Research On Human Action Classification Algorithm

With the development of human motion data acquisition technology,the research of human motion based on data has attracted more and more attentions.The research of human motion has great application space in fields of medical rehabilitation,virtual reality,film and television,etc.The research of this paper mainly includes two problems: visualization of human actions and classification of human actions.The visualization of human actions aims to show the human actions represented by data with rendering engine,compared with the traditional video recording and display of human actions,visualization of human actions has the advantages of a small amount of data,good portability and convenient observation.The classification of human actions aims to classify unlabeled human actions,and it is based on a large number of labeled human actions,in this paper we use a human action classification algorithm based on dynamic time warping(DTW)algorithm and a human action classification model based on the Long Short-Term Memory network(LSTM)to realize the classification of human actions respectively.In the part of the visualization of the human actions,we display the human actions with the Ogre rendering engine,and provide the data interface to control the movement of the virtual 3D human model.After soluting the probloms of action display,coordinate transform,gesture calibration,model calibration,and camera control,we realized the visualization of human actions in the format based on quaternions and the format of BVH.And with the data interface we provide,the visualization system has a good scalability for other human action formats.In the part of the human actions classification algorithm based on DTW,we selected 5 bones in the human skeleton model to quickly calculate the distance between two human actions,the 5 bones are throax,left lower arm,right lower arm,left lower leg and right lower leg.we calculate DTW distances between two human actions on each bone in the 5 bones,and define the distance of two human actions with the average DTW distance on the 5 bones.Then,we use the nearest neighbor algorithm based on DTW distance to classify human actions,the algorithm classify each unmarked human action into the class of its nearest neighbor in labeled actions.In addition,in order to compute the distance between two sequences of quaternions,we propose the definition of the distance of quaternions with relative rotation angles.And in the part of the human actions classification model based on LSTM network,we represent human actions as a form of time series and input one human action by frame order into two LSTMs without output layer,one is forward LSTM and the other is backward LSTM,and the output of hidden layer of LSTMs will pass the mean pooling layer and the logistic regression layer to get the final classification result.We implement our classification model and train it with the popular deep learning platform Tensor Flow.We use human motion capture database HDM05 to validate our two classification methods,and the accuracy rates of the two methods reach 94.23% and 94.84% on test set.