Research On The Key Technologies Of Point Clouds Processing In 3D Reconstruction

Moving target detection and tracking are the important researche content in the filed of computer vision. It’s also a multi-disciplinary crossed research topic. Target detection and tracking are the key steps of the following video analysis. At present, there are many theories and systems about target detection and tracking, however their real-time performance and accuracy has a certain difference with reality. Target detection and tracking results under complex environment are not ideal, this paper proposed the following strategies to solve the above problem. Adjust the parameters of the Gaussian mixture model based on its background modeling, and then construct adaptive update background model about image sequence, in order to provide basis for target detection. In order to improve the accuracy of target segmentation, we combined local energy information with improved signed distance regularization term based on level set theory. What’s more, realizing the target exactly tracking by anisotropic kernel function model that expressed by level set, this based on mean shift framework. Main research contents are as follows.Firstly, a modeling and target detection algorithm based on adaptive adjustment K -r for Mixture Gaussian background is proposed for complex scenes with non-stationary background. The Gaussian Mixture Model(GMM) is applied to learn the distribution of per-pixel in the temporal domain, then a method is constructed for adaptively adjusting the number K of Gaussian components, and the number K will be added, deleted, or merged with similar Gaussian components according to different situation. Furthermore, two new parameters are introduced in the adaptive parameter model, and the parameter ris adaptively adjusted according to the actual situation, which assures that the background modeling and target detection real-time changes with the pixel. The property of real-time and accuracy reduces the loss of information for moving target and improves the robustness and convergence.Secondly, the uneven color image cannot be segmented successfully by the traditional C-V model, and the C-V model is sensitive to the initial contour and the location. The existing signed distance regularization term has disadvantages, such as the periodic oscillation and the local extremum. This paper proposes the target segmentation algorithm, which combines the local energy information with improved signed distance regularization term(LDRCV model). This algorithm adds local information energy, curve length constraint and signed distance regularization term into the global image information of traditional C-V model. The new algorithm inherits the advantages of global and local energy function adequately, and drives the level set evolved to the target border accurately. The global image information be expanded to the HSV space, and each pixels and its statistical properties are analysed by the local energy information within the neighborhood, which can effectively realize the uneven distribution of color image segmentation in less iteration.Thirdly, the improved signed distance regularization term avoids re-initialization of level set function, improving the computational efficiency, and maintains stability in the level set function evolution process.in order to improve the accuracy and real-time of target detection by LDRCV model, the improved signed distance regularization term avoids re-initialization of level set function, as well as solved periodic oscillation and local extremum, improved the computational efficiency, and maintains stability in the level set function evolution process. The author introduced threshold evaluation method as the termination criterion of level set function evolution in LDRCV model. The author proposed threshold evaluation method as the termination criterion for level set function evolution in LDRCV model. Therefore, we detect the varied length of each iterative curve L(C), when the evolution cure approximate to the target contour, the difference of adjacent two evolution curve’s length becoming smaller and smaller. If the difference is small than a certain threshold, we can take the evolution curve as the target contour, and terminate evolution.Finally, traditional mean shift tracker by isotropic kernel founction often loses the target in the process of the target tracking. This paper proposed target tracking algorithm, which is an adaptive model of anisotropic kernel mean shift, based on level set theory and improved signed distance regularization. This algorithm introduced signed distance constraint function into signed distance kernel function, and then constructed anisotropic kernel function.The anisotropic kernel function keeps the function value in the external region of the target contour is zero. The tracking window contains all sample points, and a few or no background sample points. Constructing window centroid calculation method of the anisotropic kernel function mean shift based on traditional mean shift framework. Similarity threshold thrr was introduced to limit the change of the target model between two sequential pictures and tracking the target accurately.