基于GAT和Transformer的车辆行为预测  

Vehicle Behavior Prediction Based on GAT and Transformers

作　　者：王昀 蔡英[1] 范艳芳[1] 柳军杰 张哲 WANG Yun;CAI Ying;FAN Yanfang;LIU Junjie;ZHANG Zhe(College of Computer,Beijing Information Science&Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学计算机学院,北京100101

出　　处：《小型微型计算机系统》2025年第1期23-32,共10页Journal of Chinese Computer Systems

基　　金：北京市自然科学基金-海淀原始创新联合基金项目(L192023)资助。

摘　　要：车辆行为预测可以辅助自动驾驶系统进行决策,提高自动驾驶的安全性和效率.然而在不同道路场景下,周围交通参与者(如汽车、自行车及行人等)之间动态的变化会导致预测车辆位置信息存在较大误差,这可能使自动驾驶车辆无法及时采取避让或紧急制动等措施.本文旨在针对结构化和非结构化道路场景,构建交通参与者之间的动态交互时空图,并运用深度学习技术设计一种基于GAT和Transformer的GAN模型,用于车辆行为预测.GAT被用于学习不同参与者之间的相关性和交互规律,而Transformer被用来提取交通参与者运动状态信息的时序特征.分别在NGSIM和ApolloScape数据集进行仿真实验.结果表明,本文模型在长时域的预测表现出更高的精度,同时还具备更轻量级的特点.Vehicle behavior prediction can assist autonomous driving systems in making decisions,thereby enhancing the safety and efficiency of autonomous driving.However,in different road scenarios,dynamic changes among surrounding traffic participants,such as cars,bicycles,and pedestrians,can lead to significant errors in predicting vehicle position information.This may prevent autonomous vehicles from taking timely evasive or emergency braking actions.This paper aims to construct dynamic spatiotemporal graphs of interactions among traffic participants for structured and unstructured road scenarios.It utilizes deep learning techniques to design a GAN model based on GAT(Graph Attention Network)and Transformer for vehicle behavior prediction.GAT is employed to learn the correlations and interaction patterns among different participants,while Transformer is used to extract temporal features of traffic participants′motion states.Simulation experiments are conducted on the NGSIM and ApolloScape datasets.The results demonstrate that our model exhibits higher accuracy in long-term predictions while maintaining a more lightweight profile.

关 键 词：自动驾驶 车辆行为预测 深度学习 图注意力网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]