Action Recognition Algorithm Based On Skeleton And Pose Information

Action recognition technology integrates the research results of many subjects such as computer vision,deep learning,pattern recognition,and it has been widely used in the fields of video search,auxiliary medical care,video surveillance and so on.The skeleton data containing the key information of human action not only can describe the human movement posture in three-dimensional space,but its rigid connection structure is also robust to various external interference factors.Therefore,the action recognition technology based on skeleton and pose information has also attracted the attention of academia and industry at home and abroad,and is one of the research hotspots in the field of computer video analysis.This thesis focuses on the research of human skeleton information extraction and action recognition based on skeleton.The main works and innovations are as follows:(1)In this thesis,skeleton extraction and pose estimation algorithms were studied.Firstly,the four main skeleton extraction and pose estimation algorithms are compared and analyzed on the data set.Secondly,in order to improve the accuracy of skeleton pose information,this thesis proposes a denoising method based on data association and skeleton energy model based on the research of Open Pose,which based on convolutional neural network and PAFs features.This method associates the skeleton data with the human target,and counts the cumulative energy of skeleton points in time series.It reduces the interference of irrelevant background and environmental noise.Finally,this thesis proves the effectiveness of the algorithm,after qualitative experiment and quantitative analysis.(2)This thesis studies the action recognition algorithm based on the space-time graph convolutional neural network.Above all,the space-time graph convolutional neural network was tested and evaluated on the cross-view data by combining the Kinect data set and the self-built data set.Then,according to the recognition accuracy of the algorithm for different action types,a voting classification strategy combining skeleton information and attention model is designed to output the action recognition result.Experiments show that this method can improve the recognition accuracy of long-term action.(3)Based on the above research,this thesis designs and establishes a skeleton extraction and action recognition system for basketball.The system includes three modules: video acquisition module,skeleton extraction module and action recognition module.The video collection module sorts out data according to 6 categories of basketball actions,and establishes a corresponding human skeleton database.The skeleton extraction module completes real-time detection of skeleton and posture information of sports on the court.The action recognition module processes the function of obtaining the sports action category of a specific player by the action recognition algorithm based on the space-time graph convolutional neural network.The system has good application prospects in real scenes such as sports video live broadcasting and intelligent teaching.