Lightweight sit-ups recognition and counting method based on OpenPose

The study of accurate, fast and automatic sit-ups, mainly depends on human bone point recognition and supine sitting judgment algorithm. In terms of human bone point recognition, the existing models depend on the main network with a large amount of parameter computing, which cannot be adapted to computers with poor performance and some embedded devices. In response to the above issues, on the basis of the OpenPose model, this paper adds a multi -scale fusion module and spatial attention module to the feature extraction network, using the method of structured knowledge transfer (SKT) The identification module proposes a lightweight SKT-OpenPose network. The network parameters are only 6.1% of the original OpenPose model, which greatly shortens the operating time and reduces the requirements for hardware. Regarding the judgment algorithm of sit-ups, the existing method pays attention to whether the angle of the pelvis meets the conditions of sitting up and lying down, and needs to shoot testers from a specific perspective. The angle limit at the knee joint is used to determine the standardability of the current action; add an angle judgment mechanism to judge the motion of sit-ups of different angles. It has been proved by experiments that the lightweight sitting recognition and counting method proposed in this article have a good effect in identifying the sitting action recognition.