Deep squat training plays a crucial role in the treatment of elderly patients. In this paper, we introduce the target detection algorithm YOLOv5 into MediaPipe, a human pose estimation framework, based on machine vision detection, and propose a method for detecting deep squatting movements. By modifying the YOLOv5 feature extraction network, the human target position is accurately detected, and MediaPipe obtains human skeletal information to mathematically model the deep squat pose and find out the trunk angle, hip angle, and knee angle. The experimental results show that the method can effectively detect deep squatting movements, eliminate false detection rates and improve the robustness of the algorithm in complex environments with an accuracy rate of over 96%.