Image Segmentation And Object Tracking In Computer Vision

The thesis focuses on the image segmentation and object tracking in computer vision. In the image segmentation, several improved segmentation methods are proposed based on Active Contour Model (ACM), Gradient Vector Flow (GVF) and Watershed and Mean Shift algorithm. Furthermore, we apply two algorithms among them to segment tongue body in computer assisted tongue diagnosis in Traditional Chinese Medicine. In the object tracking, based on the Mean Shift tracking framework, three improved tracking methods are developed, whose aims are to solve the problem of scale and orientation estimation of the moving object and the target representation under the complex conditions. Generally, the main contributions of the thesis are as follows.(1) Two kinds of improved external force fields for ACM are proposed, which are called NGVF (GVF in the normal direction) and Anisotropic GVF respectively. NGVF replaces Laplacian diffusion item in GVF by the Laplacian diffusion operator along the normal direction while Anisotropic GVF chooses the diffusion speed of Laplacian operator along the normal and tangential direction automatically according to the local structure of the image. Compared with the GVF, these two proposed models make progresses on entering into the long and thin concavity and preserving the edge map of the image.(2) An image segmentation method is proposed by combining the 1D-GVF with watershed algorithm. The diffusion procedure of 1D-GVF is that edge information is interpolated into the smooth regions and small scale features are removed while large scale features are preserved well, which makes the image favorable for watershed to segment image. Therefore, over-segmentations of watershed are significantly reduced. It provides the reliable basis for next post- processes such as region merging.(3) An interactive region merging method is proposed. It extracts the object from the initial result of the region based image segmentation algorithm. The proposed method is a region merging mechanism based on maximal similarity. With the guidance of the marker inputted by the user, the proposed method automatically extracts the object.(4) We integrate the 1D-GVF image segmentation method with the proposed region merging method, and apply them to segment tongue body in computer assisted tongue diagnosis in Traditional Chinese Medicine. Firstly, we get the initial segmentation of the tongue image by combining the 1D-GVF and watershed. Secondly, after the object marker and background marker are automatically set by analyzing the structure feature of the tongue image, the proposed region merging method extracts tongue body. In the end, the obtained tongue contour by using ACM is optimized.(5) An improved Mean Shift tracking algorithm called as Scale and Orientation Adaptive Mean Shift Tracking (SOAMST) is proposed. Based on moment analysis and Bhattacharyya coefficient, the proposed SOAMST solves the problem of estimating the scale and orientation changes under the classical Mean Shift framework. It strengthens the adaption of the Mean Shift algorithm when the object has big deformation.(6) Two target representation methods are presented by combing the background and texture feature for improving mean shift tracking. The first method is called as Corrected Background-Weighted Histogram (CBWH). It integrates the background features into target model and enhances the salient features of the target by decreasing the probability of those features which are prominent in the background. The second method combines the color with texture feature to represent the target based on the key points extracted by Local Binary Pattern (LBP), which enhances the separation between the object and the background.