Moving Object Detection And Tracking From Airborne Videos

With the increasing deployment of aerial platforms,including unmanned aerial vehicles(UAVs)and intelligent missiles,aerial video processing technologies are drawing more and more attentions.Among them,the moving object detection and tracking technology has a broad range of applications including human-computer interaction,emergency rescue,automatic attack of battlefield targets,and so on.This paper researches on the problem of moving object detection and tracking based on two typical full motion video platforms,i.e.satellites and UAVs.(1)A highly efficient online detection and tracking algorithm based on trajectory search is proposed for detecting an unknown large number of small moving targets in satellite images.A fast trajectory growing algorithm is proposed to efficiently distinguish true targets from a large set of noisy detections.We then propose to enable continuous target labels by simultaneously tracking all the detected trajectories through bipartite graph matching.At last a temporal window based online multiple small target detection algorithm is proposed.Compared to existing methods,the proposed method is capable of finding a large number of moving targets at high efficiency.Extensive experiments show that our method has the advantage of low parameter dependence and high robustness to noise.(2)A real-time moving object detection algorithm based on geometric constraints is proposed for detecting moving objects from a low attitude UAV platform.We start from the two-frame based motion segmentation algorithms.We firstly propose to improve the dense optical flow based algorithm by introducing a graph-based image segmentation algorithm which is demonstrated to be able to effectively find the blurred motion boundaries in the flow field.We then propose to improve the frame difference based algorithm by fusing frame differencing results with image segmentation results,thus to recover the whole target region as well as remove ghosting artifacts.Finally,a new UAV moving object detection algorithm based on geometric constraints is proposed.Based on an improved frame differencing method,we realized reliable detection of motion pixels using the epipolar constraint and the homography constraint inherent in the UAV images.The proposed method is then utilized for fast and accurate sequential moving object detection.Experiments show that our method can handle large parallax in the scenario and output faithful detection results.(3)A dense structural learning method is proposed to track the moving objects in aerial imagery and low resolution videos.Traditional tracking algorithms are faced with the problems of small search range,low robustness and low computational efficiency.In this paper,an online structural learning algorithm with dense sampling is proposed.By introducing the Fourier techniques and optimizing the learning procedure,the proposed method achieves very high efficiency.We then propose to improve the robustness of the tracker without bringing with extra computational burden by introducing the weighting kernel.Finally,an effective object representation method is proposed to deal with the thermal-infrared object tracking problem.Extensive experiments on public tracking datasets suggest high performance of the proposed tracker.Particularly,the best overall performance is observed on a public infrared dataset and an UAV dataset respectively when compared with the state-of-the-arts.