Emotion Recognition Based On3D Human Motion Data

Emotion recognition based on3D human motion data is a new cross subject ofkinesiology and pattern recognition, and its research is in beginning stages currently.This thesis studies emotion recognition thorough exploration and research, includinghuman motion modeling, generation of Period sequence, action recognition andemotion recognition.Firstly, based on investigation of human motion modeling technology, ahalf-sphere assumption has been proposed, which pointed out every joints’ relativemotion is limited on a half-sphere surface, then utilized half-sphere modeling methodto standardize coordinate of joints. And10typical direction vectors were chosen, inorder to represent instantaneous human motion state.Secondly, Period sequence was proposed to approximate joints’ motion trajectory,in order to reduce data’s scale efficiently and separate the information of trajectoryand velocity, which corresponding to arguments of action and emotion. Then based onDouglas-Peucker algorithm, a3D curve approximation algorithm to realize thegeneration of asynchronous Period sequence was proposed. Period synchronizealgorithm was proposed to solve asynchronous one’s insufficient performance in theprocessing of recognition.Finally, based on varied human motion data, action-emotion dataset was built,including typical fifteen action label and five emotion label. Continuous SubsetMatching algorithm was proposed to recognize action or composition of actions.Signal waveform was used to represent emotion arguments, and then classify theemotion.Experiments show that, compared to conditional recognition method in frame unit,the method proposed in this thesis can reduce computing complexity of recognitionprocedure efficiently, and eliminate the disturbance of sampling deviation, whichresult in higher accuration.