Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : September 15, 2021 - September 17, 2021
In recent years, application of virtual reality (VR) for work-related training at production sites has progressed rapidly, but there is a problem of insufficient training owing to the lack of educators. The aim of this study is to verify automatic motion analysis using a VR-based training system such that trainees can recognize and learn critical information for self-improvement at work. In this study, we used video cameras to capture an operator’s motions and actions during the outer rounding and end-face cutting operations on a lathe. For the motion analysis, we used a learning model for time-series segmentation, which extracts features from video images and predicts the class of each frame. We prepared six labels, including four labels for advanced and penetration-type movements as well as two labels for static and transporting empty. A model was created and trained with 90 video data samples of several minutes each, and the training results showed more than 80% success with several metrics, such as accuracy, edit distance, and segmentation F1 score with an overlapping threshold of 50%. These results suggest the feasibility of application of the proposed scheme to training systems.
