The Research Of Action Recognition Based On Deep Learning

The research of action recognition is a central issue in the field of computer vision,because it's widely utilizing in video surveillance,motion analysis,aided medical,virtual reality,human-computer intelligent interaction,etc.Action recognition is enabling computers to identify and analysis features extracted from videos or images by some methods.While there are lots of methods of action recognition,developing a robust method is still a challenging problem.The main difficulty of action recognition is designing an effective model which should not only distinguish the object and action from the background but also be robust to its variations of the action,such as actions are partial occlusion or the same action occurs in different environments.At present,many hand-designed models have a certain ability of recognition for specific actions but are restricted in more general actions of the various changes.Now,building a network based on deep learning to automatically extract features is receiving attention in action recognition.We propose a convolutional neural network model based on deep learning for action recognition in this paper.Firstly,we learn a very deep two-stream convolutional neural network model which includes spatial and temporal very deep convolutional neural network by training the UCF101 dataset and test its performance.The input to a spatial convolutional network is RGB image,and the spatial network captures appearance information of objects from still images;the input to a temporal convolutional network is optical flow image,and the temporal network extracts motion information of actions.Secondly,we present multi-stage training strategy for very deep convolutional neural network,jointly classifying actions in videos using several classifiers.Besides,after we get features from very deep convolutional neural network,we use support vector machine classifier to classify actions,improving recognition accuracy efficiently.Finally,we conduct experiments on the HMDB51 dataset through transfer learning method.In this paper,the proposed action recognition algorithm based on deep learning has been evaluated on two benchmark action datasets.Experimental results demonstrate that the proposed algorithm gets higher accuracy and more robust than the state-of-the-art methods,especially suitable for recognizing actions in complex environments.