Research On Human Action Recognition Based On Neural Network

In recent years,the amount of video data has increased explosively,and it has been used in many fields,involving security,monitoring,entertain ment and so on.Faced with a seemingly endless stream of video data,the traditional manual data processing method cannot meet the demand of researchers.Therefore,it is of great value in both theoretical research and actual practice to realize the recognition and comprehension of video information with the help of the powerful storage and computation ability of computers.This paper firstly describes human action recognition algorithms from two aspects: feature extraction and action classification,especially focusing on the detection and description of spatio-temporal interest point as well as action classification methods.Then this paper introduces the method of human action recognition based on neural networks from these two aspects respectively.(1)Traditional convolutional neural network(CNN)performs convolution and pooling operations on two-dimensional images to extract feature vectors,which cannot preserve the temporal information of input video sequences.In order to make use of the temporal information,3D CNN performs 3D convolution and 3D pooling operations on three-dimensional spatiotemporal cubes,obtaining spatio-temporal information efficiently and effectively.Based on that,this paper proposed an algorithm framework of human action recognition based on 3D CNN.(2)For different videos,the number of feature vectors extracted by the CNN is also different.Supervised classification method requires a unified expression of these feature vectors,which may dilute the key features in this process.To solve this problem,this paper presented a multi-instance learning approach to train the classification model to make the classifier able to recognize tinny differences between similar actions.Finally,we trained a 3D CNN and analyzed the experimental results.