Panorama Stitching, Moving Object Detection and Tracking in UAV Videos

Abstract

Unmanned Aerial Vehicles(UAV) are more and more wildly used recent years in many fields. It's convenient to acquire more static and dynamic information by uav aerial videos to grasp the scene situation. Frames registration, panoramic image mosaic, moving objects detection and tracking are the key and foundation of the aerial video analysis and processing. Firstly, we use a l_q-estimation method to remove the outliers and match the feature points robustly. Then we utilize a Moving Direct Linear Transformation (MDLT) method to find the homography of the frames more accurately, and stitch the frame sequence to a panorama. Finally, we apply a 5-frame difference method on the warped frames to detect the moving objects, and use a long-term visual tracking method to track the object of interest in complex scenes. The experiments show that our method achieve good results in different conditions.