Automatic estimation of body regions from video images

Abstract

In our approach video-based recognition of sign language requires the extraction of sign parameters. Each sign can be characterised by means of manual (handshape, hand orientation, location and movement) and non-manual (trunk, head, gaze, facial expression, mouth) parameters. This paper introduces a software module which is as a part of the developed automatic sign language recognition system able to extract relevant body regions from digitised video images. The recognition of body regions is crucial for determining location of signs. The proposed software module uses a rule-based system for analysing the body contour in order to compute the 2D position of the shoulders, the top of head and the vertical axis of the body. Based on these results the position of the eyes are calculated directly from the segmented face of the signer. The position of the remaining face- (nose, forehead, mouth, cheek, chin) and trunk regions (shoulder belt, chest, belly, hip) are determined by means of two estimators, where a-priori known geometric data of the face and fuzzy technique are used. Experiments indicate that our approach leads to good estimation of body regions, which we all compute in real time.