| 引用本文: | 王天保,刘昱,郭继昌,晋玮佩.图卷积神经网络行人轨迹预测算法[J].哈尔滨工业大学学报,2021,53(2):53.DOI:10.11918/202006051 |
| WANG Tianbao,LIU Yu,GUO Jichang,JIN Weipei.Pedestrian trajectory prediction algorithm based on graph convolutional network[J].Journal of Harbin Institute of Technology,2021,53(2):53.DOI:10.11918/202006051 |
|
| 摘要: |
| 针对行人轨迹预测任务中行人间的交互模式难以被有效构建的问题,提出了一种基于图卷积神经网络的算法TP-GCN来建立行人交互模型并进行轨迹预测.首先对行人的轨迹序列使用长短期记忆网络提取轨迹运动特征；随后将行人视为图结构中的顶点,创建表示相互关系的邻接矩阵,并根据视觉盲区范围筛除无关顶点间的连接权重；然后对于轨迹运动特征,使用图卷积神经网络提取不同轨迹间的交互信息,同时增加顶点对自身所隐含交互信息的提取,并使用长短期记忆网络将交互信息编码为轨迹交互特征；之后通过深度图信息最大化方法,对图卷积神经网络的权重进行优化,使得个人的运动模式符合场景内所有行人共有的运动模式；最后将轨迹运动特征和轨迹交互特征使用长短期记忆网络进行解码,完成轨迹预测.在公开数据集ETH和UCY上的实验结果表明,所提算法能够根据行人间的交互模式做出与真实行为接近的符合行人习惯的预测,整体预测精度高.同时,消融实验和预测轨迹的可视化也显示了算法的有效性及良好的可解释性. |
| 关键词: 轨迹预测 交互模式 图卷积神经网络 长短期记忆网络 互信息 |
| DOI：10.11918/202006051 |
| 分类号:TP391 |
| 文献标识码:A |
| 基金项目:云南省重大科技专项计划项目(202002AD080001) |
|
| Pedestrian trajectory prediction algorithm based on graph convolutional network |
|
WANG Tianbao1,LIU Yu1,GUO Jichang2,JIN Weipei2
|
|
(1.School of Microelectronics, Tianjin University, Tianjin 300072, China; 2.School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China)
|
| Abstract: |
| To solve the problem that the pedestrian interaction model is difficult to be effectively constructed in the pedestrian trajectory prediction task, a trajectory prediction algorithm based on graph convolutional network (TP-GCN) was proposed to establish pedestrian interaction and predict future trajectories of pedestrians. First, the long short-term memory was used to extract the trajectory motion features of the trajectory sequences of pedestrians. Then, the pedestrians were considered as the nodes on the graph, and adjacency matrix was built to represent the created interactions. Next, the connection weights between unrelated nodes were screened out according to the blind zone. For trajectory motion features, the graph convolutional network was applied to extract the interactions between the trajectories and increase the extraction of the interaction in each trajectory, and the interaction was then encoded as trajectory interaction features by using long short-term memory. Furthermore, the weights of the graph convolutional network were optimized by the Deep Graph Info method to ensure that the motion pattern of individual accords with those of all the pedestrians in the scene. Finally, the trajectory motion features and trajectory interaction features were decoded using long short-term memory to complete the trajectory prediction. According to the experiment on the public datasets ETH and UCY, the proposed algorithm could make the predictions of pedestrian habits close to the real trajectories based on the interaction model between pedestrians, and the overall prediction accuracy was high. In addition, the ablation experiment and the visualization of the predicted trajectory also verified the effectiveness and interpretability of the algorithm. |
| Key words: trajectory prediction interaction model graph convolutional network long short-term memory mutual information |
