Multiple Camera Target Recognition Research Based On Incremental Learning Method

In recent years, the increasing demand of security monitoring among many fields, the intelligent monitoring system based on multi-camera has attracted more attention from the researchers. With all kinds of video analysis algorithm, the intelligent monitoring system based on multiple-camera has a good application prospect to realize target tracking in the area of the multiple points of view, based on motion target feature extraction and matching. However, because of the complexity of the realistic environment, it is easy to cause faulty match between multi-cameras. How to extract more accurate target information and to improve the matching success rate with more features information becomes the research key point.The main contents are as follows:(1) The algorithm for target detection and feature information extraction based on single camera was analyzed. The background modeling of color information and the contour edge information has been finished respectively, and made the comparison analysis. From above, what has been concluded is that the background modeling based on color feature is more sensitive to the change of light in the scene and greatly influenced by scene disturbance. And the contour extracted from the background modeling based on edge character of target area is not connected. Through the analysis, it is concluded that the background modeling based on the feature fusion of color and edge information can reduce the effects of illumination change and disturbance in the scene and that has been verified by experiment.(2)The algorithm for feature matching are studied, which contains the main color spectrum histogram feature matching and SIFT feature matching, and the new algorithm put forward, based on many feature matching, is the incremental learning method. The algorithm idea is:first to master color spectrum histogram as matching feature of target, according to the result of the match, if the color spectrum histogram matching is not successful, the joint SIFT feature matching and finally on the compatibility of both weight addition, matching through fuzzy judgment, and it has been verified by experiment.(3) The description for multi-camera feature matching based on the part overlapping and the non-overlapping scene has been made in detail. The target has been calibrated based on the method which is to calculate the distance between the objectives and boundary of calibration. And that has been verified by the simulation with the experiment.In this paper, the innovation is that the method based on multiple feature matching algorithm of incremental learning method has been put forward, which is to improve the matching rate between multiple cameras and to reduce the computation complexity.