Pedestrian Trajectory Prediction Based on Multi-head Self-attention and Sparse Graph Convolution

In response to the problem that existing methods cannot effectively model inter-row interactions and find it challenging to extract meaningful interaction information, a trajectory prediction method using Multi-head Self-attention (MSA) and Sparse Graph Convolution is proposed. First, we use MSA to get the spatial and temporal graphs' interaction and motion trend scores. Then, on this basis, we use the DFC mechanism and GhostNetV1 to construct the directed spatial and temporal graphs and extract pedestrian interaction and motion trend features through sparse graph convolution to obtain information on pedestrian interaction and motion trend information. Finally, the model utilizes TCN to fuse the pedestrian interaction and motion trend features and generate the predicted trajectories. The model performs better on ETH and UCY datasets, with the ADE reaching 0.35 and the FDE reaching 0.62.