Validation and Use of Markerless Captured Joint Kinematics to Drive a MSK Model For Knee Force Estmation

The present study validates joint kinematics and kinetics captured by a deep learning-based markerless motion capture system, Theia3D, with a standard marker-based system, Qualisys. Furthermore, this work investigates the use of the markerless capture system as an alternative to traditional marker-based methods in musculoskeletal (MSK) modelling for Knee Joint Reaction Forces (KJRF) estimations. Because no methodology for applying the Theia3D capture data to drive MSK is defined in literature, a pipeline was developed to input the markerless data to successfully run MSK simulations and estimate KJRF. Ten healthy subjects were recruited to perform ten walking trials and were instrumented with reflective markers. Data was simultaneously recorded using eight video cameras for the markerless capture system, and twenty-four cameras of an optoelectronic system. Both sets of processed data were exported for kinematics and kinetics analysis in Visual3D. Joint angles showed similar patterns for flexion/extension at all joints and adduction/abduction for both hip and knee. Highest differences were found for internal/external rotation angles, at the hip (RMSE = 6.0°), knee (RMSE = 10.7°) and ankle (RMSE = 12.0°). Joint moments’ RMSE was < 0.1 Nm/kg for all joints, except for hip flexion/extension (0.14 Nm/kg) and ankle inversion/eversion (0.13 Nm/kg). The KJRF predicted by the markerless and marker-based MSK models showed highest similarities in the medio-lateral direction, with RMSE = 0.09 [N/N]. In the antero/posterior direction the average RMSE was 0.2 [N/N], increasing in correspondence to the peak KJRF, whose value was higher for the marker-basd estimations (1.11 [N/N]) compared to the markerless (0.82 [N/N]) ones. Greater RMSE was measured in the proximo/distal direction (0.6 [N/N]), with greater peak value predicted by the markerless model compared to the marker-based one, amounting to 4.94 and 4.45 [N/N] respectively. These results show comparability of lower limbs’ kinematics and kinetics estimations between capture systems, and hold promise for the applicability of the markerless capture data in the estimation of joint reaction forces using MSK models. However, questions remain on the accuracy of frontal and transverse plane kinematics, as well as limitations related to the proposed markerless-driven MSK model.