Abstract: In order to solve the problem that pedestrian trajectory prediction research only focuses on the interactive information of historical trajectories and ignores the interactive information of endpoints, a pedestrian trajectory prediction model based on Graph Convolutional Network and Endpoint Induction(GCN-EI) was proposed. First, the classification method was used to learn the weighted endpoint distribution of pedestrians in the future on the training set, and then the possible endpoint was connected with its corresponding historical trajectory. The interaction features of pedestrians are extracted over a longer time span using the spatiotemporal Graph Convolutional Network (GCN) of attention mechanism and endpoint conditions, and their internal motion features were extracted using the individual feature module. Finally, the future trajectory of the pedestrian is predicted by the convolution in time. The model was tested in detail on the ETH/UCY dataset, and the Average Displacement Error (ADE) and Final Displacement Error (FDE) decreased by 4.5% and 5%, respectively, compared to the STITD-GCN model, and the final displacement error (FDE) decreased by 9.5% compared to the PCCSNet model of the classification method.

Key words: Pedestrian trajectory prediction, Attention mechanism, Endpoint induction, Graph Convolution Network (GCN), Classification method

摘要： 针对行人轨迹预测研究中仅关注历史轨迹的交互信息，而忽略了终点交互信息的问题，提出了一种基于图卷积网络和终点诱导的行人轨迹预测模型(GCN-EI)。首先在训练集上使用分类方法学习行人未来可能的加权终点分布，然后将可能的终点与其对应的历史轨迹连接，使用注意力机制和终点条件的时空图卷积网络(GCN) 在更长的时间跨度上提取行人的交互特征，同时使用个体特征模块提取其内在运动特征，最后通过时间内推卷积预测行人的未来轨迹。在ETH/UCY数据集上对模型进行了详尽的测试，相较于STITD-GCN模型，平均位移误差(ADE)和最终位移误差(FDE)分别下降了4.5%和5%，相较于分类方法的PCCSNet模型，最终位移误差(FDE)下降了9.5%。

关键词: 行人轨迹预测, 注意力机制, 终点诱导, 图卷积网络, 分类方法