| Moving object detection and tracking have always been a hot topic in the field of computer vision and intelligent monitoring, and moving object detection is the basis for some subsequent processing carried out in a video sequence like target recognition, tracking as well as behavior analysis. In most real scenes, the camera is mostly stationary, which is so-called "static background". Nevertheless, with the development of technology and requirement of cost reduction, more and more moving cameras are needed to achieve continuous tracking for the moving targets, which aims at increasing the scope of monitoring. However, the movement of cameras would result in constant change of the whole background, which would increase the difficulty of moving target detection compared to a static background. Therefore, moving target detection technology under dynamic background is quite a challenging research topic.In this thesis, the available algorithms of moving target detection are intensively studied, especially for these under dynamic background caused by surveillance cameras. Global motion compensation algorithm is usually used to detect foreground objects in a dynamic background, namely converting dynamic background into static background. Hence, this thesis studies three object detection methods for static background, Gaussian mixture background modeling, SOBS, Vi Be and the improved algorithm, which are analyzed by comparision using test videos in different motion scenarios. The experiment showes that Vi Be algorithm not only has strong adaptability for various moving scenes,but also possesses a high accuracy, especially with strong competitiveness in the scene of camera shake.The core content of global motion compensation method is to extract and match image feature points, and it is critical to select a suitable feature to represent a image. This thesis conducts an experiment to compare the extraction and match ability for typical features: SIFT, SURF and ORB, furthermore, some analysis and research is made. It can be seen from the experiment that as a scale invariant feature, SIFT feature has a more stable feature matching capability. For this reason, moving target detection algorithm based on SIFT feature points matching is deeply analyzed in this thesis, in addition, an improved scheme is proposed to the deficiency of existing algorithm.Traditional SIFT feature matching algorithm is not very efficient, especially that the process of generating feature vector takes a long time, therefore two new SIFT feature descriptor extracting methods are put forward in this thesis. One descriptor is reduced to 32-D from 128-D directly, and the other uses a circular area around of the feature points to construct a 64-dimensional SIFT feature vector descriptor, both of which cut down the feature vector generation and matching time. Meanwhile, using partial matching algorithm instead of the original global matching algorithm for feature points shortens the time further, which enhances the efficiency of the SIFT algorithm.The accuracy of the moving target detection also depends on the performance of those algorithms under static background after converting the dynamic background to a static background. For the deficiency of frame difference method of existing "holes" in the extracted moving objects, Vi Be(a better performance method) is used to detect moving targets, which improves the integrity of the targets.Experimental result shows that the proposed method not only significantly shortens the matching time for SIFT feature points, but also ensures the integrity of the detected target. That is, our method accelerates the speed of moving object detection under the dynamic background and improves the accuracy as well, which has a certain theoretical value in the surveillance video field with moving cameras. |
PDF Full Text Request