2025 Volume 76 Issue 2 Pages 54-63
Ergonomic assessment tools are used to evaluate working postures in workplaces. This study proposes a method to easily evaluate working postures based on the Ovako Working posture Analysis System (OWAS), which is one type of ergonomic assessment tool. Specifically, 2D coordinates for a worker's joints were manually obtained from an image, and these were input into a machine learning model created in advance to classify a posture code for these based on the OWAS. To achieve this, Experiment 1 was conducted in order to obtain learning data for the machine learning model and Experiment 2 was conducted in order to obtain manually digitized coordinates, and then this data was used to perform posture classification based on the OWAS. In Experiment 1, 10 participants were asked to perform various poses based on the OWAS, and measurements were taken using 16 cameras and an optical motion capture system. As a result, 3,322,672 frames of images and joint information labeled with OWAS posture codes were obtained. In Experiment 2, the joint coordinates were obtained by having the 10 participants manually tap on them, and the accuracy of the manually digitized coordinates was verified by comparing the results with the joint information obtained in Experiment 1. As a result, it became clear that there was a deviation from the joint coordinates acquired by the optical motion capture system, and that the deviation became larger when self-occlusion occurred. Then, as a result of carrying out posture classification based on the OWAS using the manually digitized coordinates acquired in the experiment, the accuracy rate of the AC (Action Category), which is the overall evaluation value used in the OWAS, was 56.7%. Furthermore, in order to improve the accuracy rate of the AC, the class composition ratio in the machine learning data was balanced, and noise was added (variation in coordinates due to manual digitization) to the machine learning data. As a result, the accuracy rate of the AC became 71.7%.
