A Gaussian mixture based hidden Markov model for motion recognition with 3D vision device

Abstract

Recognizing human motion is an import research area in computer vision. Hidden Markov Model (HMM) is widely used in object recognition and tracking due to its richness in mathematical theory. However, the poses of artistic performance are more complex compared to conventional human motion. To address this issue, in this paper, we present a Gaussian Mixture based Hidden Markov Model (GMM-HMM) for the artistic motion recognition of Peking opera. In order to filter the abnormal data and repair the missing data, a local weighted linear regression method is designed for improving the motion data captured by the 3D vision device OptiTrack. We construct a GMM-HMM model as the recognition method, which can get the exact number of hidden states based on the key frames to achieve high accuracy of recognition. In addition, we apply multiple multi-dimensional gaussian distribution functions to train the motion data, avoiding the large computational load and the discretization error. The results show that our method can identify and enable the interactions with some important motion movements in Peking opera performances.