An LVQ-based technique for human motion segmentation

This paper describes a novel approach for human motion segmentation from digital color video sequences. The problem is to separate the human image as target object from its background image in a color video sequence. In our approach, every pixel of a video frame is considered to be a 5-dimensional vector consisting of x-y coordinate components plus 3 color components in HSV color space. The basic idea is to use learning vector quantization (LVQ) defined in 5-dimensional vector space to distinguish the target human object from its background image. We assume that the target human object and its background are classified by hand at the first frame. The initial classification data are used to train the system for generating the initial codebook vectors. These codebook vectors define class regions in the 5-dimensional vector space. For tracking the target human object class in succeeding frames, LVQ codebook vectors are updated periodically by feeding back the result of classification into the training step. This paper also presents performance evaluation of the proposed LVQ-based segmentation algorithm.