基于时空图的行人轨迹预测  

Pedestrian trajectory prediction method based on spatial-temporal graph

作　　者：鹿雷 丛屾 韩纪龙 王书生 LU Lei;CONG Shen;HAN Jilong;WANG Shusheng(College of Electronic Engineering,Heilongjiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080

出　　处：《黑龙江大学自然科学学报》2024年第5期519-525,共7页Journal of Natural Science of Heilongjiang University

基　　金：国家自然科学基金资助项目(62373137)。

摘　　要：在自动驾驶技术中,行人轨迹预测的结果往往会影响到自动驾驶的安全性。行人轨迹预测技术目前面临着在实际场景中应用时与他人的交互问题,需要在预测轨迹的同时考虑社会交互性与逻辑自洽。因此,在之前的预测模型上做改进,得到了更好的预测效果。具体思路就是,首先用一个长短时记忆网络(Long short term memory network,LSTM)模型建立行人之间相互作用的时间相关性;然后,用图注意网络(Graph attention network,GAT)将长短时记忆网络的隐藏状态信息进行融合,最后得到所需要的预测轨迹信息。本文的轨迹预测模型在苏黎世联邦理工大学行人数据集(Eidgenossische Technische Hochschule Zürich pedestrians dataset,ETHD)和塞浦路斯大学行人数据集(University of Cyprus pedestrians dataset,UCYD)上进行模型训练与测试,实验结果表明,所提出的方法优于现行方法且平均精度提高13.6%。In autonomous driving technology,the results of pedestrian trajectory prediction often affect the safety of autonomous driving.Pedestrian trajectory prediction technology currently faces the problem of interacting with others in real-world scenarios,and needs to consider social interaction and logical consistency while predicting trajectories.Therefore,the previous prediction model improves is improved in this paper to obtain better prediction results.The specific approach is as follows:first,a Long short term memory(LSTM)network model is used to establish the temporal correlation of interactions between pedestrians;then,the hidden state information of the LSTM network is fused using a Graph attention network(GAT)to obtain the desired prediction trajectory information.The trajectory prediction model proposed in this paper is trained and tested on the Eidgenossische Technische Hochschule Zürich pedestrians dataset(ETHD)and the University of Cyprus pedestrians dataset(UCYD)provided by University of Cyprus,and the experimental results show that the method proposed in this paper is superior to the current methods and the average precision is improved by 13.6%.

关 键 词：自动驾驶 行人轨迹预测 LSTM 图注意网络 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]