基于有向图的强化学习自动驾驶轨迹预测

Reinforcement Learning Autonomous Driving Trajectory Prediction Based on Directed Graph

轨迹预测作为自动驾驶中的重要组成部分,旨在对车辆进行行驶估计,以便车辆根据行驶估计进行路径规划,从而做出安全准确的决策。首先,为提升车辆轨迹预测精度,采用有向图方法构建高清驾驶场景地图,有向图方法将地图信息矢量化,以便有效提取地图拓扑结构;其次,采用生成对抗模仿学习(GAIL)通过生成器与判别器的对抗博弈学习数据集驾驶策略,从而根据当前状态采取对应驾驶行为;最后,通过采样遍历得到多模态预测轨迹方案。在nuScenes运动预测数据集上进行仿真,量化结果显示相比于其他方法,K=5时,最小最终位移误差MinFDE_(5)提高了10.8%;K=10时,最小最终位移误差MinFDE_(10)提高了17.53%,最小平均位移误差MinADE_(10)提高了9.52%,失误率MissRate_(10)减少了28.26%。评估结果表明:生成的轨迹多模态符合场景基本结构,且准确度得到提高。

As an important part of autonomous driving,trajectory prediction aimed to estimate forcast the vehicle′s driving path,so that the vehicle could make path planning according to the driving estimation,so as to make safe and accurate decisions.Firstly,in order to improve the accuracy of vehicle trajectory prediction,the directed graph method was used to construct a high-definition driving scene map,and the directed graph method vectorized the map information to effectively extract the map topology.Secondly,GAIL was used to learn the driving strategy of the dataset through the confrontation game between the generator and the discriminator,so as to adopt the corresponding driving behavior according to the current state.Finally,the multimodal prediction trajectory scheme was obtained by sampling traversal.Simulation was carried out on the nuScenes motion prediction dataset.The quantitative results showed that compared with other methods,when K=5,the minimum final displacement error MinFDE_(5) was increased by_(10).8%;when K=10,the minimum fianl displacement error MinFDE_(10) increased by 17.53%,the minimum average displacement error MinADE_(10) increased by 9.52%,and the error rate MissRate_(10) decreased by 28.26%.The evaluation showed that the generated trajectories were multimodal,could conform to the basic structure of the scene,and have with improved accuracy.