Study On Human Action Recognition In Videos Based On Convolutional Neural Network

The study for human action recognition is one of the most important parts in real time monitoring applications with intelligent video technology,which has a great significance for the maintenance of public safety,fighting criminals and national defense.Human action recognition study involves various fields of knowledge such as image processing,computer vision and pattern recognition.Conventional human action recognition technologies need to extract image features in advance,and then use machine learning algorithms for the recognition and classification course,whose main shortcomings are the prior knowledge is required in depth for each feature of the given objects to be extract.Therefore,the extraction of different features can only be implemented by human design in different application scenarios.With the development of deep learning theory,which brings new ideas and infinite possibilities for the image processing implementations.One of the most typical algorithms is the convolutional neural network,which takes the advantages that it can load picture as an input directly,and is no need to design any extra features,now has become a new research focus.By making use of the partially connected network and the weight share methods in the convolutional neural network algorithm,the neural network complexity can be effectively reduced and the training effect of the model will be enhanced consequently,which can be extensively used as solutions for computer vision fields,to cope with the object recognition,face recognition and image classification issues.The main research contents and work of this paper include:In this paper,a stochastic dropout neural network model is developed for human action recognition tasks.Due to the overfitting issue of the convolutional neural networks that exists in the condition of less training samples,several implementations of model averaging are summarized for the depth of learning theory,among which the dropout method is then improved.When using the method to select the hidden layer nodes in the network,a random factor should be set as the selection ratio.It is said that the improved method is random dropout.The stochastic dropout is used to the stage of the classifier for the convolutional neural network model,the certain proportion of neuron weights are randomly selected to be "frozen" in the network training process,to change the the connection order of the neurons as the network is updated,so that the update of the network weight is no longer dependent on the combined effect of the hidden nodes with a fixed relationship.he convolutional neural network based on random dropout is applied to the Weizmann database to test.The experimental results show a higher recognition rate with this method than that of no-modified network,which improves the generalization ability and reduces overfitting greatly.On the basis of the random dropout convolutional neural network model,the network structure is then further optimized,where there are two aspects included.On the one hand,the neuron type of the network model in the feature extraction stage should be adjusted,and by combining with the maxout activation function,the length of the convolutional neural network model then can be increased;On the other hand,a comparison experiment is implemented with different pooling methods in the part of sampling layer of the proposed convolutional neural network model,in order to analyze the effect of different pooling methods of the convolutional neural network on recognition results.The empirical study is then carried out by making use of the Weizmann human action recognition data sets with the proposed network model.Experimental results show that a higher recognition rate can be obtained by using the convolutional neural network model with random dropout+maxout method compared to the basic dropout convolutional neural network model for the human action recognition implementation.