Pedestrian Trajectory Prediction Based on Multi-Relational Graph Convolution and Dynamic Attention

To address the issues of insufficient interaction modeling and error accumulation in pedestrian trajectory prediction under complex scenarios, a two-stage prediction method integrating multi-relation graph convolution and dynamic attention mechanism is proposed. This method captures local dense interactions and global motion trends in the spatial dimension through a multi-relation graph convolution network, and designs a global temporal aggregation module in the temporal dimension to integrate historical trajectory features through LSTM and introduce a direction change detection mechanism to alleviate error accumulation. Additionally, a probabilistic trajectory generation and optimization module is designed in this paper, which determines the optimal endpoint based on Gaussian mixture model and maximum likelihood estimation and embeds a dynamic attention mechanism to predict vector offset fields and optimize the initial trajectory, reducing cumulative errors. Compared with the average displacement error (ADE) and final displacement error (FDE) of some mainstream models, the effect of this method is improved.