Pedestrian trajectory prediction model based on self-supervised spatiotemporal graph network

• Hop Interaction for Feature Update: • The model introduces hop interaction instead of traditional node interaction to update node features in the spatiotemporal graph. This approach significantly reduces the number of graph convolution operations, mitigates the issue of feature smoothing, and enhances the accuracy of trajectory predictions. • Self-Supervised Module for Multi-Trajectory Commonalities: • A unique self-supervised module is incorporated into the model to discover common patterns among multiple pedestrian trajectories. This self-supervised learning mechanism further improves the precision of the predictions by leveraging the shared characteristics across different trajectories. • Superior Performance on Benchmark Datasets: • The experimental results on the ETH, UCY, and UNIV public datasets show that the proposed model outperforms state-of-the-art (SOTA) models. Specifically, it achieves a 15% reduction in Average Displacement Error (ADE) and a 10% reduction in Final Displacement Error (FDE). In dense pedestrian scenes, such as those in the UNIV dataset, the prediction errors are reduced by 25% and 22%, respectively. • These highlights underscore the model's effectiveness in improving the accuracy and robustness of pedestrian trajectory predictions. To improve the accuracy of pedestrian trajectory prediction, the graph - based pedestrian trajectory modeling method in the pedestrian trajectory prediction scenario is effective. Thus, a pedestrian trajectory prediction model based on a self - supervised spatiotemporal graph network is proposed. Firstly, in the process of spatiotemporal graph modeling, this model introduces hop interaction instead of node interaction to update node features, which greatly reduces the times of graph convolution operations, alleviates the problem of feature smoothing, and greatly improves the accuracy of prediction. Secondly, a unique self-supervised module is added to the model to mine commonalities between pedestrian’s multi-trajectories through self-supervised to further improve the accuracy of prediction. The experiment uses ETH and UCY public datasets to train and evaluate model performance. The experimental results demonstrate that this model exhibits enhancements in both ADE and FDE metrics when compared to the SOTA model, with an average prediction error reduction of 15% and 10%, respectively. In scenes with dense pedestrians such as the UNIV dataset, the prediction errors are reduced by 25% and 22%.