The Study Of Human Tracking With Multi-cameras

With the development of computer technology, object tracking has been an important research subject in such domains as pattern recognition, image processing, computer vision and arm guidance. Object tracking promises a wide range of applications in military affairs, visual surveillance, intelligent traffic and medicine. This dissertation focuses on some key technologies about human tracking from surveillance video with multi-cameras. The subject aims for attaining some information intelligently, such as the size, the locality and the color information of each person by analyzing the surveillance video. It could provide evidence for the high level analysis and understanding of human behavior through persistent human tracking.This dissertation covers the four aspects, that is, human region extraction, human feature extraction, human occlusion disposal and human tracking. The multi-cameras are used to extend surveillance range. The human tracking is accomplished by the analysis of multi-camera videos. This dissertation focuses on tracking the same person in a single camera and across cameras.The main contents of this dissertation are as follows:1. With respect to the human region extraction, background model is built firstly, and then Graph Cuts theory based object extraction method is applied. Afterwards, human region is detected by object classification method. At last, shadow is deleted by the principle that shadow has similar chromaticity but lower brightness than those of the same pixel in the background image.2. The approach that combines color and space information for tracking human in a single camera is proposed. The color model in this paper is built by combining the improved mean shift algorithm with the non-parametric kernel density estimation theory based color density function estimation method for the first appearance person in the field of view. Head detection is used to segment groups of people into single person if they exist at the first time. The improved mean shift algorithm is then done to the separate person. Color density function is estimated through the non-parametric kernel density estimation theory. Thus, the color information based human tracking is realized. Simultaneously, the distance matrix is calculated to realize the human match between adjacent frames. It could realize the human tracking when persons have on the clothes with the same color.3. Human tracking with multi-cameras in the visual surveillance system is equal to human match in essence. It is suitable to assume that the human surface is local plane due to its small size in the video image. MSER (Maximally Stable Extremal Regions) based approach is proposed to accomplish the match of the same human among different cameras and increases the accuracy of the match. It firstly does ellipse fitting to each MSER, and then selects some ellipses which meet constraints. Furthermore, these ellipses are normalized to unity circles by whitening of covariance matrix. At last, right matched ellipses are gotten by rotation invariant vectors calculation and histogram density estimation. Thus, the human tracking across cameras is realized.4. In order to solve the occlusion problem with multi-cameras, the proposed approach uses 3D information. It makes the most of space information. It firstly samples the pixels from the extracted human blob in one of the video images, and then finds the pixels that match with each of the samples in the other camera video image. Afterwards, the 3D points corresponding to each pair of the matches are found in the world coordinates. At last, the 3D points are clustered and the Gaussian smoothed histograms are created. This approach segments the occluded persons by the Gaussian smoothed histogram model, and then performs human match. In addition, the human tracking is realized efficiently.