Collaborative Trajectory Prediction via Late Fusion

Researchers from Khalifa University, Khalifa University Center for Autonomous Robotic Systems, and other institutions have published a paper on arXiv proposing a novel late-fusion framework for collaborative trajectory prediction in autonomous driving. The work, led by Nadya Abdel Madjid and seven co-authors, addresses a key limitation in current vehicle-to-vehicle (V2V) approaches: the massive communication overhead required to share high-dimensional feature maps at the perception stage. Instead of building a holistic scene view through early fusion, their method shifts collaboration to the prediction module, allowing vehicles to exchange only final trajectory forecasts. This design is model-agnostic and treats each vehicle as an independent asynchronous agent, making it far more practical for real-world deployment where bandwidth and synchronization are often constrained.

The team evaluated their approach on three major datasets: OPV2V, V2V4Real, and DeepAccident. Across all benchmarks, the late-fusion framework consistently reduced miss rates and improved the trajectory success rate (TSR0.5)—defined as the fraction of ground-truth agents with a final displacement error below 0.5 meters. On the real-world V2V4Real dataset, collaborative prediction using this method boosted success rates by 1.69% and 1.22% for the two intelligent vehicles compared to individual forecasting. This modest but consistent improvement, combined with the significant reduction in communication overhead, suggests that late fusion could enable more scalable and efficient collaborative driving systems. The paper is available on arXiv under the subject Robotics (cs.RO) and can be accessed via DOI: 10.48550/arXiv.2604.22973.