Human Action Recognition Based On Deep Convolutional Neural Networks

The research on human action recognition in the aspects of intelligent monitoring,advanced human-computer interaction,automatic labeling,three-dimensional games,and medical diagnosis,has a wide range of application prospects and potential economic value.Human action diversity,scene noise,the camera motion angle changes and other factors increase the difficulty of human action recognition.Therefore,there are important practical significance to study human action recognition.This paper mainly studies the deep learning methods to recognize human action,the research contents are as follows:Aiming at the problem of human action recognition,this paper mainly introduces the feature extraction and classifier.Firstly,there are non-deep learning methods and deep learning methods in terms of the feature extraction.Then introduce the related knowledge of deep learning and the advantages of deep convolutional neural network.Secondly introduces the classifier based on temporal features and the classifier based on fixed dimension features and explains the advantages of the softmax classifier.Aiming at the problem of Long-term Recurrent Convolutional Network(LRCN)algorithm,its timing learning structure is redundant,and it has long running time.Firstly introduces the Long-term Recurrent Convolutional Network.Then elaborates the gated recurrent unit shorter training cycle and faster convergence rate.we replace the long short-term memory unit with the gated recurrent unit to improve,and The classification result is obtained by the softmax classifier.Experiments show that the improvement improve the accuracy of human action recognition,shortening the running time.Aiming at the problem of Posed-based Convolutional Neural Network(P-CNN)algorithm,that can not better characterize high-level action information,and it has long running time.This paper introduces the Posed-based Convolutional Neural Network,elaborates convolution parameters and pooling parameters impact on human action recognition.We proposes a 3D deep convolutional neural network method to recognize human action.We use relevant techniques to identify the classification results.Firstly,successive 16 frames of the video are divided into a group.Secondly,we process the group data by the gray,gradient-x,gradient-y,optflow-x and optflow-y as multichannel.It effectively training network parameters.Thirdly,the extracted features are obtainedusing 5 layers 3D convolution,5 layers 3D pooling to increase time dimension information,Finally,the recognition results are obtained by 2 layers full connection and softmax classifier.Meanwhile,we regularize the 3D convolutional neural network by scale-invariant feature transform descriptor and motion history edge image.Compared with the other three typical algorithms: i DT,P-CNN,LRCN,this algorithm effectively improves the accuracy of human action recognition and has faster running speed.