八段锦视频的人体姿态估计与相似度计算方法OA

Human pose estimation and motion assessment can effectively assist individuals in fitness activities.This paper proposes a computer vision-based human pose evaluation algorithm for Baduanjin exercise assessment.In the pose estimation stage,to address the high computational cost caused by standard convolutions in the YOLOv8-Pose model,a multi-scale ghost residual convolution(MSGRConv)module is designed.This module integrates Ghost convolution with depthwise separable convolution and introduces multi-scale convolution kernels.It reduces the number of model parame-ters by 3.6%while improving model accuracy(with an increase of 1.2 percentage points in mAP@50 and 1.6 percentage points in mAP@50-95).In addition,the PIoUv2 loss function is introduced to replace the original CIoU,resulting in more efficient model training.In the motion evaluation stage,a scoring method combining global and local similarity is pro-posed.Global similarity is computed using the FastDTW algorithm,which measures the temporal alignment of pose sequences based on Euclidean distance.Local similarity is evaluated using 16 weighted limb vectors constructed from COCO dataset keypoints,and key frame similarity is analyzed accordingly to enhance pose evaluation accuracy.Experi-mental results show that the proposed scoring method is highly consistent with expert ratings,can effectively distinguish pose similarities and differences,and exhibits strong robustness and discriminative ability.This highlights its potential for application in the intelligent evaluation of traditional fitness exercises.