JCSE, vol. 20, no. 2, pp.89-101, 2026

DOI: http://dx.doi.org/10.5626/JCSE.2026.20.2.89

Cross-Vehicle Attention Network for Distributed Multi-Modal Trajectory Forecasting and Collision Prediction

Balla Uma Mahesh Babu, Giri Babu Kande, and Battula Tirumala Krishna
Department of Electronics and Communication Engineering, Jawaharlal Nehru Technological University Kakinada (JNTUK), Andhra Pradesh, India Department of Electronics and Communication, Vasireddy Venkatadri International Technological University (VVITU), Guntur, Andhra Pradesh, India

Abstract: Cooperative trajectory forecasting helps autonomous vehicles avoid collisions. Existing methods, however, face problems due to limited sensor views. A new method, Cross-Agent Attention Transformer (CATT), addresses this issue by using data shared between vehicles through V2X communication. CATT employs a Transformer-based attention mechanism to selectively fuse LiDAR and camera data from connected vehicles, capturing hidden hazards more accurately. On DAIR V2X, CATT reduced average displacement error (ADE) by about 17% relative to Co-MTP and reduced final displacement error (FDE) by about 16% relative to V2X Graph. Collision prediction accuracy improved by nearly 5%, with collision recall increasing approximately 10%, meaning fewer collisions were missed. Ablation studies revealed performance dropped significantly without cross-agent attention, suggesting this module plays an essential role. All experiments assume ideal V2X connectivity without latency or packet loss; the reported gains therefore reflect best case performance. Future work should evaluate robustness under communication latency and packet loss.

Keyword: No keyword

Full Paper:   8 Downloads, 4 View