A Multi-Level Group Characterization and Sparse Interaction Network for Pedestrian Trajectory Prediction

Accurate pedestrian trajectory prediction is crucial for autonomous systems, yet remains challenging due to the complexity of social interactions, particularly within groups. Existing methods often overlook the nuanced influence of velocity on group behavior or inefficiently model interactions, this paper proposes a novel group-aware graph network that integrates multi-level group features such as position and velocity and a sparse interaction module. Our key innovations include a learnable grouping mechanism and efficient feature fusion strategy. We conducted a series of experiments on the walking pedestrians dataset (ETH), crowds dataset (UCY), Stanford Drone Dataset (SDD) and ApolloScape dataset. On the SDD dataset, the average displacement error (ADE) and final displacement error (FDE) of the model reached 0.31/0.51; on the ETH/UCY dataset, the average ADE and average FDE of the model reached 0.24/0.41; on the ApolloScape dataset, the model achieves optimal trajectory prediction performance in the pedestrian class, with ADE/FDE of 0.58/0.83, respectively. Ablation studies have shown that our group feature inscriptions and sparse interaction modules have respectively reduced the average ADE by 15% and 8%, and constructed a deployable, lightweight framework.