A deep learning-based indoor acceptance system for assessment on flatness and verticality quality of concrete surfaces

Abstract

Flatness and verticality quality assessment (FVQA) must be strictly controlled during the indoor acceptance testing. Existing FVQA methods using terrestrial laser scanning (TLS) rely on the as-designed model to assist in locating inspected objects, while sophisticated deep learning techniques excel in the object recognition. Therefore, based on the TLS and deep learning technique, this study presents a general indoor acceptance system including the indoor semantic segmentation, component surface segmentation, and FVQA. In particular, a dataset consisting of 145 sets of room point cloud data (PCD) is created to train the adopted point-based neural network. An image processing-based method is proposed to check the flatness and verticality of identified concrete surfaces. Experiments are conducted on two completed residences to verify the practicality and convenience of the proposed indoor acceptance system by comparing the calculated flatness and verticality with the field measurements. The experimental results show that the deep learning technique used in this study has a good recognition accuracy (over 85%) for large concrete surfaces in residential PCD, and the proposed indoor acceptance system provides more rigorous inspection results compared with the manual inspection method.