Research On Image Matching And Moving Target Tracking In Video

Since it was proposed from the research of aircraft auxiliary navigation system in the United States in1970s,’image matching’ has been one of the hot research spots for the scholars around the world. With the rapid development of science and technology, image matching techniques have been more and more important in the field of image processing. At present, it has been widely used in target identification and tracking, as well as many other important fields.The so-called image matching, can be described as, according to the actual situation of two or more images to be matched, select the appropriate feature vectors, similarity criteria and search strategy, and then to determine the match states through the cross-correlation information.As another branch of computer vision, target tracking has also been put more and more focus.Target tracking is to achieve the parameter estimation of motorized and non-motorized target according to discrete-time system’s filtering and prediction algorithms. That is, the main purpose of target tracking is to estimate the moving status of the target route. Now, how to use the stable and effective feature information to characterize the target is the common goal of all researchers. The process of target tracking under video surveillance includes:image pre-processing, target detection, target tracking, target behavior understanding and so on. Moreover, Target tracking is the core of the entire process.Image matching and target tracking algorithms, especially the latter, is the main research topics in the thesis.We briefly describe the classification of current existing image matching and object tracking algorithms, and elaborate on some of the mainstream algorithms, such as SIFT algorithm. SURF algorithm. Kalman filter. Particle Filter and Mean Shift, besides, analyze and compare the advantages and disadvantages.In the thesis, we proposed a combined algorithm, that is, an improved SIFT algorithm based on weighted principal component analysis. We know, the classical SIFT algorithm has good matching effect, but complex calculation, therefore, we introduce weighted PCA to reduce the dimensionality here. And simulation results show that the keypoints abandoned in dimensionality reduction aren’t the stable and effective ones, and that the matching accuracy has been improved.In addition, an improved Camshift algorithm based on Kalman has been raised. Although compare with Mean Shift, Camshift has some advantages, such as it can adaptively adjust the target area when the size of the moving object is changed. But it doesn’t overcome mistakenly tracking when target occlusion. So, we adopt Kalman prediction to ensure the accuracy and the robustness. Furthermore, in order to slove the false tracking in the case of the target obscured by other fast-moving objects, we add a judgment sector in the original combined algorithm. When the situation occurs, use the target position of a previous frame to replace the position of the target in the current frame until the target obscured disappears. But the premise is the movement of the object that to occlusion the target cannot be too fast. More precisely speaking, the relative speed of the tracking target and the occlusion objects can not be too fast, otherwise the proposed algorithm will not be effective.