Detection and Posture Recognition of Operators at Nuclear Power Plant Operations Training

Abstract

Operations trainings of nuclear power plants are performed in full-size power plant simulators. The training is always recorded on videotapes. Many examples such as optimal and not optimal operations are included in the video images. This paper proposes a detection and posture recognition method of nuclear power plant operators from the video images. Posture recognition is one of the most difficult tasks in image processing, because an operator usually takes various postures during power plant operations. To recognize their postures, the method uses the four postures that have been classified by the cognitive scientists engaged in human factors research of nuclear power plant operations. In the method, silhouettes of the operators are detected by several image processing techniques such as a background subtraction, noise reductions and etc. The detected silhouettes are classified into the four postures by their moving speed and a machine learning. Their operations are summarized and visualized with human body computer graphics (CG). In evaluation experiments with over twenty thousand images, the detected silhouettes were classified to the four postures successfully and the operations training were summarized with the human body CG based on the classified postures.