Towards Edge-Assisted Trajectory Prediction for Connected Autonomous Vehicles

Trajectory prediction has been identified as a challenging critical task for achieving full autonomy of the connected and autonomous vehicles (CAVs). Despite the advancement of communication technologies, only few studies include the connectivity and data exchange aspects. Thus, we introduce a novel Edge-Assisted clustering architecture that takes advantage of recent deep learning models and the evolution of edge technologies to achieve better forecasting. First, the historical positions of the target vehicles are fed into the base models of all CAVs in the scene, resulting in multiple generated predictions. Then, each prediction is transmitted to an edge server where trajectories clustering is performed using DBSCAN algorithm to obtain multiple partitions with similar trajectories. The largest cluster is averaged then broadcast back to all CAVs in the scene. Our proposed method surpasses state-of-the-art results on the real world trajectory prediction nuScenes vehicles dataset, obtaining better predictions up to 21%. We also demonstrate the robustness of our method against single-agent system failures, succeeding to get very satisfactory results due to our ability to detect outliers. System practicality is studied under the current 5G/6G capabilities.