Multiple Aerial Videos-Based Long-Distance Vehicle Trajectory Construction With Spatiotemporal Continuity

The high-fidelity trajectory data extracted from aerial videos can provide reliable trajectory source for traffic-related researches and promote autonomous driving researches. However, compared to the current studies focusing on the single video-based trajectory extraction, the long-distance trajectory extracted from multiple aerial videos still suffer from trajectory unrealistic and discontinuity problems. Existing long-distance trajectory construction method can be categorized into two classes: trajectory-based and video-based methods. The former significantly changes the original trajectory data, leading to trajectory distortion and cannot restore vehicle travel videos with corresponding trajectories, thus falling short of meeting research demands in traffic filed. The trajectory data extracted by the latter approach still exhibit noticeable discontinuity at videos stitching area, leading to the biases of vehicle motion parameters, which further cause the errors of studies such as vehicle model calibration and safety evaluation. Therefore, this study introduces a novel long-distance trajectory construction method which can extract long-distance trajectories from multiple aerial videos and mitigate the discontinuity problem in video stitching area. The proposed method contains four modules which are multi-video registration, vehicle detection and tracking, long-distance trajectory connection and fusion, trajectory smoothing and parameters extraction. Experimental results show that the proposed method can obtain more continuous motion parameters in spatiotemporal distribution, and have more advantages in traffic-related studies. This work may provide a new research perspective for current trajectory construction researches in traffic field.