PSI - Issue 64

Jack Smith et al. / Procedia Structural Integrity 64 (2024) 220–227 Smith, J., Paraskevopoulou, C./ Structural Integrity Procedia 00 (2019) 000 – 000

222

3

Fig. 1. Typical masonry walls showing analysis complications.

2. Background Automation of masonry block semantic segmentation from 3D point cloud data has only become technically feasible since developments in both computer vision algorithms and lidar technology. While not previously used on masonry structures, Hackel et al. (2016) created a popular method for semantic segmentation of 3D point clouds that does not use neural networks. Using local features obtained by eigenvalue analysis at varying radii as inputs, a machine learning classifier was trained to determine a segmentation from the image contours. To demonstrate the advantages of deep learning methods, we used this method as a baseline in this study. XGBOOST (Chen and Guestrin, 2016) was used as the classifier with inputs set as the features suggested by (Hackel et al., 2016). XGBOOST (Extreme Gradient boosting) is an ensemble decision tree method that iteratively generates new trees to compensate for errors in the previous trees. One of the first automated methods for masonry joint segmentation from 3D location data was created by Valero et al. (2018) who developed a workflow that operated directly on 3D point clouds using continuous wavelet transforms. While their method has the flexibility to deal with surfaces in any orientation and works well on rubble masonry, performance decreases when joints form only small deviations from the masonry block surface, such as found in many brick walls. More recently, deep learning using neural networks has proven itself as a valuable tool for automating repetitive tasks in the construction industry (Dong and Catbas, 2021; Sony et al., 2021; Sabato et al., 2023). Since the development of VGG by Simonyan and Zisserman, (2014), convolutional neural networks have surpassed traditional computer vision techniques for image analysis. By training image convolutions, effectively image filters applied over consecutive pixel neighbourhoods, convolutional neural networks achieve higher performance than fully connected networks. The memory requirements are reduced, so deeper networks are possible that can achieve a higher accuracy. Encoder-decoder style convolutional neural networks have achieved state of the art performance for 2D semantic segmentation tasks (Ronneberger et al., 2015). They consist of an image encoder, which builds a description of the image, and a decoder that recreates the image with the target areas segmented. The U-Net adds skip connections to this design which improves the accuracy of the segmentation by combining unaltered data from an equivalent stage of the encoder with fully encoded feature representations. The U-Net design has proven enduringly popular due to its excellent transfer learning performance. This enables it to be pretrained on a general task, before being fine-tuned using only a small amount of data from the target domain. Recent studies on damage and joint segmentation of tunnel linings have first unrolled a tunnel 3D point cloud before rasterizing it into a 2D depth map image ready for a 2D U Net style neural networks (Ji et al., 2022; Zhang et al., 2022; Smith et al., 2023a, 2023b). This method takes advantage of a tunnel lining forming a single surface, with joints and damages forming only small offsets that can be fully represented on a 2D depth map. The latter enables it to leverage the relatively developed field of 2D computer vision