Foreground-background segmentation and obtaining robust real-time foreground matting in computer visuals is important and is difficult to achieve. Various bag-of-words methods have been applied in this area, such as optical flow, temporal differencing, and background subtraction. Noteworthy among them, background subtraction can efficiently segment the foreground image from a dynamic video sequence. Most of the aforementioned methods are weak in achieving real-time segmentation. In this paper, we propose a real-time foreground-background segmentation method using a Codebook (CB) model based on GPU programming. Firstly, the background model learns from a video sequence recorded using a static camera. Next, the foreground pixels are subtracted by comparing the incoming video frames against the background model in the GPU, after which the background is updated. Finally, we apply erosion and dilation operations to remove noise from the segmented results. We execute our improved method for greater time efficiency through GPU utilization. Experimental results demonstrate the advantages of our improved Codebook model.