多行人轨迹多视角三维仿真视频学习预测法  

Multi-pedestrian trajectory prediction method based on multi-view 3D simulation video learning

作　　者：曹兴文 郑宏伟[1,2] 刘英 吴孟泉[3] 王灵玥 包安明 陈曦 CAO Xingwen;ZHENG Hongwei;LIU Ying;WU Mengquan;WANG Lingyue;BAO Anming;CHEN Xi(State Key Laboratory of Desert and Oasis Ecology,Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,Urumqi 830011,China;College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China;School of Resources and Environmental Engineering,Ludong University,Yantai 264025,China)

机构地区：[1]中国科学院新疆生态与地理研究所荒漠与绿洲生态国家重点实验室,新疆乌鲁木齐830011 [2]中国科学院大学资源与环境学院,北京100049 [3]鲁东大学资源与环境工程学院,山东烟台264025

出　　处：《测绘学报》2023年第9期1595-1608,共14页Acta Geodaetica et Cartographica Sinica

基　　金：新疆维吾尔自治区重点研发专项(2022B03001-3,2022B03021-3);国家自然科学基金(NSFCU-U1803120);中国科学院“西部之光”人才培养计划(2021-XBQNXZ-012);国家自然科学基金(42071385)。

摘　　要：多行人轨迹预测是城市地理信息系统与智能交通融合的关键因素之一。针对现有行人轨迹预测方法训练数据量不足、标注难、对多视角场景轨迹预测精度低等问题,本文提出一种多行人轨迹多视角三维仿真视频学习预测方法。首先,通过仿真模拟器生成所需多视角行人轨迹标注数据,利用凸函数组合原始视角对抗轨迹和选定的多视角轨迹,生成增强对抗轨迹,接着使用检测跟踪算法对行人特征信息进行编码;然后,将增强轨迹和编码信息作为图注意力循环神经网络的特征输入,对行人交互信息建模;最后,通过位置解码器对行人轨迹进行解码并提取行人运动特征,完成多行人轨迹预测。本文方法在ETH/UCY固定视角数据集上的ADE和FDE精度分别为0.41和0.82。在ActEV/VIRAT和Argoverse多视角数据集上的ADE精度为17.74和65.4,FDE精度为34.96和172.8。Multi-pedestrian trajectory prediction is one of the key factors in integrating urban geographic information system and intelligent transportation.To address the problems of insufficient training data,difficult labeling,and low accuracy of pedestrian trajectory prediction in multi-view scenes for existing methods,we propose a novel multi-pedestrian trajectory prediction method based on multi-view 3D simulation video learning.First,a simulation simulator is used to generate the required multi-view pedestrian trajectory annotation data.Then,we mix up the trajectory of the selected view and the adversarial trajectory by a convex combination function to generate the enhanced adversarial trajectory.Next,an advanced detection and tracking algorithm is used to encode and track pedestrian appearance information.Furthermore,the enhanced trajectory and coding information are used as the feature input of a graph attention recurrent neural network to model pedestrian interaction.Finally,the pedestrian trajectory is decoded by a position decoder to extract pedestrian motion characteristics,and multi-pedestrian trajectory prediction is completed.The ADE and FDE accuracies of our method on the ETH/UCY fixed-view dataset are 0.41 and 0.82,respectively.The ADE accuracy on the ActEV/VIRAT and Argoverse multi-view datasets is 17.74 and 65.4,and the FDE accuracy is 34.96 and 172.8.

关 键 词：三维仿真 深度学习 轨迹预测 目标跟踪 城市地理信息系统 

分 类 号：P208[天文地球—地图制图学与地理信息工程]