Crosswalk Trajectory Prediction via an Intent-Trajectory Integrated Framework for Autonomous Driving

Abstract

The increasing complexity of automated urban environments demands accurate modeling of pedestrian crossing behavior. Existing approaches often rely on human annotations or vehicle-centric views, limiting their ability to predict pedestrian intentions in dynamic scenarios. In this paper, we propose an Intent-Trajectory Integrated Pedestrian Crossing Prediction framework (ITI-PCP) that explicitly couples high-level semantic intention inference with multimodal trajectory generation. The proposed framework introduces a Dynamic Bayesian Network to model pedestrian crossing intent in an interpretable and temporally evolving manner, enabling early detection of intention shifts prior to significant trajectory deviations. To effectively capture heterogeneous interaction densities in complex traffic scenes, we further design an Adaptive Spatio-Temporal Encoder with dynamic grid resolution selection, combined with a coarse-to-fine multi-scale grid-based prediction scheme that balances multimodality and spatial continuity. Experiments on the ETH/UCY datasets show that ITI-PCP reduces ADE and FDE by 39.1% and 33.3%, respectively, compared to Trajectron++, and achieves a lower KDE-NLL score (–1.24) than AgentFormer and PECNet. On the NuScenes dataset, the model maintains strong performance over a 4-second prediction horizon, demonstrating robust generalization to realistic autonomous driving scenarios. Ablation studies further confirm the effectiveness of each proposed module.