Real-time motion recognition based on skeleton animation
Hong Chen, Shuangjiu Xiao, Zehong Tan, Jianchao Lv · 2012
We propose a novel real-time motion recognition method based on hierarchical skeleton model. Its key modules include a self-adaptive training algorithm to boost a strong classifier among the features of rotation quaternions and a dynamic time warping algorithm based scoring method to pyramid match with standard motion class's classifier. For a sequence of recognized candidate motion class, a HMM-based most likely tagging algorithm is proposed in the end of recognition pipeline to work as a smoothing filter. Our method has a remarkable performance as it has high sensitivity, specialty and precision.