Multiview Human Pose Estimation and Corrective Feedback

Exercise posture is a basic essential activity in people's spare time. Proper and correct exercise pose can not only help people to get a healthy body, but also help to reduce pressure and relax their mood. So, for practicing exercise a trainer is important who can guide and monitor the perfectness of different exercise poses. However, the wrong way of exercising not only makes the exercise fall short of expectations, but also causes muscle damage. Therefore, we have developed a set of human exercise posture analysis and guidance software to analyze the user's exercise posture through the video captured by the camera. This software chooses OpenPose as the basic pose estimation network, and two methods are proposed to improve the model effect. The software firstly conducts the human body posture detection, obtains the coordinate information of key points, then carries on the exercise posture analysis, and gives relevant suggestions according to the definition of the correct exercise form. Integrating computer vision techniques, the proposed system analyses the user’s human pose then based on the main knowledge of exercise, the user is guided to correct the pose. Precise recognition of exercise pose is a difficult task because of high computation and lack of availability of dataset. For the accurate detection of the exercise pose, different feature extraction and preprocessing methods are applied to the dataset which results in 97.4% accuracy just by using machine learning algorithms.