Home // International Journal On Advances in Life Sciences, volume 15, numbers 3 and 4, 2023 // View article
Authors:
Riku Nishimoto
Kyoko Shibata
Keywords: Deep learning, Single camera, Estimation, Musculoskeletal model simulation, Lumbar load
Abstract:
To prevent lumbago, it is effective to have a system that enables people to improve their habitual bad posture. Therefore, we will develop a method of estimating body load without user burden for constant observation of posture. Hence, this study proposes the use of a web camera, which everyone has and can acquire images on a daily basis without any burden, as a non-contact sensing method, and the use of deep learning as a means of estimating body load from web images. Deep learning models are created by deriving body load values using musculoskeletal analysis based on skeletal position coordinates extracted from posture images and labeling the images with these as true values. Thus, if a pre-trained deep learning model is created in advance, body load can be estimated from images alone, without the use of specialized software or cloud communication. If it is possible to easily visualize one's own body load in daily life, the system can be developed to provide feedback on posture evaluation and improvement plans based on the estimated body load. We consider that this will further increase the users' awareness of improvement and lead to the maintenance and promotion of health. In this paper, as the first step, a deep learning model is created for a stationary standing forward bending posture, and the accuracy of the lumbar load estimation by the deep learning model is evaluated. The results of individual learning using untrained data allowed us to estimate the lumbar load with high accuracy. Hence, the possibility of applying the proposed method to certain individuals is indicated. The other is, the results of ensemble learning confirmed models with high and low accuracy. Hence, the deep learning models that estimated untrained participants showed large variations in accuracy and insufficient generalization performance. Discussion of the results confirms that data bias is a contributing factor to the accuracy loss and indicates the possibility of obtaining generalization performance by improving data bias.
Pages: 62 to 71
Copyright: Copyright (c) to authors, 2023. Used with permission.
Publication date: December 30, 2023
Published in: journal
ISSN: 1942-2660