AI-based Injury Prevention Assistance System for Automated Motion Analysis of the Lower Extremities to Prevent Injuries Caused by Improper Loading – A Pilot Study
Sabine Bauer, Ivanna Kramer
Introduction
The knee joint is prone to injuries, which often occur after jumping landings, changes of direction, or abrupt braking. Improper loading or habitual improper movements can also cause damage. Prevention programs can reduce the risk of injury, such as reducing anterior cruciate ligament ruptures by 51 %. In this study, we investigated the potential of marker less measurement of human posture from RGB images using different AI-based pose models.
Methods
We analyzed the lower extremity kinematics of 10 subjects (4 male, 6 female) performing two-legged squats (3 repetitions, 3 sets). The movements were recorded with Intel RealSense D435i cameras (848x480 px, 30 fps) placed in front and to the subjects’ right. Three experts annotated the knee, hip, and ankle in each image. We also used real-time AI pose models (OpenPose, YOLO pose, MediaPipe pose) to detect key points and automatically calculate knee-hip-ankle angles (figure 1).
Fig. 1: The knee-hip-ankle is calculated using the kinematic data from Open Pose (purple), YOLO (blue) and Media Pipe (red) in frontal ((a)-(c)) and lateral view ((d)-(e)). Ground truth is marked green.
Results
We analyzed 3,630 images to compare angles calculated from AI-detected key points with expert-annotated ground truth data. The mean absolute error (MAE) for OpenPose was 5.0° ± 3.68° in lateral images 8.65° ± 8.36° and in frontal pictures for the right leg, and 8.13° ± 8.06° for the left leg. YOLO pose showed an MAE of 5.08° ± 3.34°in lateral images and 6.44° ± 6.39° in frontal pictures for the right leg, and 6.31° ± 6.48° for the left leg. MediaPipe had an MAE of 5.49° ± 4.08° in lateral images and 6.43° ± 7.24° in frontal pictures for the right leg, and 7.13° ± .23° for the left leg.
Discussion and Conclusion
The analysis showed that the MAE values for each model differed across different views and legs. OpenPose had a higher MAE in the frontal view, especially for the right leg. YOLO pose performed consistently between lateral and frontal views. MediaPipe had the lowest MAE in the lateral view but more variability in the frontal view. These results suggest that the view and specific leg significantly influence model performance. Additionally, the pose models exhibited higher MAE in images of participants wearing loose clothing.
Key Message
The results show that lower joint angles can be reliably calculated using landmarks from pose models, supporting their use in real-time squat motion analysis for injury prevention. We suggest using a combination of pose models, such as OpenPose or YOLO pose, for the lateral images and YOLO for the frontal images.
Authors
Dr. rer. nat Sabine Bauer
Ivanna Kramer, MSc
University of Koblenz
Institute for Medical Technology and Information Processing
Postbox 20 16 02 | D-56016 Koblenz
E-Mail: bauer@uni-koblenz.de