Feasibility of markerless motion capture in clinical gait analysis in children with cerebral palsy

Abstract

The main purpose of this study is to explore the feasibility of deep learning based markerless motion capture in clinical settings. This study compares the markerless motion capture (Blazepose) to the gold standard marker-based system (Vicon) by assessing the hip, knee, and ankle joint flexion angles of cerebral palsy (CP) patients and their typically developed (TD) peers. The participant group included six CP patients and six TD individuals. The participants walked on an 8-meter gait path at a self-selected pace. 11 Vicon cameras (200 Hz) and 3 GoPro (60 Hz) cameras were used for the marker-based and markerless system setup. Both systems were synchronized and recorded simultaneously. The keypoint trajectories from Blazepose were obtained by feeding the images collected with GoPros as an input to the algorithm. Further analysis included calibration, 3D reconstruction, and data filtering in Matlab. Skeletal modeling and joint angle calculations were conducted in OpenSim for both systems to eliminate the methodological difference. SPM1D Matlab package was used for statistical analysis. Significant differences were observed in ankle and hip joint angles between the Blazepose and Vicon systems at specific gait cycle phases in both the CP and TD groups. The ankle angle showed significant differences in the CP group at 0.7–1.3% (p<0.016) of the gait cycle and in the TD group at 38–46% (p<0.016). For hip flexion, significant differences in the CP group were noted at 0.81–1.81% (p<0.016), 13–40% (p<0.016), and 89–93% (p<0.016) of the gait cycle, while in the TD group, a significant difference was observed at 75–84% of the gait cycle (p<0.016). No significant difference was observed for the knee angle in both groups. The results of this study highlight the potential use and certain limitations of markerless motion capture systems like Blazepose in clinical settings. While showing promise in certain aspects of joint angle tracking, the study emphasizes the need for more refined datasets and advanced algorithms to enhance the accuracy and reliability of such systems, especially for clinical applications involving CP patients.