Most encoder-decoder structure based predictions models usually predict trajectory according to the position and historical movement of nearby pedestrians. Their input range (receptive field) is small. They often ignore some specific information such as the speed and direction of pedestrians’ movement or the temporal attention. This leads to detailed pedestrian interaction that cannot be obtained. Therefore, we propose a novel spatio-temporal graph attention network (GAT) called GSTA. In the spatial domain, GSTA obtains complex interaction by spatial attention (SA) based on multi-feature fusion, and expands the receptive field through feature updating mechanism (FUM). In the temporal domain, we design temporal attention module (TAM) and feature selection module (FSM). TAM is used to discover the internal relationship of historical trajectory and solve the problem that temporal attention is averaged. FSM overcomes the adverse effect of small temporal perceptual range and reasonably controls the flow of feature information. Experimental results on 5 commonly used pedestrian trajectory datasets show that the prediction accuracy of our proposed model is further improved.

大多数基于编码器-解码器结构的预测模型通常根据附近行人的位置和历史运动来预测轨迹。它们的输入范围（感受野）很小。他们经常忽略一些特定的信息，例如行人运动的速度和方向或时间注意力。这导致无法获得详细的行人交互。因此，我们提出了一种新的时空图注意网络（GAT），称为 GSTA。在空间域，GSTA通过基于多特征融合的空间注意力（SA）获得复杂交互，并通过特征更新机制（FUM）扩展感受野。在时间域中，我们设计了时间注意力模块（TAM）和特征选择模块（FSM）。TAM用于发现历史轨迹的内部关系，解决时间注意力平均化的问题。FSM克服了时间感知范围小的不利影响，合理地控制了特征信息的流动。在 5 个常用行人轨迹数据集上的实验结果表明，我们提出的模型的预测精度进一步提高。