Extraction and recognition of components from point clouds of industrial plants

Abstract

Point clouds of industrial plants are very useful for supporting renovation planning, production and product design, asset management, and so on. However, point clouds of an industrial plant contain a large number of components. In order to utilize point clouds, it is necessary to extract each component from point clouds and identify its type. In this paper, we discuss methods for identifying component types in industrial plants using machine learning. In our method, cylinders and planes are detected from point-clouds and candidate component regions are extracted. Since point clouds captured using the terrestrial laser scanner can be mapped on the 2D grid, convolutional neural network (CNN) designed for images can be applied. Three types of 2D images are generated from point clouds, and they are used for classification. To increase the numbers of training data, depth images are augmented using CAD models. In evaluation, nine classifiers were trained and evaluated. By comparing the nine CNN models, we discuss classifiers suitable for recognizing components in industrial plants.