Human Action Recognition Based On Convolution-GRBM

With the rapid development of the technology of pattern recognition and machine vision,fast and accurate identification of human behavior in the video has become an important research direction,which has a broad application prospect in intelligent surveillance,human-computer interaction,medical care and other fields.Through the analysis of the relevance of human images in the video sequence,human behavior recognition process extracts the characteristics of change,in order to determine the human behavior patterns.Feature extraction and behavior classification are the two key steps of human behavior recognition.In practical applications,due to the coverage of the target,dynamic background,moving camera angle and illumination changes,there are many factors effecting feature extraction.The existing methods based on traditional manual design have been unable to apply the increasing requirements of feature extraction.However,based on the depth learning model Gate Restricted Boltzmann Machine(GRBM),we can learn the high-level abstraction of data from the input video,so it has become a hot research direction of human behavior recognition.But due to the large number of model parameters,the deep learning method suffers a huge amount of complex computation during the training process.Under this circumstances,it can only be applied in simple human action and can not meet the requirements under complex action and background.According to the feature of human behavior recognition that the statistical property between adjacent frames and some parts of one image in video is stable,a better idea with convolutional neural network model is proposed here.It can overcome the weakness of the existing methods and shows a better ability to extract features than other models.The main contents of this paper are included as follows:Firstly,by grasping knowledge of GRBM model in deep learning and the convolutional neural network technology,combine the two technologies together.Getting different characteristics of input layer through different convolution kernel extraction in order to improve the extraction ability of specific feature model,and achieve a better recognition of human behavior.Then,joining the pool operation,we aggregate statistics on the characteristics of the convolution output at different locations to reduce the output characteristic of convolution layer dimension.As a result,it can overcome the shortcomings of original models such as too many parameters and over-fitting,reducing the complexity of human behavior recognition.In the end,build the model and deduce the formula,then begin to train the model with the test sets.Secondly,we study human behavior recognition process and select the support vector machine(SVM)as the classifier model to complete the system.Then we test it in human behavior test library.Finally,we can find the C-GRBM model proposed in this paper can be proved more effective and advanced than GRBM model in human behavior recognition.