A markerless system for gait analysis based on OpenPose library

The paper reports the performance of a low-cost markerless system for 3D human motion detection and tracking, consisting of the open-source library OpenPose, two webcams and a linear triangulation algorithm. OpenPose is able to identify anatomical landmarks with a commercial webcam, using Convolutional Neural Networks trained on data obtained from monocular images. When images from at least two different points of view are processed by OpenPose, 3D kinematic and spatiotemporal data of human gait can be also computed and assessed. Despite its potential, the accuracy of such a system in the estimation of kinematic parameters of human gait is currently unknown. With the aim to estimate OpenPose accuracy in 3D lower limb joint angle measurement during gait, two synchronized videos of a healthy subject were acquired, with two webcams, in a walking session on a treadmill at comfortable speed. 2-dimensional joint centers coordinates were assessed by OpenPose, and computed in 3D by triangulation algorithm. The resulting angular kinematics was, then, compared with inertial sensors outputs. Results showed that the system was generally able to track lower limbs motion, producing angular traces representative of normal gait similar to the ones computed by IMUs. However, OpenPose approach showed inaccuracy, mostly in the computation of maxima and minima joint angles, reaching error values up to 9.9°.