The Research On Tracking In Intelligence Videos Surveillance Based On Multi-Camera

The intelligence videos surveillance is an important application of computer visual technology. Moving object tracking and re-identification are an important research topic in the intelligent video processing field, and is a theme of frontier research that needs multi-subject theory. Due to the broad application prospect of intelligent video tracking, it has become a central issue of research for the field of computer vision, also caused the high attention of researchers at home and abroad. The main research content of it is to estimate the state or the trajectory of moving objects from the sequence of video. In recent years, the rapid development of computer technology has improved greatly large-scale computing speed, this makes the mathematical theory of video tracking be applied practically, and it makes the intelligent video tracking technology be widely used.The mean shift algorithm and particle filtering algorithm for non-parametric density estimation method are used to deal with problem of video tracking. But because of the char-acteristics of video and the defect of the tracking algorithm itself, some inherent problems still exist. Aiming at these problems, this thesis makes some researches as following:In order to improve the accuracy and robustness of the tracking, we introduce the con-cept of a geogram that includes more rich features of the image. The spatiogram contains some moments upon the coordinates of pixels corresponding to each histogram bin while the geogram contains information about the perimeter of grouped regions in addition to fea-tures in the spatiogram. The perimeter of the given region has capability to represent the geometrical compactness on the distribution of the given feature. The geogram-based fea-ture descriptor increases the accuracy of tracking because it can capture the features in lower levels. We test our feature descriptor and measure in object tracking scenario.From the point of view of the automatic control theory, we consider that a convergence of the mean shift algorithm for the spatiogram is divided into the obvious dynamic state and the steady state, and introduce a hybrid technique of the geogram and the histogram to control the convergence of the mean shift algorithm. Moreover, we derive a mean shift procedure for the proposed geogram. Then, the accuracy and the robustness of the proposed algorithm is evaluated by comparison of experimental results.From the point of view of the similarity calculation for particle filter, we explain com-putational instability of IBD(Incremental Bhattacharyya Dissimilarity) by analyzing do-main and range properties of cross-bin method. Motivating by the transportation problem, we introduced the normalization method of an AISM (Asymmetric Incremental Similarity Matrix) and AIBS(Asymmetric Incremental Bhattacharyya Similarity) as non-symmetrical similarity. We test the similarity measure in object tracking scenario.In order to improve the performance of the particle filtering, we introduce the concept of a spline resampling in particle filter to deal with two problems of particle filter:the high accuracy and the sample impoverishment. The spline resampling consists of two parts:the spline transformation of weights and the spread transformation of states. Two transforma-tions are sequentially implemented to incorporate with each other. Then, we propose a global transition model of estimating the object position. Also, we propose the refining-resampling algorithm to advance the performance of the framework based on association of PF and KBOT. Finally, we test the stability and the accuracy of the association approach-es that we proposed in this paper, on a synthesized image sequence and several real image sequences.Based on the intergrating framework of improved mean shift and the particle filter-ing, we analyzed of the controllability of all the particles and inappropriateness of tracking method based on elimination of ill-positioned particles. Then, in order to control the con-vergence of all the particles, we proposed boosting-refining approach based on condition number, and in order to get the estimation result close to the global maximum, we pro-posed the refining-resampling approach. Finally, we test our approaches in object tracking scenario.Based on the mean salience designed especially for person re-identification, human salience is calculated across non-overlapping camera views and is used as a meaningful representation of human appearance in recognition. Then, person re-identification is defined as a mean salience matching problem. Dense correspondences between local patches are used to calculate visual similarity. Finally, global bi-directional matching is proposed to guarantee robust mean salience matching, and mean salience matching and global bi-directional matching are tightly integrated into a unified framework based on a combined salience.