Study On Color Iamge Segmentation And Vision Target Tracking Under Complicated Environment

With the development of digital signal processing theory and computer technology, motion-target tracking has become an important researching topic in many study fields, such as pattern recognition, image processing, computer vision, and weapon control-guide, etc. Combining image processing, automatic control and information technology, target tracking has developed to be an advanced technology, which can automatically distinguish objects in real-time from the video image, acquire and predict the target position information, and track the target movement. It plays very important roles in military, industry, safety guard, intelligent transform, medical, and science researching, etc, and it is of a wide development prospect as well.Target tracking is a very important but difficult problem in computer vision, where theory and application technology are not well-considered, many problems and difficult points are still unsolved. First of all, the image is a 2-D plane projection from a 3-D real object, that is an ill-condition problem. Next, the target movement often behaves very complicated in real environment, such as target size, shape changing, move speed and orbit, ray changing, the similar degree of target color and background color, and background stability, etc. Therefore, the researching on target tracking under complicated background is of both important theory significance and application value.In this paper, we studied the target tracking under complicated environment from the main tracking-related four areas (movement target picking up, target area segmentation, target feature acquiring, target describing, and target tracking process). A series of positive results have been obtained, but still there are some problems and inadequacies. The main work and contributions are summarized as follows:1.A new method of color image segmentation based on self-organizing feature map neural network is proposed and studied. The first step includestraining the net with RGB values of the image elements and getting the density distribution graph of whole feature mapping points. In the next step, by means of self-organizing map analysis, the clustering number and individual centers are found. Finally, every element is calculated and classified according to the distance competition. In addition, an adaptive self-organizing feature map (SOFM) network is constructed by using clustering validity function based on Fisher distance, and applied to image segmentation. The experiment results show that this method is effective on color image segmentation, which is satisfactory for extracting color feature for target tracking.2.Genetic Algorithm and two-dimensional entropy method are studied, and a new method of image segmentation is presented. For one situation of the number of segmentation regions is knowable, a fixed code-length GA is applied to image segmentation. For another situation of the number of regions is unknowable or uncertain, an improved Genetic Algorithm of changeable code-length is given. In the improved GA, a fitness function includes classes of segmentation in it, which is the length of one chromosome. Then it realizes optimizing classes of segmentation as soon as optimizing every threshold. Therefore we attained self-adaptive segmentation based on changeable code-length GA. According to the experiment result, the conclusion can be made that the applied segmentation method shows highly validity and speediness.3.The object extraction based on the moving area is studied. In this section, two consecutive frames subtraction is used to calculate the moving area, and the algorithm of combining the two consecutive frames subtraction and the background subtraction are used to detect the moving object. Then the feature extraction of the detected moving objects is done. The edge is the feature for rigid objects (the edges are used to the edge matching tracking algorithm) and the color for the flexible object (the color is used to the mean-shift tracking algorithm and the particle filter tracking algorithm). The color feature is extracted by clustering segmentation based on the SOFM network, so that the moving objects is segmented using the color feature, in which major color is selected automatically,partly solving the problem of the tracking initialization .4.Several visual target tracking algorithms under complicated environment, such as occlusion, illumination varying and so on, are proposed.1) A feature-correlation-matching tracking algorithm is proposed to solve occlusion problem for the rigid target. Choosing the edge or intensity to be the matching feature depending on the intensity characteristic of the target, then edge matching algorithm or intensity matching algorithm based on multi-block are adopted. The edge matching method determines the displacement vector between two successive frames by the optimal matching between the edge template and the current unoccluded edge. On the other hand, The intensity matching algorithm based on multi-block is used to get blocks adaptively dividing with distinct feature which can estimate the occlusion region accurately, overcoming the false occlusion estimation of using fixed block, since it may has similar intensity with the occluding object. Then target is tracked by the remained unoccluded blocks. The algorithm can track the rigid target accurately in real-time, but with limitation of tracking non-rigid target which deformed greatly.2) To solve the above limitation, an improved adaptive particle filter tracking algorithm is proposed. The remarks of the algorithm are as follows: (1) An efficient occlusion detection method is proposed, changing the dynamic model under occlusion, each particle undergoes its own independent Brownian motion to ensure the reliability of tracking under occlusion. (2) The algorithm adaptively changes the variance of Gaussian in the dynamic model and the particle number to guarantee the validity of particles and overcome high computational cost and particle degeneration problems in some degree. (3) Updating the color template real-time in non-occlusion condition to adapt the illuminating variation and guarantee the accurate tracking. The algorithm ensures the reliability of tracking in the complicated environment especially under occlusion, but the real-time property is not satisfactory.3) Aiming at solving the problem in 2), an algorithm which combining the improved mean shift algorithm and particle filer is proposed. Overcoming thedeficiency of the generic mean shift algorithm, the improvement is to let the target point which accords with the color template have the largest weight (regardless of its probability in the histogram), and to make the target density at peak and to overcome the problem of serious clutter and occlusion. Meanwhile combining the above algorithm and particle filter, an improved mean shift algorithm is used to track accurately and quickly in the usual condition. Only in the serious different color occlusion that the target has little points left, the improved particle filter algorithm is adopted to guarantee the robust and real-time tracking.At last, the main work of this paper is summarized. Also, the inadequacies and further research directions are proposed.