考虑货车因素图注意力网络的自动驾驶车辆轨迹预测方法  

Vehicle Trajectory Prediction Based on Graph Attention Network Considering Truck Factors

作　　者：罗懿斓 陈垦 李志斌[1] 刘攀[1] 杨洋 谭屈山 LUO Yian;CHEN Ken;LI Zhibin;LIU Pan;YANG Yang;TAN Qushan(School of Transportation Southeast University,Nanjing 211189,China;Sichuan Digital Transportation Technology Co.,Ltd,Chengdu 610041,China)

机构地区：[1]东南大学交通学院,南京211189 [2]四川数字交通科技股份有限公司,成都610041

出　　处：《交通与运输》2024年第6期49-55,共7页Traffic & Transportation

基　　金：新一代人工智能国家科技重大专项(2022ZD0115600)。

摘　　要：为提高车辆自动驾驶的安全性并减少拥堵的可能性,创新性地考虑货车因素对车辆交互感知的影响,提出一种基于图注意力网络的车辆自动驾驶轨迹预测算法。该算法首先采用编码器-解码器架构,其中Encoder利用全连接层和LSTM层,将输入的低维特征转换为高维表示,并初步提取轨迹的时序特征;其次,以GAT作为Encoder的输出,通过图注意力网络动态捕捉车辆间的空间关系,增强了模型对复杂交通环境的感知能力;最后,利用GAT的结果,结合一个全连接层和一个可训练权重矩阵进行驾驶意图识别,并将识别结果与GAT输出串联,再次输入LSTM网络以生成最终轨迹预测。通过在NGSIM数据集上的实验验证,数值结果表明,考虑货车因素的模型在轨迹预测结果上得到更小的均方根误差,显著提升了预测精度。To enhance the safety of autonomous vehicle driving and mitigate the potential for congestion,this study innovatively considers the impact of truck factors on vehicle interactive perception and proposes an autonomous driving trajectory prediction algorithm based on Graph Attention Networks(GAT).The algorithm initially employs an encoder-decoder architecture,wherein the Encoder leverages fully connected layers and Long Short-Term Memory(LSTM)layers to transform low-dimensional input features into high-dimensional representations,thereby preliminarily extracting the temporal features of the trajectory.Subsequently,utilizing GAT as the Encoder's output,the algorithm dynamically captures the spatial relationships between vehicles through the graph attention network,thereby enhancing the model's perceptive capabilities in complex traffic environments.Finally,the GAT results are combined with a fully connected layer and a trainable weight matrix for driving intention recognition,and the recognition results are concatenated with the GAT output and re-input into the LSTM network to generate the final trajectory prediction.Experimental validation on the NGSIM dataset demonstrates that the model,which accounts for truck factors,achieves a smaller root mean square error in trajectory prediction results,significantly improving predictive accuracy.

关 键 词：图注意力网络 货车因素 动态交互 轨迹预测 智能驾驶 

分 类 号：U491[交通运输工程—交通运输规划与管理]