Bag Of Words Based Background Modeling

With the improvement of processor performance and the development of video pro-cessing technology, video surveillance will gradually realize high-definition and intellec-tualization in the future. However, due to the constraints of computing resources, termi-nal surveillance equipment cannot meet the processing requirements of high-resolution video. In this paper, we adopt the ideas of Bag of Words (BoW) models to solve the background estimation for an image sequence captured by a stationary camera, and de-sign a new method of BoW based background modeling. This method not only saves memory, and improves the accuracy of the background model.At first, we present a method of background model based on local BoW. This method establishes an independent BoW model for each block through incremental K-Means clus-tering, as the clustering can properly distinguish foreground from background, its results of foreground-background segmentation are good. However, this method will face a com-plex problem of words mergence when it updates the models.Therefore we propose another method of background model based on global BoW.This method trains a global BoW using K-Means clustering prior to background es-timation, in particular, the training samples can be content-independent. Then each block shares a global BoW model to establish its own background model. In the estimation pro-cess, it only needs to update the global BoW without adding new words. So this method not only greatly reduces the consumption of memory, avoids insertion and deletion oper-ations on BoW, and still improves the results of foreground-background segmentation. It is well suited for high-definition video surveillance.The comprehensive experiments show that the proposed global BoW based back-ground modeling method has many good characteristics, such as low memory consump-tion, high accuracy of the segmentation, versatility of the BoW, etc. It also can inhibit the effects of shadow and luminance variance. Of course, this method has its disadvantages too. For instance, the block matching on the global BoW is inefficient, so an optimization algorithm for searching is necessary.