面向未知域场景的车辆轨迹预测模型

Vehicle trajectory prediction model for unseen domain scenarios

自动驾驶技术随着科技革新迎来蓬勃发展,轨迹预测已成为智能汽车软件系统不可或缺的关键组成部分。为了解决传统车辆轨迹预测模型中存在的泛化能力不足的问题,提出一种基于泛化终点预测和地图场景的车辆轨迹预测方法。该方法采用基于不变风险最小化的条件变分自编码器生成轨迹终点,并结合时序网络编码的地图场景特征,提升了模型预测未知域数据的准确率。在交互式道路场景数据集INTERACTION上的实验结果证明该模型具有良好的泛化性能。本方法与效果最好的方法REx相比1、2、3 s处的mADE值(越小越好)分别下降0%、36.59%、50.68%,在未知测试域的预测轨迹准确度得到显著提升。

With the rapid development of autonomous driving technology,trajectory prediction has become an essential component of smart car software systems.To address the limitations in the generalization of traditional vehicle trajectory prediction models,this study proposes a vehicle trajectory prediction method based on generalized endpoint prediction and vector maps.This method employs a conditional variational autoencoder based on invariant risk minimization to generate trajectory endpoints.The accuracy of the model in predicting unseen domains is improved by integrating map scene features encoded by a time-series network.Experiments were conducted using the interaction dataset,which includes interactive driving scenarios.Experimental results showed that,compared with the best-performing state-ofthe-art(SOTA)method REx,the mADE values(where lower is better)decreased by 0%,36.59%,and 50.68%at 1,2,and 3 s,respectively.The accuracy of the predicted trajectories in the unseen test domain was significantly improved.