ScholarWorks@경상국립대학교

초록

Musculoskeletal injuries require continuous rehabilitation even after surgery. Recently, pose estimation models such as OpenPose have been used to extract joint-based postural information. However, these models have limitations in accurately detecting detailed joint positions. To address this, we developed a lower limb joint detection model using SSD (Single Shot MultiBox Detector) and various YOLO (You Only Look Once) nano and small models, optimized for real-time object detection. The model was designed detect key joints during three rehabilitation exercises: ankle pump, heel slide, and hip abduction. Training results showed that YOLO models accurately predicted joint positions, with nano models demonstrating high accuracy despite low computational demand. A rehabilitation assistance program was constructed by integrating the models trained for each exercise. The system was evaluated using real rehabilitation videos, and predicted joint angles were compared with actual angles. The YOLO small model recorded standard deviations of 1.61° and 0.7° for specific movements. In the ankle pump exercise, which involves detecting small joints such as the toe, the YOLO v5 small model achieved a 98.7% success rate, with a standard deviation of 5.19° and a mean error rate of 4.19%. The proposed program enables non-invasive, low-cost monitoring of rehabilitation performance without the need for wearable sensors or expensive equipment. This study demonstrates the feasibility of using lightweight, real-time object detection models for rehabilitation support and offers potential for broader application in medical and home settings.

키워드