Enabling Continuous Drone Tracking across Translational Scene Transitions through Frame-Stitching

Abstract

This paper introduces a method for enabling the continuous tracking of drones across translational scene transitions through the introduction of a stationary global frame. The main benefit of such a method is that it allows current pan-tilt-zoom (PTZ) cameras commonly used in security applications to be used in the domain of drone tracking. The first step in the process is to correct for lens distortion with the intrinsic parameters of the camera obtained through calibration. The motion of the camera frame is then estimated using the optical flow of tracked features from frame to frame. Information obtained about the translational motion of the frame is then used to construct a global frame on which object tracking is performed. Results from the comparison of object detection and tracking performed on two clips with and without pre-processing to construct the global frame show an improvement in tracking performance when the technique was applied to the video. The method was shown to successfully compensate for the relative velocity of objects in the frame relative to the frame, with no sudden changes of velocity during the scene transition. In addition, the introduction of a global frame enabled the continuous tracking of a stationary drone that was reassigned to a new track in the original video.