Research On Human Motion Pose Matching Based On Deep Learning

Under the background of the whole society vigorously carrying out national fitness,people pay more attention to their own health conditions,and more and more people participate in fitness sports.Exercise and fitness can enhance people’s body and help them maintain a good life.With the increasing number of people participating in sports,the shortage of professional sports coaches is a prominent problem.If sports personnel carry out sports activities with an irregular movement posture for a long time,not only can not play the role of strengthening the body but will make the body suffer certain damage.In order to help people correct irregular motion posture,a human motion posture matching system combining human target detection and human posture estimation is proposed.The main research contents are as follows:(1)The human object detection model based on YOLOv7 network was established to locate the specific position of a single human body in the input image,so as to help realize the posture estimation of multiple people.The attitude matching system is mainly human oriented,and only keeps the human tag category of YOLOv7 target detection model,which has fewer parameters compared with the original multi-target detection model.(2)Establish the human posture estimation network based on the improved HRNet to obtain the key points of human bones.The improved HRNet network is based on HRNet_W32,and the fourth stage redundancy is reduced by adjusting the network structure.In order to make up for the network’s insufficient ability to acquire feature information of important key points,CBAM attention mechanism is embedded in the network to improve the network’s ability to extract important key points from channel dimension and spatial dimension respectively.Depth-separable convolution is used to replace common convolution in the network Bottleneck module and BasicBlock,which bottleneck makes the model overall lightweight.(3)Build a human motion posture matching system based on dynamic time warping algorithm DTW,achieve motion posture similarity score,and extract video frames with low similarity.By constructing the skeleton diagram of human body,the Angle between the limb Angle and the limb orientation vector was used as the similarity judgment factor.The internal structured path search method of DTW algorithm is redefined,and the obtained distance is converted into a similarity scoring mechanism.Visual presentation of the whole system,more intuitive and convenient for motion detection personnel to provide detection.The experimental results show that the improved HRNet human pose estimation network can achieve 73.8%AP value in COCO2017 data set and 90.1%PCKh@0.5 value in MPII data set on the basis of model parameter only 13.2M.The improved network can ensure high accuracy while lightweight.The motion posture matching system based on DTW can effectively help the detection personnel to correct the irregular motion posture.