Segmentation of moving human body parts by a modified MAP-MRF algorithm

Abstract

Using a dense motion vector field as the main information the authors develop a region segmentation algorithm in which each region is matched to a four-parameter motion model. Based on Markov random fields the segmentation model detects moving parts of the human body with different apparent displacement such as the hands. The motion vector field has been estimated by a Baaziz pel-recursive method and considered together with others sources of information such as intensity contours, intensity values and non-compensated pixels as inputs of the Markov random field model. The maximum a posteriori criterion (MAP) is used for the optimization of the solution, and performed with a deterministic method: iterated conditional modes (ICM). Results on segmenting and classifying real sequences are shown and, based on a roughly defined directional dictionary, one application is pursuing the use of the segmented regions as commands for a virtual robot. The classification is based on the correlation coefficient (between the trained sequences and others) of wavelet coefficients, of the projected sum of the intensity of the segmentation field (in its binary version).