3D Markerless clinical gait analysis based on RGB and Depth sensing technology for children affected by Cerebral Palsy and Clubfoot

Marker-based stereophotogrammetry (MB) is the gold standard for gait analysis, but its clinical use is limited due to high costs, long set-up times, and patient discomfort. Recent advances in depth sensing and machine learning have made markerless (ML) gait analysis a promising alternative for clinical applications, especially when ease of setup is essential. The development of RGB-Depth technology further enhances ML capabilities by integrating color and depth data, thus allowing for a color point cloud reconstruction. In the literature, various single-camera ML algorithms have been proposed, including deterministic and deep learning-based approaches. However, these algorithms rely on 2D video analysis, requiring manual identification of anatomical landmarks and failing to capture out-of-plane movement. Recent advances in computer vision have led to 3D statistical models, such as the Skinned Multi Person Linear (SMPL) model, which realistically represent diverse body shapes and poses. However, its use in clinical gait analysis remains limited. This thesis proposes an original ML protocol based on a single RGB-D camera and the 3D SMPL model. Five cerebral palsy and nine clubfoot patients performed three self-selected gait trials per side, and a static posture of each participant was acquired from three different camera views (frontal, posterior, and sagittal). A SMPL consisting of foot, shank, thigh, and pelvis interconnected by ankle, knee, and hip joints was calibrated to each participant’s static posture. Then, the SMPL was aligned to each dynamic frame of the gait cycle using the articulated iterative closest point algorithm to estimate 3D joint kinematics and extract seven clinical gait features. Validation was performed against a 3D MB clinical gait analysis protocol. Assuming movement repeatability, gait trials were recorded separately to avoid IR sensor interference between systems. The accuracy of the ML protocol was evaluated using Mean Absolute Difference (MAD) with respect to the MB system, while reliability and variability of both ML and MB protocols were assessed via Intraclass Correlation Coefficient (ICC) and Gait Variability Standard Deviation (GVSD). MAD values were: for the knee kinematics, 3.5° during stance, 2.8° in swing; for the ankle kinematics, 4.1° in stance, 4.5° in swing; for the hip kinematics, 5.4° in stance. ICC for ML and MB were: for the knee kinematics, 0.84 vs 0.89 in stance, 0.82 vs 0.96 in swing; for the ankle kinematics, 0.91 vs 0.86 in stance, 0.87 vs 0.90 in swing; for the hip kinematics, 0.90 vs 0.95 in stance. Mean GVSD values for ML and MB were: 4.6° vs 3.6° for the knee, 3.4° vs 1.9° for the ankle, 3.2° vs 2.3° for the hip. Residual MAD values are mainly due to asynchronous acquisitions and different anatomical axis definitions between ML and MB protocols. ICC values for the two protocols are comparable, with the largest discrepancy observed in the knee kinematics during the swing phase due to the depth sensor's limited ability to reconstruct depth values at high speed. Mean GVSD values for ML and MB protocols are comparable, with an average difference of 1°, proving ML protocol to have similar variability to MB. In conclusion, the proposed ML protocol provides 3D joint kinematics by leveraging the SMPL model to enhance automation and overcome 2D analysis limitations. The proposed ML protocol shows strong reliability (ICC > 0.8), making it a promising solution for future clinical applications.