ADAPTIVE LONG TERM TRACKING AND AUTONOMOUS FOLLOWING USING STEREO-CAMERA OF AN UNMANNED AERIAL VEHICLE WITH COLLISION AVOIDANCE
Unmanned aerial vehicles commercially called quadcopters or drones have become increasingly popular over recent years, delving into wide range of fields from medicine for providing immediate health care or in agriculture for locating damaged crops using special sensors to being used in quarries for 3d mapping. We focus on the application of drones in adaptive long term tracking of an object-of-interest and following it with necessary collision avoidance. For this we have implemented a tracking framework called TLD, employing an integrated stereo camera on-board the commercial drone Spiri as the sensor to perform long-term tracking of a target object and use the depth map generated from the disparity of the stereo camera to maintain necessary distance from the target. This is built over the ROS framework. We examine and demonstrate this design in real-time on a commercial drone with monocular camera and in simulation on a model drone integrated with stereo camera. We further refined the tracking process by remodeling TLDs tracker to work with SIFT features supplemented by depth information. We present the evaluation results to show the improvements achieved by our algorithm to autonomously maneuver the drone in making smooth and rapid transitions and then provide comparisons to show improved tracking resilience against modest change in object appearance immediately following system initialization.