結構裂縫偵測與三維空間資訊萃取

Abstract

建築物是我們生活和工作的重要場所，其結構健康直接關係到我們的安全和舒適。然而，若結構元件損壞或變形，建築物的整體穩定性和安全性就會受到影響。特別是台灣位處地震帶，降雨量豐富且集中，建築物的結構健康更需重視。裂縫是建築物受外力破壞的表現之一，目前建築物裂縫的偵測主要依靠結構技師的現地評估，是高成本且危險的。本研究旨在解決現有自動化裂縫偵測研究中缺乏完整幾何資訊的問題，提出基於運動恢復結構（Structure from Motion, SfM）和YOLOv8影像分割模型的裂縫幾何資訊計算方法。實驗中印製四組常見室內牆面裂縫形式的真實圖片，並模擬真實應用情境，以手機拍攝目標牆面。實驗結果顯示，本方法對於裂縫長度的誤差約為1.5 cm（3 %）、傾斜角度的誤差約為1°（4 %）、寬度的誤差約為0.43 mm（26 %）、中心位置的誤差約為2 cm。除了計算裂縫的幾何資訊，本研究將裂縫邊界框範圍內的像素點映射回場景點雲中，以彌補牆面在點雲上無法清楚顯示的缺陷。本研究提供了一種快速、低成本的自動化裂縫檢測方法，具有一定的準確性和穩定性，可供專業人員進行進一步的評估參考。

Buildings are essential places for our daily life and work, and their structural health is directly related to our safety and comfort. However, if structural elements are damaged or deformed, the overall stability and safety of the building will be affected. This is particularly important in Taiwan, which is located in a seismic zone with abundant and concentrated rainfall, making the structural health of buildings a critical concern. Cracks are one of the manifestations of external force damage to buildings. Currently, crack detection in buildings mainly relies on on-site evaluations by structural engineers, which are both costly and hazardous. This study aims to address the lack of complete geometric information in existing automated crack detection research by proposing a method for calculating crack geometric information based on Structure from Motion (SfM) and the YOLOv8 image segmentation model. In the experiment, four sets of real images of common indoor wall crack forms were printed and simulated in real application scenarios by using a mobile phone to capture the target wall. The results showed that the proposed method achieved an error of approximately 1.5 cm (3 %) in crack length, 1° (4 %) in crack angle, 0.43 mm (26 %) in crack width, and 2 cm in the center coordinates. In addition to calculating the geometric information of cracks, this study also mapped the pixel coordinates within the crack bounding box back to the scene point cloud to compensate for the inability to clearly display the wall in the point cloud. This study provides a fast, low-cost automated crack detection method with a certain degree of accuracy and stability, offering a reference for further evaluation by professionals.