Research Of Background Modeling Algorithm

As a key point and prior step in most video analysis, background modeling is one of the most meaningful research topics. According to different capture method,s videos can be divided into fixed-view video (captured by a standing camera) and changing-view video (captured by a moving camera). Because of the significant difference of these two video types, background modeling methods are accordingly divided into two categories. Background modeling methods in these two categories are discussed based on the analysis background pixel values of two different video types.In algorithms of background modeling on fixed-view videos, most of the traditional methods are based on statistic, they only analyze pixel separately in modeling process, leading to the misjudging in process slow-moving object and reciprocating-moving object. According to the defect and advantage of these existing algorithms, some new methods are presented. One method evaluates the credibility of every pixel values as background by analyzing the stability and the frequency of appearance, and then fixes the credibility by considering spatial relations of pixels. The background value can be found by comparing the credibility of pixel values. At the same time one background estimate method based on image repairing is also presented. This method judges the moving objects area roughly and estimates the background which part blocked by foreground objects according to self-similarity of the image. Compared to existing algorithms, this algorithm adapts to not only environment changes/background disruption but also adapt to slow-moving objects and reciprocating-moving objects.About videos captured by moving cameras, one possible modeling method has been proposed. This method calculates the transformation matrix on the basis of SIFT feature matching then finds pixel correspondence between images on different views by affine transformation. According to this the modeling algorithms used on fixed-view videos will be able to apply to the modeling in changing-view videos.