Motion Detector Training with Virtual Data for Semi-Automatic Motion Analysis-Elimination of Real Training Data Collection using 3DCG Synthesis

Abstract

A visitor's interest in the exhibits at a museum can be understood by analyzing his or her behavior. Conventionally, behavior analysis is performed manually, but it is troublesome to do this in practice. Therefore, our aim is to reduce this trouble by the automation and semi-automation of behavior analysis. We focus on motion detection by machine learning as the first step in this process. To detect motion using machine learning, it is necessary to collect training motions in advance and create a detector. In this study, human models and motions were generated using three-dimensional computer graphics. Then, we synthesized them and collected training motions using a virtual environment. We used the collected training motions to create a detector and detected motions collected from people in a real environment. The results of the experiments indicate that this training data collection method is effective.