适用于多车交互场景的车辆轨迹预测模型  

Vehicle trajectory prediction model for multi-vehicle interaction scenario

作　　者：黄玲[1,2] 崔躜 游峰 洪佩鑫[1] 钟浩川 曾译萱 HUANG Ling;CUI Zuan;YOU Feng;HONG Pei-xin;ZHONG Hao-chuan;ZENG Yi-xuan(School of Civil Engineering and Transportation,South China University of Technology,Guangzhou 510640,China;Jiangsu Key Laboratory of Urban ITS,Southeast University,Nanjing 210096,China;Pazhou Lab,Guangzhou 510330,China)

机构地区：[1]华南理工大学土木与交通学院,广州510640 [2]东南大学现代城市交通技术江苏高校协同创新中心,南京210096 [3]人工智能与数字经济广东省实验室,广州510330

出　　处：《吉林大学学报（工学版）》2024年第5期1188-1195,共8页Journal of Jilin University:Engineering and Technology Edition

基　　金：广东省基础与应用基础研究基金项目(2023A1515010742)。

摘　　要：提出了一种具有动态交互感知池化层的多长短期记忆神经网络(DIP-LSTM)模型结构,使得场景中相邻的车辆通过池化(Pooling)共享各自LSTM网络隐藏态,获取历史轨迹特征,进而实现自车与周围多车的时间空间关系的交互性建模,并输出车辆未来的预测轨迹。使用美国的NGSIM和德国的High-D自然驾驶车辆轨迹数据集对模型进行训练与测试,并对模型的精度、鲁棒性和迁移性(普适性)进行验证。研究结果表明:与传统模型的预测方法相比,考虑多车交互信息的DIP-LSTM网络的预测方法在预测精度与长时域预测上具有优势,且模型具有良好的迁移性和鲁棒性,显著提高了车辆轨迹预测模型的实用性和普适性。A DIP-LSTM model with dynamic interactive poling layer is proposed, which enables neighboring vehicles to share hidden states of LSTM network by pooling to get the characteristic of historical trajectory, and then realizes interactive modeling of time-space relationship between target vehicle and surrounding vehicles. NGSIM from USA and High-D from Germany are used to train and test the model, and the accuracy, robustness and transferability of the model are verified. The results show that compared with the traditional model prediction method, the DIP-LSTM network show advantages in prediction accuracy and long-time prediction considering multi-vehicle interactive information, and the model has good transferability and robustness, which significantly improves the practicability and universality of intelligent vehicle trajectory prediction model.

关 键 词：车辆工程 自动驾驶 轨迹预测 多车交互 长短期记忆神经网络 

分 类 号：U495[交通运输工程—交通运输规划与管理]