A deep-learning pipeline to extract clinical gait metrics from single-camera recordings.

Abstract

Single camera markerless motion capture is a low-cost, accessible motion capture technology that could be used to improve clinical decision making. However, current models lack the spatial and anatomical fidelity required for research-grade analysis, which prevents accurate modelling of 3D human gait dynamics. The purpose of the current study was threefold, 1) to develop a means of predicting 3D gait kinematics using consumer-grade video inputs, 2) to assess the accuracy of the method across a variety of ecologically relevant perturbations such as camera location, body size, and clothing characteristics, and 3) to assess the utility of the 3D gait kinematics to score walking gait quality. Treadmill walking data were recorded from four different camera angles, and a diverse range of participant characteristics and clothing conditions. A single LSTM deep neural network was trained from each of the four camera angles to predict the 3D anatomical landmarks recorded from a research-grade motion capture system consisting of 17 body-worn inertial measurement units. The model trained on data from the camera positioned on the front right (i.e., 2o'clock) of the treadmill demonstrated the best-performing model (RMSE of 2.97 cm and 4.24 degrees). Significant effects of clothing characteristics, gait strategy, and anthropometrics on model performance were observed. Acceptable agreement was observed between the gait profile scores calculated from the LSTM and research-grade motion capture system (RMSE = 2.07, ICC = 0.65). These results demonstrate the potential for single-camera markerless motion capture to produce clinical grade data in a healthy population, thereby improving the accessibility of clinical gait assessments.