Action recognition and quality assessment of postpartum rehabilitation training are of great significance to improve the efficiency of rehabilitation guidance and reduce the risk of wrong training. The traditional postpartum rehabilitation evaluation mainly relies on the observation of therapists and the subjective feedback of puerpera, which is susceptible to differences in experience, training environment and feedback delay. To solve this problem, this paper proposes an intelligent evaluation system for postpartum rehabilitation training based on action recognition. Based on posture estimation, the system extracts key points such as shoulder, hip, knee, ankle and other bones, constructs joint Angle, pelvic stability, action amplitude and left-right symmetry features, and completes the recognition and quality score of bridge training, pelvic forward and backward tilt, lateral lying opening and closing, kneeling leg raising and supine leg raising through the time sequence action recognition model. At the same time, a lightweight action recognition network is constructed to support real-time feedback from the end side. Experimental results show that the average action recognition accuracy of the system is 93.3%, the quality assessment accuracy is 91.0%, and the real-time frame rate of the INT8 quantization model reaches 30.7 frame·s⁻¹, which verifies the effectiveness and application feasibility of the system in the intelligent evaluation of postpartum rehabilitation training.
