Research On Human Detecting And Matching Technology

In recent years, based on the intelligent visual surveillance system of single camera, the development of human movement detection and tracking technology have become increasingly mature, at the same time, because of increasingly requirement of monitoring network of multi-cameras, we paid more attention for the research of intelligent visual surveillance system of multi-cameras. At present, we can divide surveillance system into overlapping view and non-overlapping view based on multi-cameras of view domain. Considering the actual difference between the reality of monitoring scope and camera resources, this thesis choose to research on human detection and matching technology under non-overlapping view of visual surveillance, the main research contents are as following:(1) Research based on background of human detection technology. This thesis used background difference algorithm to detect human movement, combined with the adaptive background modeling to update the background of the current image automatically. But it has the effect of light and noise in the actual scene. This thesis removes the shadow and searchs connected areas to extract clearer human movement.(2) Extraction of human body characteristics. This thesis introduced the common features of human movement, such as spatial characteristics, geometric characteristics and statistical characteristics, and analyzed advantages and disadvantages on describing each human body characteristics and the extraction methods. It focused on the human body statistical characteristics, such as color histogram, gradient histogram. It had a profound impact for moving target tracking and matching.(3) Research on human movement tracking algorithm. This thesis based on the analysis of thought and the advantages and disadvantages of the Meanshift algorithm which has been widely used in human movement track, used Camshift algorithm to track movement human body of single camera, it overcome non-accurate tracking situation caused by the Meanshift algorithm due to target dimension changes.(4) Research on the human body matching technology. For probability model of transfer time of Multi-cameras of non-overlapping view, we use Gauss probability density curve estimates, in a way not only eliminate mismatch of between objects, and greatly reduce the amount of matching. Finally, we use the weighting sum algorithm to merge the transfer time probability model and color model, enhance the accuracy of matching.