A Background Modeling Approach Based On Neighborhood Correlation

Background modeling is the basis of detection, tracking, behavior learning and recognition in the field of computer vision, which the Mixture of Gaussian(MOG) and Codebook are currently popular algorithm based on pixel-based background modeling, are widely used in moving object detection video surveillance. However, these methods usually assume the information between pixels is independence, leaving only the time domain information while ignoring the spatial information, which makes the model assumptions background description ability has certain limitations.In this paper, we mainly study the improvement method of pixel-based background modeling. The main work is as follows:1) Introduces the test image dataset of this paper and the method of extracting the model of the neighborhood, analysis the neighborhood model in spatial distribution relationship. It is proved that the “similar when adjacent” and continuity between the models of neighborhoods, and the correctness and reliability of the current pixel model by using the neighborhood information.2) Proposed a neighborhood correlation(NC) background modeling method: according to the “similar when adjacent” principles, use the pixel models in the neighborhoods and itself models of “sharing” mode, so as to achieve the pixel distribution in temporal and spatial better description ability, overcome the defects as the MOG and Codebook pixel-based background modeling method in the complex background description ability, so as to improve the effect for the prospects detection.3) Add the adaptive neighborhood range mechanism to the NC method, the paper proposes an adaptive neighborhood enhancement background modeling method(ANC): establish the confidence model for each pixel, and to adjust the neighborhood range adaptively. Solved the problem of over modeling when in the simple background scene and the lack of modeling when in the complex background scene in the NC algorithm. Experimental results show that compare to the general algorithm, the improved algorithm not only more accurately describe the change in pixel value under a plurality of modes, but also has good robustness, noise immunity, and more adaptive to the complex background.