Human Action Recognition Based On RNN Framework And Inference Network

Human action recognition has always been a very important subject in the computer field.It has extremely important research significance for human-machine interaction,video surveillance,medical assistance,and analysis of abnormal human behavior.In recent years,the emergence of various short video software has caused the video data on the Internet to explode.The traditional methods have been unable to meet the demand for data growth,which provides an opportunity for the development of deep learning.However,human body movements are more complicated.The use of deep learning methods to identify human movements still has certain challenges.Based on the deep learning method,the main work of thesis is as follows:Firstly,aiming at the influence of occlusion and other external factors on human motion recognition in RGB video,and the problem that the recognition accuracy needs to be improved,a two-stream independent recurrent neural network human action recognition algorithm is proposed.In terms of feature extraction,the temporal network uses hierarchical Ind RNN to extract the feature of the 3D skeleton coordinate information.The spatial network uses deep Ind RNN to extract the feature of the spatial position relationship of the skeleton.The spatial structure of the skeleton adopts the graph traversal.The method uses a weighted summation method for the feature fusion of the spatial network and the time network,and finally classifies the actions with Softmax.Secondly,context information is of great significance for improving the accuracy of visual recognition.In the human action recognition,if the influence of the interactive object on the human body motion is not considered,some action misjudgment often occurs.Aiming at this problem,a human action recognition algorithm based on inference network is proposed.In order to improve the recognition accuracy of human motion,the algorithm adds LSTM to the Faster RCNN and combines the object information related to human motion in the scene.Firstly,the feature information with human as the main area and the object as additional area is extracted through Faster RCNN,and then it is input into LSTM for border regression and action classification.By combining Faster RCNN and LSTM,rich spatial and temporal features are obtained.so as to obtain more accurate action classification.