Trajectory Prediction Network of Autonomous Vehicles With Fusion of Historical Interactive Features
Zhiqiang Zuo, Xinyu Wang, Songlin Guo, Zhengxuan Liu, Zheng Li, Yijing Wang
Abstract
In order to improve driving safety and achieve more accurate decisions and planning, autonomous vehicles are required to predict the future trajectories of the surrounding vehicles. In this article, a novel trajectory prediction network with the fusion of historical interactive features is proposed. The encoder for interactive features based on the multi-headed attention mechanism extracts the historical interaction between the target vehicle and its neighbouring vehicles quickly and comprehensively. Further, the multi-modal factorized bilinear (MFB) pooling achieves full fusion of historical trajectory features and interactive features. Finally, the distribution of future trajectories for the target vehicle is output by the multi-modal recognition module and the trajectory generation module. During the training process, K-Means is used to reclassify the lateral classes of vehicle trajectories to solve the problem of data imbalance in the NGSIM dataset. Several quantitative and qualitative experiments illustrate the superiority of the network proposed in this article, and the influencing factors in the network are also interpreted through ablation experiments.