Zhejiang GongShang University, 310018 Hangzhou, China

---

Abstract—Human pose recognition algorithm for monocular video was proposed to model human part parameters using video features combination with 3D motion capture data. Firstly Constant Part Appearance (CPA) Models and threedimensional motion data projection constraint graph structure was defined. To simplify the reasoning process, a constraint graph of the spanning tree construction algorithm and the balancing algorithm were proposed. Combination with the proposed function mechanism, spanning tree of constraint graph and Metropolis-Hastings method, human motion under monocular video can be tracking and recognition, and inferring the 3D motion parameters. By using data-driven (Markov chain Monte Carlo MCMC) and constrain map, human motion limb recognition algorithm is proposed, and the method can be applied to data-driven online human behavior recognition. Experimental results show that the proposed method can recognize human motion action automatically and accurately in monocular video.

Index Terms—human action, Markov chain, belief propagation, 3D human action estimation .

[PDF]

Cite: Yi Ouyang and Jianguo Xing, "Human Action Recognition algorithm based on Minimum Spanning Tree of CPA Models," Journal of Software vol. 7, no. 7, pp. 1577-1584, 2012.