基於空間時序隱馬爾科夫模型的人體三維動作捕捉數據行為識別

This paper presents a novel Hidden Markov Models, called spatial-temporal hidden markov models, for the recognition of human actions from 3D motion capture (MoCap) data. We exploit both the spatial dependency between each pairs of connected joints of the articulated skeletal structure and the continuous temporal movement of the individual joints. Each action is represented by a sequence of 3D joints positions, and a single spatial-temporal hidden markov model is learnt to capture these two types of dependencies for each action class. Results of recognizing 11 different action classes show that our approach outperforms traditional HMM.