Self-adaptive background modeling research based on change detection and area training

Abstract

The detection result of traditional Gaussian mixture model algorithm easily becomes fragmentary and exists shadow, the fixed number of Gaussian component leads to bad performance. Aiming at building a robust background, using background difference method, self-adaptive threshold segmentation and Gaussian mixture background modeling algorithm, a self-adaptive background modeling method is proposed. The background difference method and self-adaptive threshold segmentation classify the pixels in each frame into moving targets and background area. When training the background model, this new algorithm keeps the Gaussian mixture background model of the pixels in moving target area unchanged and never build new Gaussian component for this area. Background area is updated in regular way, make the number of Gaussian component for each pixel in this area to be self-adaptive, keep the Gaussian component of background model only updated by the real background pixels, improve the performance of the algorithm and validity for background constructing. Experiments show that the background model built based on the proposed algorithm has good adaptability for video sequences with uncertainties, it can eliminate the shadows and quickly response to the change of actual scene, the computing speed of this model improves a lot as well.