A novel multi-agent trajectory prediction method for maneuver and collision optimization

Accurate multi-agent trajectory prediction, facilitated by bird’s-eye view from Unmanned Aerial Vehicles (UAVs), is fundamental to proactive traffic management and intelligent driving. However, dense traffic introduces pronounced maneuvering and collision challenges that complicate accurate trajectory forecasting, especially during lane changes and merges. We propose MGC-HCA, a multi-agent trajectory prediction model capable of achieving high prediction performance while addressing maneuverability and collision constraints. To enhance temporal feature extraction, particularly for maneuver-related features, a novel Hybrid Channel Attention (HCA) mechanism is employed. Furthermore, the Unscented Kalman Filter (UKF) is integrated to refine uncertainty estimation in multimodal predictions. Additionally, a collision-cone loss function is introduced to optimize collision avoidance. Comprehensive experiments on three UAV-recorded road traffic datasets (highD, inD, and rounD) demonstrate that MGC-HCA significantly outperforms state-of-the-art methods in trajectory prediction accuracy and collision mitigation, thereby provided a robust tool for enhancing road safety.