地上レーザースキャンデータを用いた鉄筋コンクリート建築物の損傷検出フレームワーク

• Material Type: 博士論文
• Title: 地上レーザースキャンデータを用いた鉄筋コンクリート建築物の損傷検出フレームワーク
• Author/Editor: 邵, 万朋
• Author Heading: 邵, 万朋
• Publication Date: 2023-12-27
• Publication Date (W3CDTF): 2023-12-27
• Alternative Title: A Framework for Damage Detection of Reinforced Concrete Buildings Using Terrestrial Laser Scanning Data
• Degree Grantor: 九州⼯業大学
• Date Granted: 2023-12-27
• Date Granted (W3CDTF): 2023-12-27
• Dissertation Number: 甲第389号
• Degree Type: 博士（情報工学）
• Text Language Code: eng
• Subject Heading: Earthquake; Reinforced concrete building; Damage detection; Terrestrial laser scanner; Reflection Noise Removal
• Target Audience: 一般
• Note (General): 九州⼯業⼤学博⼠学位論⽂ 学位記番号：情工博甲第389号 学位授与年⽉⽇: 令和5年12⽉27⽇
• DOI: 10.18997/0002000371; https://doi.org/10.18997/0002000371
• Persistent ID (NDL): info:ndljp/pid/13343393; https://dl.ndl.go.jp/pid/13343393
• Collection: 障害者向け資料
• Collection (Materials For Handicapped People:1): テキストデータ
• Collection (particular): 国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文; https://dl.ndl.go.jp/collections/A00014
• Acquisition Basis: 博士論文（自動収集）
• Date Accepted (W3CDTF): 2024-03-01T22:05:23+09:00
• Format (IMT): application/pdf
• Access Restrictions: 国立国会図書館内限定公開
• Service for the Digitized Contents Transmission Service: 図書館・個人送信対象外
• Availability of remote photoduplication service: 可
• Periodical Title (URI): https://doi.org/10.18997/0002000371; http://hdl.handle.net/10228/0002000371
• Data Provider (Database): 国立国会図書館 : 国立国会図書館デジタルコレクション; https://dl.ndl.go.jp

• Summary, etc.: Damage assessment is critical important for post-earthquake buildings, especially for the reinforced concrete buildings. It is first to know and detect local damage of damaged buildings (damage types, locations, etc). Normally, visual inspection is conducted for investigation of post-earthquake buildings. However, it need a significant amount of time and manpower to survey a city-level damaged buildings, as well as the high risk of collapse such as aftershocks. Terrestrial laser scanning (TLS), a non-contact measuring instrument, can capture detailed shape of the target building in point clouds with an error of a few millimeters at certain distance. So it is possible to confirm local damage such as concrete spalling and cracking by visualizing the damaged building in 3D point cloud. However, even we have damaged buildings in point clouds, we still need to manually label local damage of them by visual inspection, which is time consuming and have a risk of overlooking local damage in a citywide damaged buildings. To detect local damage of the building surface in 3D point clouds quickly and accurately, automatic damage detection is needed. Recently, learning networks has achieved huge success in classification of various 3D models. If the deep learning technique can be applied for damage detection, which will help us to detect local damage of the building surface quickly and accurately. Although various deep learning methods have been studied for deep learning using 3D point clouds, direct learning from point clouds like PointNet, PointNet++ without preprocessing of the 3D point clouds has become the mainstream method. Therefore, the objective of this study is automatically detecting the visually observable damage in the reinforced concrete building in point clouds by 3D deep learning methods. For automatically detecting local damage of reinforced concrete building in TLS point clouds, our research is divided into 3 parts: ・Application of 3D DNN for damage detection ・Improvement of damage detection by 3D DNNs investigation ・Multi-type damage detection with color information by 3D DNN For application of 3D DNN for damage detection: We solve the local damage detection problem not as a detection problem but as a classification problem. Before feeding the damaged building into 3D DNN, the whole building surface is divided into small grids. Then, all samples are classified by 3D DNN one by one. As a deep neural network to classify the voxel grids, we used PointNet++. Then, we apply our damage detection model to detect “spalling” areas of a post-earthquake building model at first. In order to detect more “spalling” areas, we propose a shifting technique. With shifting, 92.9% of “spalling” samples can be correctly classified. After that, we also apply our damage detection model for a damaged building on Hashima island. Even 94.8% of “spalling” samples are correctly detected, the precision is considerably low with only 21.5%. The reason is “spalling” areas of damaged building on Hashima island look as smooth as than “intact” ones due to years of corrosion. Any uneven surface on the wall could potentially be identified as “spalling”, which is undistinguishable. To improve the damage detection rate, we investigate the utility of various 3D deep neural networks for “spalling” detection of the damaged building. We compare the performance of different kinds of 3D DNNs for damage detection to find out the most effective 3D DNN for damage detection. Meanwhile, we find a pair of optimal parameters after investigating the impact of different point sizes and voxel sizes on detection results. We demonstrate that PointNet++ SSG with 0.15m voxel size and 1024 sample size can achieve the best performance in our study. For multi-type damage detection with color information by 3D DNN, in the real scene, after earthquake, the RC building will show two types of local damage: crack and spalling. We firstly apply our damage detection model for multi-type damage detection of the post-earthquake building model. Experimental results show that it is still challenging to differentiate “crack”, “spalling” from “intact” samples only based on TLS geometry data. By comparison, 88.7% of the “crack” and “spalling” areas of the damaged building in TLS point clouds with color can be correctly detected by the 3D DNN. For some misclassified cases, we also propose a four-step post-processing method based on cluster analysis including removal of isolated damaged clusters, dilation, and filling, etc. to develop the detection results. Finally, the detection ratio increase to 92.8% after the post-processing. However, when using the TLS to measure a target building with glass components. Reflection noise points will be produced at the opposite of the glass object and mixed with the real-world points. To prevent damage detection by 3D DNN from being affected by noisy points, these noisy points should be removed at first. For reflection removal of RC building in TLS measurement, our study is separated into three steps: (1) For glass objects extraction: we apply an unsupervised segmentation method (super points graph) based on the intensities and geometric features to partition the whole scene into many simple geometric shapes with similar features. Then, glass objects are extracted by modelling the distribution of the average intensities of all segments and a sequence of constraints. (2) For affected points detection: we find the points that fall in affected areas based on the laser scanner position and the four corner points of every glass component. (3) For noisy points detection: we detect the reflection noise points based on the final scores computed by the reflection symmetry and geometric similarity. Experimental shows around 90% reflection noise can be selectively removed. This paper proposes and demonstrates the effectiveness of the 3D DNN for damaged detection of the post-earthquake building. Moreover, our reflection noise removal method can selectively delete most of the noisy points reflected by glass windows and doors without deleting real-world point as much as possible.
• DOI: 10.18997/0002000371; https://doi.org/10.18997/0002000371
• Format (IMT): application/pdf
• Source: fulltext; https://kyutech.repo.nii.ac.jp/record/2000371/files/jou_k_389.pdf
• Access Restrictions: インターネット公開
• Data Provider (Database): 国立情報学研究所 : 学術機関リポジトリデータベース（IRDB）（機関リポジトリ）; https://irdb.nii.ac.jp
• Original Data Provider (Database): 九州工業大学 : キューテイカー

Table of Contents

1 Introduction | 2 3d Point Cloud Measurement Using Terrestrial Laser Scanner | 3 Application Of 3d DNN for Damage Detection | 4 Multi-Type Damage Detection of Damaged Building By 3d DNN | 5 Noise Removal Using Location Information and Shape Similarity | 6 Conclusion