基于TM-Net网络估计的三维人体姿态运动监测算法

TM-Net Based Estimation Algorithm for 3D Human Pose and Motion Monitoring

针对三维人体姿态估计的便捷性与准确性提升需求,提出一种基于TM-Net网络估计算法。该算法以MediaPipe为中心,融合帧率计算、动作检测、动作计数和真实坐标解析等多功能模块,实现对人体运动的精准检测与计数。针对公共数据集LSP(Leeds Sports Pose)和自建校园健身房运动数据集使用关键点的正确性概率(Probability of Correct Keypoint,PCK)、关节位置误差平均值(Mean Per Joint Position Error,MPJPE)和普罗克鲁斯对齐后的平均关节位置误差(Procrustes-Aligned Mean Per Joint Position Error,PA-MPJPE)等指标对该算法进行评估,并与目前先进的TP-3D网络估计算法进行对比。结果表明,TM-Net具有更高的准确率。此外,以开合跳为例进行消融实验,结果表明,TM-Net具有更强的泛化能力,能适应不同个体及拍摄角度的变化,满足了运动监测的实际需求。

To meet the demand for increasing convenience and accuracy in the estimation of 3D human posture and motion,an algorithm based on TM-Net is proposed.Centered on MediaPipe,the algorithm integrates multi-functional modules such as frame rate calculation,motion detection,action counting,and real coordinate analysis to achieve accurate detection and counting of human motion.The algorithm is evaluated on the public dataset(Leeds Sports Pose,LSP)and the sports dataset of the campus gym using the indexes such as probability of correct key points(PCK),mean value of joint position error(MPJPE)and mean value of joint position error after Procluss alignment(PA-MPJPE).Compared with the advanced TP-3D network estimation algorithm,TM-Net one is of higher accuracy.In addition,the ablation experiments are conducted taking jumping jack as an example.The result shows TM-Net algorithm has a stronger generalization ability to deal with changes of different individuals and shooting angles,meeting the actual needs of motion monitoring.