| In order to promote the intelligence of sports for all people,while reducing the space,time and economic costs required for sports,and to prevent poor sports results and muscle injuries caused by wrong sports postures,a single person sports posture evaluation system is designed and implemented.The system mainly uses computer vision technology to estimate a person’s posture and gesture in real time and make a simple effect evaluation.The whole system is based on Mediapipe framework and SVM classification algorithm for motion detection,motion counting,and motion analysis.The system uses Mediapipe,a mobile-friendly framework,which is divided into two parts: the pose estimation of the human body(Blaze-Pose),which is the basis of the system implementation,and is mainly used for the pose detection of the human body and the extraction of the corresponding key point information,and the SVM classification algorithm,which is mainly used to compare the newly acquired key point data information with the standard motion key point data preset by the system,and perform the motion classification.The SVM classification algorithm is mainly used to compare the newly acquired key point data with the pre-set standard action key point data,classify the action and determine whether the action is up to standard.The motion accuracy is also determined by adding an angle threshold to the SVM classification algorithm to determine whether the motion is standard or not.The system has a number of advantages over existing exercise applications,including real-time motion detection and counting,and off-screen non-contact gesture control at intervals.The system can be used to enhance anaerobic fitness training,including the degree of movement standardization,rational movement arrangement of movements,and movement considerations,which are several important aspects of motion monitoring.The paper mainly analyzes the current state of domestic and international research development of computer vision technology in human gesture estimation,including non-machine learning methods and machine learning methods,and briefly introduces them in the chronological order of model proposal.Then a brief introduction to the Medidapipe framework is given,describing the key point classification of the BlazePose and Blaze-Hands branch frameworks and the workflow of the solution,including lightweight detection methods and a new approach to localize targets for video streaming image frames.The basic principles and ideas of the basic SVM classification algorithm and the algorithm of the added angle thresholding method with arm movements as an example are also introduced,and the detection principles and ideas of the two classification models KNN and VGG16 for comparison tests are also briefly introduced.The human pose datasets FLIC,LSP,MPII and MSCOCO,JHMDB,which are often used in current deep learning,are also introduced.Then the proposed architecture of human pose and predicted action type of this system is introduced in detail,and the classification algorithm is compared with the accuracy of action classification on the dataset prepared by ourselves,and the experimental results are analyzed.Finally,the implementation process of this system is shown in detail and the work is summarized,and the ideas for further improvement of the human pose estimation model at a later stage.The proposed and implementation of this system applies the deep learning model to real-life fitness exercise,which is free from subjective human guidance,saves manpower,space resources,and economic consumption,and can provide better action supervision and guidance functions to reduce the risk problems of muscle strains,joint injuries,and poor exercise effects due to substandard movements when exercising alone. |
