考虑驾驶员特性的自学习换道轨迹规划系统

Self-learning Lane-change Trajectory Planning System with Driver Characteristics

为更好地实现个性化驾驶,本文中提出了一种集成驾驶员特性辨识的自学习换道轨迹规划系统。首先,在高斯分布中引入驾驶员特性系数Jc和驾驶员反应与操作时间td,建立了个性化换道轨迹规划模型,并通过DTW算法对实际轨迹和拟合轨迹进行匹配。之后,基于采集的驾驶员换道轨迹进行AP聚类,离线标定Jc和td共性化值,同时获得30名驾驶员的标签,将其驾驶特性分为舒适、一般和运动型。然后,将自由驾驶数据进行特征工况的提取,并基于长短期记忆网络(LSTM)搭建驾驶员特性在线辨识模型进行训练。最后,选取15名驾驶员进行实车验证,系统实时提取特征工况,然后基于辨识结果在线调整Jc和td,并不断更新拟合轨迹。实验结束后,其中14名驾驶员的实际轨迹与拟合轨迹平方欧氏距离小于1,拟合正确率达93.3%。因此,该系统能够良好地复现真人换道轨迹。

In order to better realize personalized driving,this paper proposes a self-learning lane-change trajectory planning system integrating driver characteristics identification.Firstly,this paper introduces the driver characteristic coefficient Jc and the driver response and operation time td into the Gaussian distribution to establish a personalized lane-change trajectory planning model,and matches the actual trajectory with the fitting trajectory through the DTW algorithm.After that,AP clustering is carried out based on the collected driver lane-change trajectories and the general values of Jc and td are calibrated offline.At the same time,the labels of 30 drivers are obtained,and their driving characteristics are divided into comfort,normal and sport type.Then,the free driving data are extracted for characteristic working conditions,and an online identification model of driver characteristics is built based on long-term and short-term memory network(LSTM).Finally,15 drivers are selected for real-car verification.The system extracts characteristic conditions in real time and then adjusted Jc and td online based on the recognition results,and continuously updates the fitting trajectory.After the experiment,the squared Euclidean distance between the actual trajectory and the fitting trajectory of the 14 drivers is less than 1,with the fitting accuracy reaching 93.3%.Therefore,the system can reproduce the trajectory of real lane-change well.