Learning

针对自动驾驶车辆换道过程中存在的车辆规划轨迹与人类驾驶员决策轨迹偏差较大问题,开发了一种基于驾驶员轨迹特征学习的换道轨迹规划算法。采集驾驶员换道轨迹曲线函数特征,在轨迹采样及成本优化相结合的轨迹规划基础上,采用最大熵逆强化学习策略迭代更新成本函数权重,并依据学习的成本函数筛选备选采样轨迹,生成反映驾驶员轨迹特征的自动驾驶车辆换道轨迹。试验结果表明,进行驾驶员特征学习后的换道轨迹基本包含在驾驶员换道轨迹区域内,且轨迹特征更为接近人类驾驶员换道轨迹特征,更能反映驾驶员主观感受。

Large deviation between vehicle planning trajectory and driver decision trajectory exists in the process of lane change for autonomous vehicles.To solve this problem,a lane change trajectory planning algorithm is developed based on learning trajectory feature.Based on the sampling and cost optimization combination of trajectory planning,the algorithm collects the driver’s lane changing trajectory function characteristics.By means of the maximum entropy inverse reinforcement learning,cost function weight is updated iteratively.According to the achieved cost function,the alternative sampling paths are designated to generate lane changing trajectory of autonomous vehicles which reflect the characteristics of drivers’trajectories.The experimental results show that the lane changing trajectory after learning of drivers’characteristics are incorporated in the lane changing trajectory area of the driver.The trajectory’s features are more similar to the real lane changing trajectory’s features of the driver,and can reflect driver’s subjective feeling.