Mean Shift Tracking Technology Based On Multi-feature Fusion

With more extensive application of Video, the explosion of video information time has arrived. With maturity and widely application of video compression technologies, more and more video are applied, transmitted and stored by compression method. What's more, people become more and more urgent for the demand for exact video content. Current a typical color-based motion tracking framework is based on the classic Mean shift nonparametric density estimation. And its improvement of tracking performance mainly depends on the accuracy of the color characterization's description of targets to be tracked. If you continue to follow this thought, it's kind of difficulty to improve the effects of Mean shift based on color characterization.For solving these theoretical and technical challenges, this article is mainly on research of multi-feature fusion key algorithm based on Compressed-domain analysis, and make use of available information of video stream's compressed domain to mine the characteristics which can be extracted from the compressed domain, such as movement characteristics, geometry, statistical characteristics of texture and so on. By integrating the pixel color characteristics of the domain, we can implement a new Mean shift of moving object tracking framework based on compressed domain feature and color feature integration. Meanwhile ensuring efficiency of operation, we can enhance the effect of target tracking.The creative contributions of this study mainly include the following four aspects:Firstly,the integrated multi-feature Mean shift tracking modelAimed at the problem that the current single color feature cannot describe moving object fully, this paper integrates the feature of the compressed domain with the motion tracking operation of the Mean shift based on the color feature and it is the first to present the motion tracking frame work of Mean shift based on the integrated feature of compressed domain and color and it analyzes the way of extracting compressed domain feature based on compressed domain information. In the given bitstream, this paper extracts the motion feature, geometric feature and statistical feature of the texture according to motion vector of compressed domain and DCT coefficient and applies the extracted compressed domain feature into the tracking operation of the Mean shift based on the color feature, which improves the algorithm tracking accuracy effectively. The result of the experiment shows the proposed multi-feature Mean shift tracking based on the compressed domain analysis can improve the tracking accuracy significantly compared with the original tracking operation of the Mean shift based on the color feature.Secondly, the construction of the renewed kernel window-width model of Mean shift based on motion vector analysisAimed at the present problem that it is vulnerable to lose track of objects which vary sizes in movement by adoption of fixed kernel window-width in Mean shift computation, this paper puts forward the Mean shift kernel window-width renewal model, which can extract the geometrical features of moving objects and obtain their sizes through motion vector analysis, by which dynamic the kernel window-width is renewed as well as the objective model of moving targets. It's on this basis to construct the kernel window-width model of objective template in Mean shift computation, to elevate its kernel window-width accuracy after size change of dynamic goals and to improve the motion tracking effect under target size varying condition.Thirdly, method for the similar color interference problem of Mean shift based on compressed domain analysisThis paper presents method based on compressed domain analysis aimed at the following problems:such as, color probability density of the current Mean shift algorithm cannot describe the moving target accurately; it is easy to lose the track under the condition of color interference and the similar color background and so on. The method extracts statistical characterization of object texture of the compressed domain, matches the texture statistical characterization of the target template and the candidate template and builds the criterion for judging whether the candidate template is the moving target to guarantee the accuracy of the tracking results of Mean shift.Fourthly, fast moving target tracking model based on compressed domain analysisAimed at the problem which is Mean shift tracking is easy to lose the fast moving target, this paper, according to compressed domain analysis, extracts the speed and direction of the moving target and evaluates the position coordinates of the moving target in the current frame which are used to amend first centralized search of the Mean shift candidate template. This method solves the defect of the Mean shift algorithm which can only search in the local area, enhances the tracking effect of the fast moving target of Mean shift algorithm and reduces the iterations of Mean shift algorithm to improve the operational efficiency of this algorithm.Above all, in order to achieve the better tracking effect, this study which starts from Mean shift tracking based on color feature blends the main features extracted after the analysis of compressed domain, introduces the new tracking frame work integrated the feature of the compressed domain with the feature of the pixel domain, presents a series of new feature extraction algorithm of the compressed domain and the tracking algorithm integrated the feature of the compressed domain with the feature of color and further eliminates the defects of Mean shift algorithm. It has significant theory value and promising future.