Cooperative Perception-Based Vehicle Trajectory Prediction and Conflict Estimation

Abstract: Intersection safety challenge has been a critical research focus in transportation field for over decades, especially as rapid urbanization introduces greater complexity into intersection management. The increasing presence of various road users, such as pedestrians, cyclists, and vehicles, has made intersections more unpredictable and hazardous. Among these, high-speed vehicles pose the greatest threat to traffic safety. Drivers often experience vision occlusion and blind spots, which can lead to serious accidents. Even with the support of Autonomous Driving technologies that incorporate remote sensing equipment, environmental perception is limited to a two-dimensional, 360-degree view. For instance, a pedestrian crossing behind a large truck may be difficult to detect, whether by human vision or current sensing techniques, making it challenging to predict potential risks.
To address intersection safety challenge, this work introduces Cooperative Perception using roadside cameras to enhance real-time vehicle conflict detection. Unlike conventional image-based vehicle trajectory prediction and risk estimation methods commonly used for dashcams and surveillance cameras, this approach proposes to convert vehicle pixel coordinates in roadside cameras to real-world geographical coordinates. Combined with actual speeds, headings, and sizes calculated through camera calibration, the system predicts future vehicle trajectory on a real-world scale and inferences Time-To-Collision for conflict estimation. Computer Vision based AI techniques such as object detection and tracking are integrated to support trajectory prediction. This advanced pipeline can transmit processed recognition, prediction and risk data via UDP protocol in real time, enabling immediate alerts to road users.
Experiments are performed at several pilot test sites featuring different settings of intersection design, vehicle types, and roadway geometry, to validate the system’s effectiveness. Furthermore, the developed models are compressed and deployed to Edge Computing environments for more efficient local processing. The proposed system has great potential to significantly improve intersection safety in a practical and efficient manner.