Multi-camera multi-target drone tracking systems with trajectory-based target matching and re-identification

Abstract

This paper presents the integrated use of multiple cameras to detect, track and localize multiple moving objects, especially drones, in real-time by computer vision. The algorithm of the integrated system incorporates target tracking, localizing and identifying schemes, with the ability to use multiple cameras from different viewing angles and simultaneously track moving objects in the cameras frames. Moving objects are detected by hybrid detection (motion-based blob detection appearance-based detection), and movements are predicted with Extended Kalman Filter. For motion-based detection, the trajectories of the tracked objects are analyzed using trajectory features variable. For appearance-based detection, motion is tracked with Yolo V3 detection algorithm. The integrated target identification algorithm matches and re-identifies targets between frames (intra-camera) and among cameras (inter-camera). This algorithm subsequently cross-correlate every tracked object in the different camera frames, and pair all the tracked targets to each other so that the cameras track the same objects for subsequent 3D localization. We tested our multi-cameras system to track multiple aerial and ground targets, and were successful in the re-identification of targets in real time.