Research On Normalization Measurement Of Broadcast Operations Based On Graph Convolution Neural Network

With the rapid development of modern society,people pay more and more attention to their physical and mental health.Exercise is an important way to improve physical fitness.However,due to the lack of professional sports guidance,the public mostly imitates and masters the essentials of movements by watching videos of coaches’ standard movements when participating in sports.Therefore,this paper applies the computer vision technology in artificial intelligence to the field of sports,explores the evaluation of the standard degree of sports actions in an intelligent way,and brings professional sports guidance to the masses.Among sports,radio exercises are popular because they do not require a specific place or time.This paper takes broadcasting exercises as an example to conduct an in-depth study on the normative measurement of human movement.The specific research contents are as follows:(1)Established broadcast gymnastics dataset.In this paper,the broadcast gymnastics data set required for this experiment was established by collecting the standard action videos of coach broadcast exercises on the Internet and using cameras to collect students’ broadcast gymnastics action videos.(2)Estimation of 3D human body pose from RGB video in broadcasting operation.The 2D human pose was extracted from the video using HRNet,and a 3D human pose estimation network model combining graph convolution and LSTM was established in the part of elevating the two-dimensional pose to a three-dimensional pose.Among them,in view of the problem that the graph convolution reduces the feature extraction ability caused by the shared feature transformation matrix of nodes,this paper uses the initial two-dimensional position prior relationship between joint points to construct a mask matrix instead of an adjacency matrix,and introduces a multi-head attention mechanism to learn joint The point global context information makes the feature aggregation process of nodes more flexible.Aiming at the problem of poor time consistency of the 3D pose sequence obtained by single frame independent estimation,the 3D pose estimation result with spatio-temporal characteristics is further obtained through the LSTM network.The accuracy of the network designed in this paper is proved by experiments on the public data set,and the 3D human pose estimation of broadcasting RGB video is realized.(3)A method for measuring the similarity of three-dimensional posture sequences in broadcasting exercises is designed.Measure the degree of standardization of student actions by calculating the similarity between the three-dimensional pose sequence of student actions and the coach’s standard actions.Among them,a bidirectional longest common subsequence least squares distance algorithm was designed to address the issues of high dimensionality,time axis and amplitude distortion in 3D pose sequence data,and to consider the interference of redundant information on measurement results.Through this algorithm,the similarity between student actions and training standard actions was obtained,which serves as a basis for determining whether student actions are standardized.The experimental results show that the algorithm calculation results in this paper have high consistency with the scores given by professional coaches,and can effectively measure the degree of standardization of students’ broadcasting exercises,and provide intuitive guidance.