Research And Implementation Of Human Activity Recognition System Based On OpenPose

Crossing the "semantic gap" and letting the machine complete the pixel-to-semantic mapping is the fundamental problem to be studied in computer vision.The research on Human Action Recognition(HAR)has gradually become a hot topic in recent years and has attracted much attention due to its broad application prospects.Aiming at the shortcomings of current action recognition methods,this thesis proposes a human action recognition method and real-time framework system based on OpenPose.Taking advantage of the small amount of skeleton data and the ability to retain the key information of actions,this thesis studies the representation method of the action model and how to effectively improve the accuracy of action recognition in video streams.The main work has been summarized as follows:(1)Design a method of construction a human action model.The model includes data regularization and feature design methods.Aiming at the problem of recognition accuracy decrease due to the difference distances between the human bodies and the camera and the difference shapes of bodies,this thesis proposes a method for normalizing bone joint point data.At the same time,in order to fully obtain the human action characteristic information,a method for fully extracting the joint angles distance and position regularization is proposed.We verify the effectiveness of the proposed method through experiments.(2)Propose an action segmentation algorithm.To solve the low accuracy and high computational complexity of current multi-frame video behavior recognition,this thesis designs an Action Segmentation Algorithm(ASA).A complete action usually consists of initial action,execution process,and end action.The essence of ASA is to cut the spatio-temporal information matrix data of the multi-frame video stream to obtain the complete action spatio-temporal information matrix and remove redundant video frame data,thereby improving the recognition accuracy.(3)Development of behavior recognition system based on OpenPose.The system is developed based on the above-mentioned research content and methods,and consists of five modules: data acquisition,action segmentation,action model construction,Opencv drawing,and neural network construction and recognition.At present,it is possible to realize the determination of human body motion tags based on data sets.This thesis mainly uses three public data sets of KTH,HMDB51 and Weizmann as the test data.The test results show that the human motion model construction method can fully extract the motion feature information;The action segmentation algorithm can effectively obtain the spatio-temporal information matrix of the action in the video stream,and further improve the recognition accuracy.The research results of this thesis can be applied to practical fields such as smart security,traffic violations,games.