Pixel-Wise Human Motion Segmentation Using Learning Vector Quantization

This paper proposed an efficient human motion segmentation algorithm with pixel-wise accuracy. Our aim is to solve the problem of separating human image as object of interest from the background image. In our approach, every pixel of a video sequence frame is considered to be a 5-dimensional vector, consisting of pixel position coordinate components (x,y coordinates) plus pixel color information in HSV (Hue, Saturation, and Value). First, the human assistant is employed to create the reference frame of desired human object of interest. This step is done only at the first frame of video sequence. The Kohonen Learning Vector Quantization (LVQ)[1] is then used to give optimal class region decision between the human object class and background class by training its codebook vectors, supervised by reference frame. The segmentation result is generated by doing vector quantization of LVQ codebook vectors to all pixels of image frame. Finally, for adapting the human object class movement in succeeding frames, LVQ codebook vectors are updated periodically by feeding back the result of the last segmentation into the training step. This paper also presents proposed segmentation algorithm performance to some MPEG-4 video test.