An Effective Framework for Human Body Pose Estimation using K-Best Viterbi Search Method

Estimating human body pose in monocular images is important for many image understanding applications. In this paper we propose a robust and efficient framework for human body pose estimation. In contrast with the conventional methods (Felzenszwalb and Huttenlocher, 2004; Lan and Huttenlocher, 2005; and Sigal, 2006), our algorithm has several advantages: 1) no learning process of human model prior from image database is needed, and no limits on the video scene are required. 2) By integrating the generalized integral image and distance transforms into the tree-structured model, the time complexity of the inference could be greatly reduced. 3) A post-process called K-best Viterbi search is proposed to add global constraints elegantly to get reasonable results and accelerate searching in the solution space. Our framework consists of three steps. The first step is to use the tree-structured human model to reduce the complexity of representation of human body. Second, the concept of the generalized integral image is applied to integrated the multi-cues, such as face and limbs, to enhance the robustness. Finally, we present a K-best Viterbi search method applied to post-process. We test our framework on synthesized and real-world database respectively, and receives satisfactory results.