The purpose of this research was to evaluate the algorithm DeepLabCut (DLC) against a 3D motion capture system (Vicon Motion Systems Ltd) in the analysis of lumbar and elbow flexion-extension movements. Data were acquired concurrently and tracked using DLC and Vicon. A novel DLC model was trained using video data derived from a subset of participants (training group). Accuracy and precision were assessed using data derived from the training group as well as in a new set of participants (testing group). Two-way analysis of variance were used to detect significant differences between the training and testing sets, capture methods (Vicon vs DLC), as well as potential higher order interaction effect between these independent variables in the estimation of flexion-extension angles and variability. No significant differences were observed in any planar angles, nor were any higher order interactions observed between each motion capture modality with the training versus testing data sets. Bland-Altman plots were used to depict the mean bias and level of agreement between DLC and Vicon for both training and testing data sets. This research suggests that DLC-derived planar kinematics of both the elbow and lumbar spine are of acceptable accuracy and precision when compared with conventional laboratory gold standards (Vicon).
Keywords: accuracy; machine learning; motion capture; precision; range of motion.