請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/86207
標題: | 基於建築資訊模型以及三維重建模型結合影像分割輔助工地墜落安全檢測 BIM and reality model-driven fall hazard recognition using semantic segmentation |
作者: | Yi-Han Lai 賴意函 |
指導教授: | 謝尚賢(Shang-Hsien Hsieh) 謝尚賢(Shang-Hsien Hsieh | shhsieh@ntu.edu.tw | ), |
關鍵字: | 墜落危險檢測,建築資訊模型,三維重建,電腦視覺, Fall Hazard Recognition,Building Information Modeling (BIM),3-Dimensional Reconstruction,Computer Vision, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 從高處墜落一直以來是全球建築業安全事故的主要原因,其中許多發生在施工架和未受保護的開口。研究人員已嘗試利用建築資訊模型(Building Information Modeling, BIM)和時程規劃表通過模擬的方式,從模型中找出潛在的墜落危險,並在安全規劃階段採取相關措施。最近也有許多研究開發了基於電腦視覺以及機器學習的方法來辨識有關工地開口的問題。然而即使倚靠建築資訊模型來進行安全規劃以及利用電腦視覺來促進違規檢測,上述兩種方法都無法改善現階段的安全檢查仍然依賴耗時且費力的人工走動檢查的特性。因此,主要的挑戰是實現建築資訊模型以及現實模型之間訊息的雙向利用,例如在建築資訊模型中加入現場的狀況或是在現實模型以及深度學習方法中加入建築資訊模型的先驗知識。本研究提出一種結合建築資訊模型、現實模型和深度學習的集合方法,以系統性地進行安全規劃和檢查。具體來說,提出的方法首先在建築資訊模型上開發基於規則的安全檢查模組,用於辨識設計規劃階段的潛在墜落危險,為整體系統提供先驗訊息。接著,利用一組有組織的一般相機、360相機、或是空拍機所拍攝的影像,透過三維重建(Structure from Motion, SfM)以及多視立體(Multi-View Stereo, MVS)將這些影像建置成現實模型。接下來,對同一組影像進行語義分割來辨識施工架以及開口,並將結果反投影到三維的現實模型中。最後,本研究提出演算法來分析這些物件的相互關係,並將分析結果可視化以供檢查。本研究經工地實際案例測試後,驗證提出之方法能成功辨識現實模型中的施工架以及開口。其中實驗結果顯示,二維的語意分割在施工架以及開口上的表現,平均精度(Average Precision, AP)可分別達到90.02% 和 89.62%,在三維的點雲模型中施工架的準確率為78%左右。此外,在估計施工架與相鄰結構體的距離成果,達到低於5公分的誤差,這些成果展現所提出的方法能夠有效支持在建築資訊模型和現實模型中的自動墜落安全檢測。 Falls from height has been the leading cause of construction safety accidents across countries, many related to scaffolding and uncovered openings. Researchers have tried utilizing Building Information Modeling (BIM) with the schedule to find potential falling hazards through simulations and react in the safety planning phase. Also, recent studies developed computer vision-based machine learning approaches to identify the opening related subjects. While BIM supports safety planning and vision-based methods facilitate violation detection, safety inspection still relies on time-consuming and labor-intensive job walks that fall short of leveraging either approach. The main challenge remains to enable bidirectional information utilization between BIM and reality models, such as onsite situation awareness in BIM and prior BIM knowledge usage in reality models and deep learning methods. This research proposes an integrated approach incorporating BIM, reality model, and deep learning to systematically conduct safety planning and inspection. Specifically, a rule-based safety checking application on BIM is developed for identifying the potential fall hazards in the design phase, providing prior information to the system. Afterward, given a set of organized point-and-shoot, 360 or drone images, the system uses a typical Structure from Motion (SfM) and Multi-View Stereo (MVS) pipeline to generate the reality model. Next, the semantic segmentation is adopted to the same images set to recognize the scaffolding and opening, and the results will be back-projected to the reality model. Finally, the interrelationships of these elements are analyzed, and the results are visualized for inspection. We have evaluated the proposed method on a real-world construction site and successfully identified the scaffolding and opening in the reality model. The experimental results demonstrate that the average precision of 2D image segmentation on scaffolding and opening achieved 90.02% and 89.62%, respectively; and the accuracy of 3D point cloud segmentation on scaffolding obtained around 78%. Also, the distance between scaffolding and structure obtained approximately 5cm error, which exhibits the proposed method's capability to support the automatic fall hazard detection in BIM and point clouds. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/86207 |
DOI: | 10.6342/NTU202202931 |
全文授權: | 同意授權(全球公開) |
電子全文公開日期: | 2022-09-06 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2908202214595700.pdf | 60.19 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。