Real-World Trajectory Extraction from Aerial Videos - A Comprehensive and Effective Solution

Dongfang Zhao, Xiaopeng Li

Abstract: In this paper, we propose an advanced real-world trajectory extraction method to extract vehicle trajectories from aerial videos. The method includes three steps: vehicle identification and tracking, camera rotation and shifting correction, and semi-automatic lane identification. We identify vehicles using the advanced Mask R-CNN model and track vehicles according to their movement properties. We calculate camera rotation and shifting parameters using a binary search correlation-based matching algorithm. And we generate lane structures based on the topographic properties of extracted trajectories semi-automatically. We tested the method using two aerial video examples, one of which captures freeway traffic and another one captures roundabout traffic. The results show that vehicle trajectories are extracted accurately. The proposed method provides a promising solution to extract real-world long vehicle trajectories from aerial videos.