Study On Service Action Recognition Based On Two-stream 3D Residual Network

The service quality of enterprises has a significant impact on the scale of customer traffic.In recent years,in order to imp rove the competitiveness of enterprises,State Grid has established the core values of "customer-centered,professional,focused,continuous improvement".Around the core values,in order to better improve customer satisfaction,many power supply business halls use the method of manual supervision to supervise the service actions of employees.However,it is very time-consuming and laborious to use the manual supervision method to evaluate the service action of employees,so it has great application value t o study a method of intelligent identification of servic e action.In recent years,due to the advantages of convolution neural network,such as convenience and high recognition accuracy,convolution neural network has gradually replaced the traditional artificial design method of action recognition,and more an d more industries have applied convolution neural network to practical occasions.However,there are two obvious differences between service action and traditional action.One is that some service act ions are very similar and the range of action is small;Second,the scenarios of service actions are similar,and the types of actions cannot be obtained by analyzing the background information.Therefore,the traditional convolutional neural network for s ervice action recognition will have some limitations.In addition to the limitation,the model size and operation speed of convolutional neural network need to be considered.In order to meet the needs of service action recognition,thesis proposes a service action recognition method based on two stream 3D residual network.In the first chapter,the research status of action recognition at home and abroad is briefly explained.The advantages and disadvantages of the current mainstream indoor action reco gnition methods are analyzed.Combined with the needs and di fficulties of current service action recognition,the convolutional neural network is selected as the basis of recognition method.Firstly,thesis briefly describes the research st atus of action recognition domestic and foreign,analyzes the advantages and disadvantages of the current mainstream indoor action recognition methods,and then selects convolution neural network as the basis of the recognition method combined with the cur rent needs and difficulties of service action recognition,and finally expl ains the main content of thesis.Then the algorithm and architecture of convolutional neural network are analyzed,and the application of transfer learning method is described.Then it introduces the basic principles of the classical 2D and 3D convolutional neural network method and optical flow method,which paves the way for the subsequent explanation of two stream network.Finally,the background subtraction method is introduced.In the third part,the effect of optical flow and image flow in service action recognition is studied,and the effectiveness of two stream network in service action recognition is verified.Firstly,the advantages of two stream convolutional neural network in action recognition are analyzed.Then,aiming at the difficulties of service action recognition,a two stream C3 D convolutional neural network is proposed.Then,the service process including say hello,business introduction,business handling and sending off customers is preset,and the service actio n data set is established according to the process.Finally,16 consecutive stacked images are input to the two stream C3 D convolutional neural network for experiment.The experimen tal results show that although the image stream shows good recognition accu racy and the optical flow reduces the dependence of convolution network recognition on the scene,the two stream C3 D convolution neural network still has some problems,such as background dependence,the speed of optical flow calculation can not reach real-time and the model takes up a large space,which need to be further solved.The limitations of two stream C3 D network are analyzed and the improvement ideas are proposed.A service action recognition method based on two stream C3 D network is proposed.There are three parts for improvement:(1)the action recognition of convolutional neural network relies on scene information,so thesis improves the KNN(k-nearest neighbor)background subtraction method to reduce the calculation amount,and uses the KNN background subtraction met hod to preprocess the video before the video input network,that is,to eliminate most of the background information.(2)Although Gunnar farneback dense optical flow method is effective,it has slow calculation speed and some noise in the optical flow image.Therefore,TV-L1(total variation-l1)dense optical flow method is used to replace it.(3)Aiming at the problem of large space occupied by C3 D network model,3D residual network with faster calculation speed,smaller model and stronger anti over fitti ng ability is used instead..Finally,it mainly tests the performance of two stream 3D residual service action network.Firstly,the experimental design is carried out,and the parameter setting and process of the experiment are explained.Compared with the C3 D network,the 3D residual network can save 41% of the model space;Compared with Gunnar farneback method,TV-L1 method improves the speed of optical flow calculation by about 60%,which can meet the demand of service action recognition;The recognition accuracy of two stream 3D residual service action network reaches 92.85%.