Model-Based Human Pose Estimation and Its Analysis Using Hausdorff Matching

This paper presents a novel algorithm for automatically locating, extracting, and recognizing human poses. The proposed algorithm consists of four stages: (i) training of typical poses in 2D and 3D configurations, (ii) Hausdorff matching for locating the pose, (iii) silhouette representation of the located pose using deformable templates, and (iv) joint extraction from the silhouette. In order to reduce the computational overhead, an a priori obtained training set, Hausdorff matching with principal component analysis (PCA), and a heuristic approach were used. The algorithm was tested on typical human poses such as standing, sitting, walking, crouching, and stretching, and experimental results show the efficiency and accuracy of the proposed algorithm.