Uav-Based Moving Object Detection Based On Sliding-Window Trajectories Analysis

Moving object detection (or say change detection) for stationary cameras has been well studied by research, however such problem for cameras with general motion still remains challenging This is because the motion of target is coupled with the motion of camera platform and the scene structure. Moreover, the problem can be further complicated if thermal cameras are used. In this work, we propose an effective moving object detection technique for image sequences captured via a UAV's thermal camera. Specifically, in each adaptive sliding window of frames, sparse feature trajectories of the thermal images are extracted, followed by a group sparsity constrained low-rank decomposition to detect foreground in the trajectory domain. Due to the unavailability of a standard benchmark dataset for the freely moving thermal camera case, in addition to the VIVID sequence, we include three challenging thermal image sequences that reflects real-world settings for performance evaluation. The results demonstrate that our method works effectively and achieves better performance than many available methods.