基于深度学习的健身动作识别系统

Fitness Motion Recognition System Based on Deep Learning

随着全民健身热潮的兴起,越来越多的人积极参加健身锻炼,但由于缺乏科学的运动指导,使健身难以取得相应的效果。据我们所知,没有产品可以自动分析健身运动并提供指导。针对这个现象,设计了一个基于深度学习的健身动作识别系统,该系统由三个部分组成:提取运动边界、人体姿态估计和动作识别/评分。首先使用边界敏感网络来生成包含动作实例的时序动作提名,这样可以确定视频中的动作边界,提取视频中做健身运动的部分,忽略无关背景的影响;然后应用AlphaPose来估计人体姿态;最后通过人体姿态构造图形,并利用图卷积网络来识别健身动作。还通过比较用户姿态和教练姿态之间的相似性来获得用户健身动作的得分。大量的实验表明,设计的系统在性能上可以满足实际应用的要求。

With the rise of national fitness,more people voluntarily participate in fitness exercises.However,due to the lack of scientific exercise guidance,fitness is difficult to obtain corresponding results.In response to this phenomenon,this paper designs a fitness motion recognition system based on deep learning.The system consists of three parts:motion boundary extraction,human pose estimation and motion recognition/scoring.This paper firsts use Boundary Sensitive Network(BSN) to generate temporal motion proposals,which contains motion instances.In this way,can determine the motion boundary in the video and extract the part of doing fitness in the video,ignoring the influence of irrelevant background,then apply AlphaPose to estimate the human pose.Finally,this paper construct the graph by human pose and utilize graph convolution network to recognize fitness motion.This paper compares the similarities between the user’s pose and coach’s pose to obtain a motion score.Extensive experiments demonstrate that the performance of our system designed in this paper can meet the requirements of practical application.