DSTIGCN: Deformable Spatial-Temporal Interaction Graph Convolution Network for Pedestrian Trajectory Prediction

Accurate and reliable pedestrian trajectory prediction can reduce the risk of human-vehicle collisions and predict accidents in advance, which is crucial for developing autonomous driving and intelligent monitoring. Previous trajectory prediction methods face two common problems: 1. ignoring the joint modeling of pedestrians’ complex spatial-temporal interactions, and 2. suffering from the long-tail effect, which prevents accurate capture of the diversity of pedestrians’ future movements. To address these problems, we propose a Deformable Spatial-Temporal Interaction Graph Convolution Network (DSTIGCN). First, we construct a spatial graph and employ the attention mechanism to preliminarily describe the spatial interactions of pedestrians at each moment. To solve problem 1, we design a deformable spatial-temporal interaction module. The module autonomously learns the spatial-temporal interaction relationships of pedestrians through the offset of multiple asymmetric deformable convolution kernels in both spatial and temporal dimensions, thereby achieving joint modeling of complex spatial-temporal interactions. Next, we obtain trajectory representation features through graph convolution and then predict the two-dimensional Gaussian distribution parameters of future trajectories using the Temporal Attention-Gated Temporal Convolution Network (TAG-TCN). To address problem 2, we introduce Latin hypercube sampling to sample the two-dimensional Gaussian distribution of future trajectories, thereby improving the multi-modal prediction effect of the model under limited samples. Experiments on ETH, UCY, and SDD datasets have verified that our method can achieve high-precision prediction of pedestrian future trajectories under limited parameters.