基于时空图联合关系路径的行人轨迹预测框架  

A New Pedestrian Trajectory Prediction Framework Based on Spatio-Temporal Graph with Joint Relationship Path

作　　者：孙科 鄢府 范勇强 牟利平 谢海兵 SUN Ke;YAN Fu;FAN Yongqiang;MU Liping;XIE Haibing(Digital Luzhou Industrial Investment Group Co.,Ltd.,Luzhou 646004,China;Sichuan JKC Geographical Information Technologies Co.,Ltd.,Chengdu 610095,China;Chengdu Ecological Environment Digital Intelligent Governance Center,Chengdu 610015,China;Sichuan Tobacco Corporation Luzhou Branch,Luzhou 646004,China;The People s Liberation Army Joint Logistic Support Force,Wuhan 430013,China)

机构地区：[1]数字泸州产业投资集团有限公司,四川泸州646004 [2]四川省金科成地理信息技术有限公司,四川成都610095 [3]成都市生态环境数智治理中心,四川成都610015 [4]四川省烟草公司泸州市公司,四川泸州646004 [5]中国人民解放军联勤保障部队,湖北武汉430013

出　　处：《无线电工程》2023年第2期281-289,共9页Radio Engineering

基　　金：四川省科技计划资助(2021JDJQ0021,2022YFG0186)。

摘　　要：时空图被广泛应用于行人轨迹预测等时间序列任务中,如何更精确地捕捉不同时间段的轨迹位置信息以及更充分地利用时空图的结构信息对于轨迹预测至关重要。传统的轨迹预测方法规则复杂、约束性强、可扩展性较差,往往只能应用于特定领域。基于学习的轨迹预测方法不依赖于专家经验的物理规则,根据观察的轨迹数据来学习不同时间段各个空间位置之间的变化规则。基于学习的方法存在一定局限性,如没有充分利用时空图的结构信息,导致轨迹预测模型的性能下降。针对上述问题,提出了一种新型基于时空图联合关系路径的行人轨迹预测框架(Spatio-Temporal Graphs with Relationship Path Trajectory Prediction Framework, STRP-TPF)。STRP-TPF主要包括EdgeRNN和NodeRNN模型。STRP-TPF基于时空图构建关系路径,基于关系路径构建因子图;构建EdgeRNN和NodeRNN模型,并将因子图作为输入;输出下一时刻行人的位置,并且预测完整的行人轨迹。STRP-TPF利用关系路径能够准确捕捉时空图的结构信息,充分学习行人在不同时间和空间点的轨迹关系。大量实验结果表明,在ETH和UCY数据集上,STRP-TPF的整体性能均优于目前最先进的方法。在平均位移误差和最终位移误差方面,STRP-TPF比目前最先进方法低32.6%和37.7%。STRP-TPF的预测轨迹能够更准确地匹配真实轨迹。Spatio-temporal graphs are widely used in time-series tasks such as pedestrian trajectory prediction. How to accurately capture the trajectory position information in different time periods and utilize the structure information of spatio-temporal graphs is crucial for trajectory prediction. Traditional trajectory prediction methods have complex rules, strong constraints and poor scalability, and can only be applied to specific areas. In recent years, learning-based trajectory prediction methods do not rely on the physical rules of expert experience, and learn the variation rules between various spatial locations in different time periods according to the observed trajectory data. However, learning-based methods also have certain limitations, such as not fully utilizing the structure information of the spatio-temporal graph, which leads to the degradation of the performance of the trajectory prediction model. To cope with the above problems, a new pedestrian trajectory prediction framework based on spatio-temporal graph with relationship path called Spatio-Temporal Graphs with Relationship Path Trajectory Prediction Framework(STRP-TPF) is proposed. STRP-TPF mainly includes EdgeRNN and NodeRNN models. Firstly, STRP-TPF constructs a relationship path based on the spatio-temporal graph;secondly, STRP-TPF constructs a factor graph based on the relationship path;then, the EdgeRNN and NodeRNN models are constructed, and the factor graph is used as input data;finally, the location of pedestrian at the next moment is output, and the complete pedestrian trajectory is predicted. By using relationship path, STRP-TPF can accurately capture the structure information of the spatio-temporal graph, and fully learn the trajectory relationship of pedestrians at different time and space points. The experimental results show that, the overall performance of STRP-TPF is better than the state-of-the-art methods on the ETH and UCY datasets. In terms of average displacement error and final displacement error, STRP-TPF is 32.6% and 37.7%

关 键 词：时空图 轨迹预测 关系路径 因子图 平均位移误差 最终位移误差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]