Foreground Segmentation for Live Videos by Texture Features
This paper presents a method to extract the foreground images from live videos by means of automatic object segmentation. The parameters such as colour, motion of the pixel and image texture or more specifically texture constraints are used for segmentation. A cellular neural network which combines both colour as well as motion of pixels which varies from frame to frame is implemented which helps in accurately separating the boundaries and thus reducing misclassifications. The global motion of pixels is calculated by computing the forward and backward displaced frame differences (DFDs) with the respect to the current frame. The texture constraint for each pixel to be labeled is calculated from the difference between their corresponding texture descriptors and the texture prior models which is provided by the Local Binary Pattern (LBP) histogram. Finally by means of the randomized texton searching algorithm and graph cut frame work the foreground is extracted from the video.
Cellular neural network, DFDs, Local binary pattern histogram, Texture, Texture prior models, Texture descriptors.