Research Of Micro-Action Recognition Technology Based On Deep Learning

With the increasing maturity of computer science and technology,multimedia understanding has become one of the most important carriers of information.The image recognition has made amazing achievements,but video understanding still faces many challenges.Micro-action recognition is a task that can recognize body language which can reflect the current psychological state of people.The analysis of micro-action can be applied to lie detection,political psychology analysis and other fields.No matter in the scene of social interview or criminal trial,it plays an indispensable role in the identification and analysis of micro-action.The traditional way to recognize micro-action relies in the analysis of psychologists.In this thesis,the micro-action recognition is based on the action recognition.The methods of action recognition are similar to natural language processing.For the commercialization of action recognition,the algorithm needs to be real-time with high accuracy.In this thesis,the videos are decoded and sampled by method in TSN network.The video background is disorderly,which affects the recognition results.YOLO-V3 network with fast speed is used to detect the human body in the sampling frames.For the balance between computing speed and accuracy of action recognition network,a network of action recognition based on the fusion of 2D convolution and 3D convolution is proposed.The spatial feature maps of RGB image sequence are extracted by 2D convolution,and then a small amount of 3D convolution is used to fuse the spatial and temporal information of the feature sequence.The model not only has high accuracy,but also has fast calculation speed.The accuracy of micro action dataset is 94.8%.Compared with TSN network,the accuracy is increased by 6.6%.Compared with I3D network,the algorithm delay is about 1/3 of I3D network.It is verified that the algorithm in this thesis is more efficient.In order to expand the application of micro-action recognition,this thesis designs and implements a micro-action recognition system,which visualizes the web page of micro-action recognition.The system can decode and analyze the uploaded offline video,and recognize nine kinds of micro-action by using the proposed micro-action recognition algorithm.Finally,the system can return the results to the web page for display.