Vehicle trajectory extraction and matching using multi-UAV cooperative vehicle video data

Abstract

Unmanned Aerial Vehicles (UAVs), known for their high-altitude operation, wide coverage, and flexibility, are increasingly used as sensors for traffic data collection. However, the limited field of view of a single UAV poses challenges in capturing comprehensive traffic information required for medium-level and macro-level traffic flow analysis. To address this limitation, this study proposes a post-processing method for multi-UAV collaborative video to extract wide-area vehicle trajectories. The framework comprises four main components: vehicle detection, tracking, coordinate conversion, and trajectory matching. Specifically, YOLOv8 is utilized for vehicle detection, while the DeepSORT algorithm is employed for vehicle tracking. To transform pixel coordinates into real-world coordinates, a conversion matrix is obtained through camera calibration. For trajectory matching, a cost matrix between adjacent videos is constructed for multi-segment trajectories, and assignments are optimized using the Hungarian algorithm. The proposed framework is evaluated using four metrics: a correct trajectory matching rate of 0.964, a Root Mean Square Error (RMSE) of $0.46$ m, a Mean Absolute Error (MAE) of $0.51$ m, and a similarity score of 0.999. Experimental results demonstrate that the proposed framework effectively overcomes challenges in trajectory extraction and matching, making it a reliable approach for wide-area vehicle trajectory analysis.