Research And Implementation Of Action Recognition Based On Serialization Convolutional Neural Network

In recent years,action recognition has more effect on our life,this thesis uses the current popular approach of deep learning for action recognition research and application.Deep learning is a hot research area recently among the researchers.and the convolution neural network(CNN)is the representative of deep learning,so a large number of researchers are working on CNN.Compared to the traditional neural network,CNN has many advantages.For example,less parameters,rotation invariance and translation invariance and so on.CNN is also a kind of end to end network,doesn't design manual features,and does not need manual features.Firstly,it needs a large amount of data samples,and then plays tag samples.Secondly,we put into the network.Thirdly,we let the convolutional neural network learn the characteristics of the sample.Finally we can get good recognition rate.However,for making good performance we need to have very good network structure design.Next we focus on building a reasonable network model structure in this thesis,then we use some related techniques to make the network structure obtain a better recognition rate.The main contents of this thesis include:(1)For the shortage of the original 3DCNN network,in this thesis the network structure is optimized.the structure of the network add the MLP layer,which is the layer of a multi-layer perceptron,and the layer without increasing the depth of the network improve the ability and abstract generalization.In addition,sampling operation is optimized,so that it can sampling operation in the dimension of time.In addition,the network has the advantages of sampling in 2D,extended to the time dimension,and the network has no deformation in the dimension of time.it also greatly reduces the time dimension calculation,and greatly improves the network the performance.In addition,the network uses Relu activation function instead of tanh activation function,so that activation function become from saturated to unsaturated activation function,the network in the process of training the training time is greatly reduced,and increase the convergence speed,and activation by relu feature map function is sparse,the generalization ability is very strong.In addition,the optimized network's input reduces the 4 channel information,increases the design of a stack of optical flow information,and reduces unnecessary inputs and the amount of computation and improves the performance of the network.(2)This thesis focuses on a multi task recursive convolution neural network model,multi task recursive convolution neural network designed in this chapter combining the multi task learning theory into the network structure establishs the structure of network.It includes a deeper improved VGG network structure and LSTM structure.For solving the neural network over fitting in the process of training,it joined the Dropout technology,so that it let network have ablity of extracting long time of video and multi-task classification.(3)Finally,in this thesis,it mainly describes a Super 3D-CNN the network behavior recognition system,then it describes the framework of behavior recognition system.In addition,in my dataset of fighting it obtains good effect.It also analyzes the performance of the system,so that this system has reached the requirement of real-time monitoring.Finally it introduces the behavior recognition system advantages and disadvantages.