A saliency based approach for foreground extraction from a video

In the field of Computer Vision, success of the most research works like detecting or recognizing some object partially or sometimes completely depends on the success of a decent segregation of foreground from background. In this paper, we describe a novel approach of foreground extraction from video based on saliences that handles the motion in foreground, background or in camera considerably well. We assume that foreground objects are less likely to be positioned in the border. We use the color contrast information to identify the foreground by using visual saliency. Using our assumption, we construct a border based saliency which improves our overall visual saliency information. We also use the optical flow information from video frames to find motion of regions. Again, using the same assumption, we construct a background model using the flow magnitude and angle information of the border pixels to suppress unwanted background movement or camera motion. We then mix up the both information of visual and motion saliency values to get a grayscale saliency map of foreground. We then apply Otsu's thresholding to extract binary foreground mask from saliency map. We remove some noises from the frames using contour detection and comparison of regions to get the final foreground mask. Both qualitative and quantitative experiments are performed on a benchmark dataset which show that the proposed approach performs quite well for practical use.