Quasi-periodic action recognition from monocular videos via 3D human models and cyclic HMMs

This paper proposes a system to recognize quasi-periodic human actions from monocular video sequences. First, each input video frame is analyzed and estimated to generate the best 3D human model pose which consists of a set of 3D coordinates of specific human joints. Next, these 3D coordinates for each frame are converted into corresponding 3D geometric relational features (GRFs), which describe the geometric relations among body joints of a pose. Finally, we train a cyclic hidden Markov model (CHMM) for each action based on the vector quantized 3D GRFs, and the trained CHMMs are used to classify different quasi-periodic human actions. The experimental results indicate the effectiveness of the proposed system in terms of the view point invariance, the low -dimensional feature vectors, and the encouraging recognition rates.