室內建築檢測用機器學習驅動擴增實境工具之開發與實作

Abstract

結構損傷的目視檢查被視為一種快速且具成本效益的建築狀況評估方法；但是，在實際操作中，目視檢查常因為其高度依賴人力與耗時的特性而效率低下，且經常導致測量不準確或損壞識別錯誤。此外，目視檢查也面臨許多挑戰，例如合格工程師人力短缺、工程師與屋主之間的排程衝突，以及居民對隱私的顧慮。這些問題綜合起來，導致維修作業延誤，並對公共安全構成潛在風險。為了解決上述限制，本研究提出一套整合機器學習(Machine Learning)與擴增實境（Augmented Reality）技術的工具，目的是提升室內結構構件目視檢查的準確性與效率。 擴增實境解決勞力密集、耗時與測量不準問題的問題。擴增實境能快速且有效地掃描室內空間，並即時生成精確的三維重建模型，取代傳統人工測量，不僅大幅減少檢查所需的時間與人力，也提高測量的準確度。透過建立沉浸式與可互動的數位環境模型，擴增實境讓檢查人員能快速評估場域並追蹤掃描進度。機器學習則用於彌補專業人員短缺的問題，透過自動辨識結構損壞，即使是缺乏經驗的實習人員也能執行初步檢查。本研究採用 YOLOv8 物件偵測模型，協助現場人員即時辨識潛在損壞並進行記錄。後續由合格工程師對檢查結果進行複核與驗證，以確保物件偵測模型所偵測之損壞準確無誤。此流程不僅提升了專業人力的使用效率，也提高了整體檢查流程的可靠性。 本工具專為 iOS 裝置開發，結合 Apple 的 ARKit 與 RoomPlan API，可建立精確的室內三維模型。並整合 YOLOv8 物件偵測模型，自動於擴增實境環境中辨識並標註結構損壞。檢查人員可於掃描過程中透過射線投影（Raycast）即時記錄損壞，並將其位置準確標示於根據三維模型所生成的二維平面圖上。擴增實境與 ML 的結合不僅簡化了檢查流程，減少人工測量與人力需求，也提升了損壞紀錄的精確度，例如損壞位置與長度的量測。根據實地測試結果，該工具可有效辨識並記錄如裂縫與混凝土剝落等結構損壞，與傳統方法相比，在檢查準確性、所需時間與整體工作流程方面皆顯著優化。 最後，本系統可將損壞照片、二維平面圖與三維模型整合儲存於同一資料夾中，實現無縫的紀錄與文件整理流程。未來研究將著重於提升裝置端報告自動生成能力、提升模型偵測準確率，以及擴展可辨識的損壞類型。此研究展示了 擴增實境與 ML 的結合如何革新傳統目視檢查流程，使其更加高效、精準與具擴展性。

Visual inspection for structural damage is regarded as a quick and cost-effective method to assess building conditions; however, in reality, visual inspections often pose inefficiencies due to their labor-intensive and time-consuming nature, frequently yielding inaccurate measurements and misidentifications. Additionally, these inspections face challenges such as a shortage of qualified engineers, scheduling conflicts between engineers and homeowners, and concerns about residents' privacy. Combined, these issues can result in delayed repairs and pose potential risks to public safety. To address these limitations, this research presents the development and implementation of a machine learning-powered augmented reality inspection tool that integrates augmented reality (AR) and machine learning (ML) to enhance the accuracy and efficiency of visual inspections for indoor structural members. Augmented reality offers a solution to the problems of labor intensity, time consumption, and inaccurate measurements. AR enables the fast and efficient scanning of indoor environments, providing precise, real-time 3D reconstructions of spaces that eliminate the need for manual measurements, thereby reducing the time and effort required for inspections while simultaneously enhancing accuracy. By creating an immersive and interactive digital model of the environment, AR allows inspectors to quickly assess the area and track the progress of their scanning efforts. Machine learning addresses the shortage of qualified engineers by automating the identification of structural defects. This research utilizes YOLOv8, a widely adopted algorithm for real-time object detection, to assist staff members by identifying potential defects in real-time, enabling them to document and record the findings. After the inspection, qualified engineers can later review and cross-check the results for accuracy, ensuring that the defects identified by the ML model are correctly assessed. This approach not only optimizes the use of skilled personnel but also enhances the reliability of inspections. The tool, designed for iOS devices, leverages Apple's ARKit and RoomPlan APIs to create accurate 3D reconstructions of indoor spaces. The YOLOv8 object detection model is integrated to automatically identify and anchor structural defects in the AR environment, allowing inspectors to raycast and document defects during the scanning process, accurately marking their locations on a 2D floor plan generated from the 3D model. The combination of AR and ML streamlines the inspection process by reducing the need for manual measurements, minimizing the number of personnel required, and enhancing the precision of defect documentation, such as the location and length measurements of defects. Field tests demonstrate the tool's practical effectiveness in identifying and documenting structural defects, such as cracks and spalls. These tests reveal improvements in inspection accuracy, reduced inspection time, and an overall more efficient workflow than traditional methods. Additionally, the tool's ability to consolidate defect images, 2D floor plans, and 3D models into a single folder on the device ensures a seamless documentation process. Future work will focus on enhancing on-device report generation, increasing model accuracy, and expanding the tool's capabilities to detect a broader range of structural defects. This research highlights how the integration of AR and ML can revolutionize traditional visual inspections, making them more efficient, accurate, and scalable.