Extending Motion Detection to Track Stopped Objects in Visual Multi-Target Tracking
Jacob H. White, Karl Salva, Randal W. Beard · 2018
Various solutions to visual multi-target tracking have been proposed, but many of them are not capable of running in real time from a moving camera on an unmanned aerial vehicle (UAV). We present a tracker that runs in real time and tracks multiple objects while accounting for camera motion on a UAV. Our algorithm is capable of processing over 10 frames per second on a 1280x720 video sequence. We utilize Recursive-RANSAC, an efficient algorithm for tracking multiple objects in clutter. Our work combines motion detection with optical flow and feature matching to allow stationary objects to be tracked. We use a feature prioritization algorithm to reduce computational complexity and spatial redundancy. We also present a ghost track reduction method which prevents tracking non-existent objects when true objects are no longer visible. We demonstrate the performance of our tracker on a moving camera video sequence. Video results are available at https://youtu.be/6bXjKb_-6qY.