Research on Lightweight Network of Human Posture Estimation for Physical Training
Yaowei Sun, Gang Li, Tongzhou Zhao · 2022
Human posture estimate based on openpose is lightweight and allows for real-time analysis on mobile devices. Contrapose the problem that the real-time of physical training and detection process based on computer vision needs to be improved. Firstly, the MobileNets network was used to achieve model lightweight. MobileNets has the advantage of reducing more parameters and faster speed. Then the image was input, and the coordinate information of human skeleton key points was obtained in real time through OpenPose network, and each key point was connected through PAFs. Finally, the joint Angle value was calculated by the selected coordinates of six bone key points, and the obtained joint Angle value was compared with the standard value to judge whether the posture detected in human physical training is normal. Through experimental verification, the lightweight network can reduce the computational complexity by 8-9 times, improve the accuracy by 3%, and realize the detection function of human body physical training posture.