Comparison of 2D and 3D Gait Kinematics in Gender Classification Using Principal Component Analysis and Convolutional Neural Networks

Abstract

Human gait analysis is of great importance and can be used in the prevention and treatment of motion abnormalities. Spatiotemporal data of gait captured by high-frequency cameras is studied in 3 anatomical planes: sagittal, frontal, and horizontal. The necessity of using several cameras in motion capture technology for capturing 3D data limits its application for clinical purposes. This study evaluated the possibility of using Principal Component Analysis (PCA) as a feature selection technique to find out which anatomical plane provides the most useful information in gait analysis. For this purpose, 3-dimensional marker trajectories of 14 healthy subjects walking on a treadmill with three different speeds were captured. Then, PCA was applied to each gait cycle data to find out variables with the most variation. Afterwards, to evaluate the accuracy and reliability of PCA results, a convolutional neural network (CNN) was used. The highest eigenvalues obtained from PCA indicated that Y-axis (forward direction) had the most variance. Based on the mentioned result, 3 different datasets were prepared as CNN inputs for gender classification: 1) marker trajectories in 3D space, 2) marker trajectories in the X-Y plane (horizontal), 3) marker trajectories in the Y-Z plane (sagittal), The classification accuracy obtained from all CNN models were higher than 95%, which confirmed the significant role of the 2D plane for some useful applications such as gender classification.