José Ferreira Pinto

Abstract
Advances in technology are leading to more and more devices integrating sensors capable of acquiring data quickly and with high accuracy. Point clouds are no exception. Therefore, there is increased research interest in the large amount of available light detection and ranging (LiDAR) data by point cloud classification using artificial intelligence. Nevertheless, point cloud labeling is a time-consuming task. Hence the amount of labeled data is still scarce. Data synthesis is gaining attention as an alternative to increase the volume of classified data. At the same time, the amount of Building Information Models (BIMs) provided by manufacturers on website databases is increasing. In line with these recent trends, this paper presents a deep-learning framework for classifying point cloud objects based on synthetic data sets created from BIM objects. The method starts by transforming BIM objects into point clouds deriving a data set consisting of 21 object classes characterized with various perturbation patterns. Then, the data set is split into four subsets to carry out the evaluation of synthetic data on the implemented flexible two-dimensional (2D) deep neural framework. In the latter, binary or greyscale images can be generated from point clouds by both orthographic or perspective projection to feed the network. Moreover, the surface variation feature was computed in order to aggregate more geometric information to images and to evaluate how it influences the object classification. The overall accuracy is over 85% in all tests when orthographic images are used. Also, the use of greyscale images representing surface variation improves performance in almost all tests although the computation of this feature may not be robust in point clouds with complex geometry or perturbations. © 2022 American Society of Civil Engineers.

Abstract
Advances in technology are leading to more and more devices integrating sensors capable of acquiring data quickly and with high accuracy. Point clouds are no exception. Therefore, there is increased research interest in the large amount of available light detection and ranging (LiDAR) data by point cloud classification using artificial intelligence. Nevertheless, point cloud labeling is a time-consuming task. Hence the amount of labeled data is still scarce. Data synthesis is gaining attention as an alternative to increase the volume of classified data. At the same time, the amount of Building Information Models (BIMs) provided by manufacturers on website databases is increasing. In line with these recent trends, this paper presents a deep-learning framework for classifying point cloud objects based on synthetic data sets created from BIM objects. The method starts by transforming BIM objects into point clouds deriving a data set consisting of 21 object classes characterized with various perturbation patterns. Then, the data set is split into four subsets to carry out the evaluation of synthetic data on the implemented flexible two-dimensional (2D) deep neural framework. In the latter, binary or greyscale images can be generated from point clouds by both orthographic or perspective projection to feed the network. Moreover, the surface variation feature was computed in order to aggregate more geometric information to images and to evaluate how it influences the object classification. The overall accuracy is over 85% in all tests when orthographic images are used. Also, the use of greyscale images representing surface variation improves performance in almost all tests although the computation of this feature may not be robust in point clouds with complex geometry or perturbations. (C) 2022 American Society of Civil Engineers.

Abstract

Abstract