Research On Background Modeling Algorithm Based On Codebook Learning

In actual surveillance video scenes, due to the changes of ambient light, the moving targets disturbances and changes of the scene itself, it is difficult to ensure the accuracy of the background model. Therefore, the establishment of a background model for changing scenes in real time is significant. In view of this, we did the following work in the thesis:1. On the basis of the classical codebook background model, to present a foreground detection system based on adaptive codebook model which contains the scene information.2. Update the codebook background model according to the HOG features of pixel neighborhood so as to eliminate spurious foregrounds caused by sudden illumination changes. HOG feature descriptor has good illumination invariance. We propose a zoning separation strategy based on a pixel neighborhood. By comparing the HOG distance, we can fand the pseudo foregoings caused by illustration changes and eliminate the spurious foregrounds by updating the codebook model.3. Combined with the gray threshold and the cross correlation coefficien feature, our system can detect the shadow, and the adaptive codebook model can update itself to eliminate the shadow of moving targets. The shadow area has a smaller gray value and has an approximate linear correlation with the scene4. According to the proposed foreground detection system, we design an overall experimental procedure. In order to test our algorithm, we compare this model with other background model(MOG, classic codebook, etc.). We carry out detailed comparative analysis of the efficiency in terms of time, memory usage, and the foreground detection effect.The improved codebook background model has a better ability to eliminate spurious foregrounds caused by sudden illumination changes and shadow.