Action Recognition Of Video Based On Distributed System Of Deep Learning

Action recognition of Video data is to identify the characters in the video which is the basis of solving video surveillance,human-computer interaction,video emotion analysis and other relating issues and it is a hot area of deep learning.With the increasing demand of video action recognition,how to construct a fast and efficient large-scale video recognition framework and how to design a better model of video action recognition are particularly important at the present stage.There has been no uniform solution yet.In view of the above two problems,this thesis makes an in-depth study to realize the distributed deep learning system based on remote GPU call,and proposes an improved video action recognition model.Finally,the thesis uses the distributed depth learning system to train the video action recognition model to verify the feasibility and validity of the two.The main contents of this thesis are as follows:1.We construct a remote GPU call using the virtualization method of API redirection,and then accelerate the training of deep neural network on the basis of remote GPU call.We use the traditional distributed system built on ZeroMQ to achieve multiple GPU remote call to form a distributed system of deep learning based on multi-GPU remote call,and improve it on the deep learning library cuDNN,P2P,network communications and other aspects.This method of building a distributed depth learning system can be used as a large-scale video action recognition framework.The biggest advantage of this framework is the single-CPU multi-GPU code can be quickly extended to the distributed environment which does not need to be modified or only a small amount of changes.2.By Improving the traditional method of video action LRCN,the new method iRCN does not need to manually design features of the original video.This method uses the global sampling feature by dividing the whole video into different time periods for image sampling.We use 3D CNN to replace the original 2D CNN to extract the motion characteristics of each stage in the vicldeo.And then we use biLSTM to replace the original LSTM.in order to obtain the correlation im provement of all the motion characteristics in the time domain,and funallvy Softmax is used as the lcoss function.At last.iRCN,obtains the Ccorrect rate of 85.6%and 56.6%respectively on the data set UCF-101 and HDMB-51.which is the highest correct rate in all video action recognition methods which do not use features manually designed so far.3.The distributed system of deep learning can quickly implement the distributed expansion of data parallel and model parallel.We use the distributed system of deep learning to accelerate the training speed of the improved action recognition model.In this thesis,the accelerated training of video action recognition includes two aspects,motion feature extraction of video based on improved MapReduce and distributed training of the complete model.Proved by the experiment,the distributed system of deep learning can effectively improve the training speed without affecting the correct rate.The experiment can both prove the feasibility and effectiveness of the distributed deep learning system and the improved video action recognition model at the same time.Finally,the advantages and disadvantages of the system are analyzed by the model parallel in the distributed environment.