Interaction-aware trajectory prediction for heterogeneous agents in shared spaces

Trajectory prediction in shared spaces represents a fundamental challenge for autonomous systems, requiring accurate forecasting of heterogeneous traffic participants including pedestrians, cyclists, and vehicles. Although deep learning methods have advanced trajectory forecasting, most existing approaches either neglect heterogeneity among agents or focus solely on interactions during the observed history, failing to account for dynamically evolving interactions that may emerge in future time steps. To address these challenges, we propose a novel encoder–decoder framework that strategically integrates cascade spatial–temporal interaction modeling in the encoder and a cross-LSTM decoder, explicitly capturing interactions in the observed history while leveraging the cross-LSTM to account for dynamically emerging interactions throughout the prediction horizon. Experiments on two datasets demonstrate that our approach achieves superior prediction accuracy(ADE/FDE) and lower collision rates compared to strong baselines. Factor analysis and ablation studies validate the effectiveness of each core module and further reveal that reducing the frequency of interaction modeling in the decoder improves both prediction accuracy and computational efficiency. Our findings provide valuable insights for designing more effective and efficient architectures for trajectory prediction in shared space.