Action recognition of motion capture data
Na Lv, Zhiquan Feng, Lingqiang Ran, Xiuyang Zhao · 2014
With the advancement of motion capture technology, 3D skeleton data is easier to be obtained. 3D skeleton data has the advantage over traditional video data for the reason that it is less affected by illumination, complex background, self-occlusion and noise. 3D skeleton data brings new opportunities and challenges to the action recognition research. In this paper, we propose a new method for action recognition of motion capture data. We use relative velocity of all the joint pairs to encode the kinematic characteristics and the primary vector decomposed from Motion Sequence Volume(MSV) to represent the distribution of joint positions in the motion sequence. The extracted features are fed into a Spectral Regression Kernel Discriminant Analysis(SRKDA) classifier to identify motion types. In the experiment, our method obtains higher recognition accuracy than the state-of-art methods.