Marker-Less Human Body Tracking Using Locally Affine Invariant Contour Matching

Abstract

This paper proposes an affine invariant contour matching method for human body tracking without markers. The contours of human bodies are extracted using image segmentation or background subtraction methods. The matching of contours is based on a locally affine invariant contour descriptor, which can approximate the articulated motion of human bodies. The matching algorithm using the descriptor is simple and fast, wherein an outlier rejection scheme using the geodesic distance along contours is also embedded, making it robust to self-occlusion and viewpoint changes of human poses. The proposed method can track articulated human motions with large degree of pose variations; this is superior to many marker-less human tracking algorithms in the literature. Real-world human motion sequences are tested to demonstrate the performance of the method.