Distance- and Azimuth-Aware Interaction Modeling for Trajectory Prediction in Unstructured Environments

Autonomous driving in unstructured environments poses significant challenges due to the dynamic and complex interactions among diverse traffic agents. Accurate trajectory prediction is critical for ensuring safety, yet existing models often struggle to effectively capture these intricate interactions. This study proposes a novel multi-agent trajectory prediction framework that integrates a plug-and-play interaction modeling layer on an encoder-decoder architecture. A distance- and azimuth-aware filtering technique is proposed to precisely identify interacting agents, followed by a Distance-Azimuth Convolutional Pooling operation to extract the fine-grained directional interaction feature. To enhance robustness, a Multi-scale Temporal Convolutional Embedding is incorporated into the encoder, enabling reliable performance even under sensor failures and incomplete trajectory data. Extensive experiments on two datasets demonstrate the effectiveness of our approach, showing promising performance while maintaining computational efficiency.