Human Tracking for Occlusion People in Video Surveillance
|Course||Applied Computer Technology|
|Keywords||human body model background edge model hierarchical dirichlet process back-propagation neural network foreground edge curve behavior model|
In the developmental process of social informatization, intelligent monitoring and tracking system for a single body have gained great development and have a certain application. However, the study of multi-body tracking technology has been a difficult problem for the reason that shadow, light and mutual occlusion between the human bodies may lead to tracking errors or missing. The thesis focuses on the problem of human body tracking with mutual occlusion in the video surveillance, and main contributions and work are described as follows:(1) For motion detection, combining background model and background boundary model, an approach is presented for moving object detection in order to avoid the impact of light and shadow on foreground detection. This model can obtain more complete foreground object and effectively avoid the impact of light and shadow on foreground detection.(2) Mutual occlusion of crowded people makes human segmentation and tracking more difficult in video surveillance. Thus, a human segmentation method combing human model with body edge curve is presented. This method draws human body boundary curves according to the boundary pixels, deals with the boundary curve to determine the width of human head, and finally determines the proportion of the various parts of the body according to the human model. The segmentation method in this thesis is to transform 2D image processing into the treatment of one-dimensional curve, so this method has the advantages of simplicity and good real-time.(3) Human body tracking model in this thesis is presented based on the combination of hierarchical Dirichlet process and back-propagation neural network. The human segmentation method in this thesis can effectively solve the problem of partial human occlusion, but this method can result in bigger defect and distortion in the value of human characteristics. BP neural network can deal with these issues with complex environmental information and fuzzy inference rules and allow samples to have defects and distortion. Moreover, the learning and training algorithm of BP neural network is relatively simple and fast. Therefore, back-propagation neural network is adopted as human body tracking model in the thesis. Tracking system in this thesis is divided into two subsystems, namely online human body tracking and offline network learning. In the process of online tracking, hierarchical Dirichlet process is used to cluster human eigenvectors in current frame and human characteristics data in knowledge base so as to decide whether new behavior patterns are generated. If a new behavior pattern is generated, it will activate offline subsystem to quantize new behavior model and update the knowledge base. As BP network is constantly learned, knowledge base becomes more and more abundant, the performance and adaptability of system tracking will be constantly improved. The introduction of HDP in the process of learning BP network has effectively improved the autonomy and efficiency of the BP neural network learning.