Home // International Journal On Advances in Life Sciences, volume 13, numbers 1 and 2, 2021 // View article
Authors:
Yuta Ono
Oky Dicky Ardiansyah Prima
Kazuki Hosogoe
Keywords: 3D human pose estimation, partial body joint, RGB-D camera, computer vision, sitting posture
Abstract:
Human pose estimation has been used to perform human motion analysis in widespread applications. Threedimensional(3D) human pose estimation from single image has attracted much attention because of its ease of measurement. Methods of this approach have become more accurate with the introduction of deep neural networks. Most of these methods are trained to estimate the body joints of the whole human body. However, when a part of the body joints is obscured by the presence of other objects or the camera position and angle, the estimation accuracy of the overall body joints may be degraded. In this study, we attempt to experimentally construct a 3D human pose estimation model for partial body joints to accurately estimate the pose of a partially human body that can be visibly measured. To evaluate the performance of the proposed model, we construct a neural network model that estimates only the 3D position of the visible upper body joints, assuming that only those joints are visible. Our evaluations showed that the partial body joint model was more accurate in estimating the posture from frontal human images. However, there was no significant difference in the accuracy between our model and previous pose estimation model when the posture was estimated from images of people taken from extreme angles. Finally, we attempt to extend our model to a system for detecting deterioration in sitting postures to verify the effectiveness of the model.
Pages: 114 to 123
Copyright: Copyright (c) to authors, 2021. Used with permission.
Publication date: December 31, 2021
Published in: journal
ISSN: 1942-2660