A markerless platform for automatic assessment of gait based on Human Pose Estimation: A proof of concept

Instrumental and observational gait analysis are time-consuming and call for adequate operator training. The instrumental analysis calls for dedicated spaces, whilst observational analysis remains strongly operator dependant and subtle information may be missed, despite standardized tools can be adopted by clinicians (i.e., tables with specific features to be looked for in people walking). Markerless motion capture systems, such as pose estimation software based on convolutional neural networks and machine learning algorithms, may overcome the limitations of both the described approaches. This article presents a Markerless Automatic video-based platform for Gait Analysis (MaGA) that, given a predefined clinical tool and starting from the video recording of a person walking, computes the joint kinematics, identifies gait cycle sub-phases, perform a feature extraction and disentangles pathological from physiological walking via support vector machine algorithms, with linear and non-linear kernels. Results reported linear and non-linear models performances in classifying and predicting severity of gait alterations at hip and knee joints, according to the Ranchos Los Amigos scale. The Mean Absolute Error ranged between 0.04 and 0.70 for linear models and was generally lower than 0.50 for non-linear models. F1-scores are generally above 0.70, with a few exceptions. The analysis of three example cases demonstrate the effectiveness of the method in evaluating sagittal hip and knee kinematics, highlighting agreements and a few discrepancies with expert evaluations taken as ground truth. The proposed platform has the potential to be customized for the automatic assessment of individuals’ gait based on various clinical evaluation tools, thereby addressing the common limitations associated with them. Future plans include conducting comprehensive technical and clinical trials to assess the platform’s sensitivity under varying data collection conditions. Additionally, efforts will be made to establish a broader reference dataset, encompassing individuals with diverse disorders and varying levels of pathology severity. • Gait analysis is key for assessing motor impairments in clinical practice. • MaGA is a low-cost, automated tool for improving observational gait analysis (oGA). • MaGA applies machine learning to enhance accuracy and reduce human error in oGA. • Non-linear SVM models in MaGA improve gait classification and reduce overfitting. • Further research will aim to improve likelihood estimates of pathology in MaGA.