Discrimination of Structures in a Nuclear Facility by Deep Learning Based on 3D Point Cloud

- Acquisition of Projection Images by Virtual Camera and Applying to Learning -

Abstract

This paper describes a method for discrimination of structures in a nuclear facility by deep learning based on projected images generated from a 3D point cloud using a virtual camera. In order to create a training dataset, we constructed a system that automatically generates both projected and labeled images from 3D point cloud data along a pre-defined trajectory path using a virtual camera. Finally, we reported the training accuracy and discrimination performance by the training dataset created by the proposed method from the point cloud measured at the test plant in the Fukushima Robot Test Field.