MSDHGNN: Multiscale Decomposition Hypergraph Neural Network for Pedestrian Trajectory Prediction

Abstract

Pedestrian trajectory prediction is vital for enhancing the safety and operational efficiency of autonomous vehicles and navigation robots. Hypergraph neural network-based methods have shown advantages in modeling high-order interactions among pedestrians, but two key challenges remain: 1) insufficient modeling in pedestrian social interactions and movement patterns at different time scales, and 2) coarse hypergraph construction and lack of mechanisms to filter redundant interactions. To address the above challenges, we propose a Multi-Scale Decomposition Hypergraph Neural Network (MSDHGNN), which employs dedicated sub-networks for each time scale to model pedestrian social interactions and motion patterns. MSDHGNN develops a kinematic parameter joint estimation module that jointly utilizes both pedestrian position and velocity information to construct the hypergraph, accurately describing individual and group interactions. Furthermore, a multi-feature guided sparse module is further applied to eliminate redundant connections. Finally, pedestrian movement patterns are captured through temporal convolution networks, and then features on multiple time scales are integrated to complete multimodal trajectory prediction. Extensive experiments on multiple datasets demonstrate that MSDHGNN achieves lower prediction errors than existing methods and can effectively model social interactions and movement patterns across multiple time scales. The code is available at https://github.com/Chenwangxing/MSDHGNN-Master.