Abstract:

In order to solve the problem that pedestrian trajectory prediction research only focuses on interactive information of historical trajectories and ignores interactive information of endpoints， a pedestrian trajectory prediction model based on Graph Convolutional Network （GCN） and Endpoint Induction was proposed， named GCN-EI. Firstly， a classification method was employed on the training set to learn the weighted distribution of potential future endpoints for pedestrians. Subsequently， the possible endpoints were connected with their corresponding historical trajectories， and the interactive features of pedestrians were extracted over a longer time span by using the GCN with attention mechanism and endpoint conditions. Meanwhile， an individual feature module was used to extract the internal motion characteristics of pedestrians. Finally， the future trajectory of pedestrian was predicted by the temporal inference convolution. Test results on ETH and UCY datasets show that compared to STITD-GCN （Spatio-Temporal Interaction and Trajectory Distribution GCN） model， the proposed model has the Average Displacement Error （ADE） and Final Displacement Error （FDE） decreased by 4.5% and 5.0%， respectively； moreover， compared to PCCSNet （Prediction via modality Clustering， Classification and Synthesis Network） model using classification method， it has the FDE decreased by 9.5% .

Key words: pedestrian trajectory prediction, attention mechanism, endpoint induction, Graph Convolutional Network (GCN)

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，提出一种基于图卷积网络（GCN）和终点诱导（Endpoint Induction）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；其次，将可能的终点与它们对应的历史轨迹相连接，并使用基于注意力机制和终点条件的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行人的内在运动特征；最后通过时间内推卷积预测行人的未来轨迹。在ETH和UCY数据集上对模型进行的测试结果表明，相较于STITD-GCN（Spatio-Temporal Interaction and Trajectory Distribution GCN）模型，所提模型在平均位移误差（ADE）和最终位移误差（FDE）上分别下降了4.5%和5.0%；相较于采用分类方法的PCCSNet（Prediction via modality Clustering， Classification and Synthesis Network）模型，在FDE上下降了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