Context-Aware Heterogeneous Fusion for Pedestrian Trajectory Prediction in Pedestrian-Vehicle Interactions

In autonomous driving scenarios, the high dynamic relative motion between vehicles and pedestrians introduces significant challenges for trajectory prediction. Rapidly changing speeds and directions often lead to unevenly sampled trajectory points, which make it difficult to capture fine-grained motion details and increasing the complexity of prediction. To address these challenges, we propose a novel context-aware hierarchical trajectory prediction framework that leverages heterogeneous fusion, bidirectional residual cross-attention (BRCA), and an adaptive stepwise decoding mechanism. The hierarchical fusion effectively extracts temporal dependencies from sequence information and spatial-dynamic features from visual data. Meanwhile, the BRCA mechanism enhances intermodal interactions by capturing hidden relationships and complementary patterns, while the adaptive decoder refines predictions dynamically, incorporating global motion trends and local dynamic details to improve accuracy. Extensive experiments on the JAAD and PIE datasets demonstrate the effectiveness and superiority of our approach.