AUTOMATIC INSPECTION OF BUILDING CRACKS FROM AERIAL IMAGES ACQUIRED BY DRONES

Resumo

The application of machine learning (ML) and computer vision (CV) in inspections of civil construction works has shown significant advances, although there is still considerable room for the exploration and development of these technologies in this field. This study proposes a computational method for automatic detection and measurement of cracks in buildings, using images captured by drones. The proposed method comprises two main stages. In the first stage, convolutional neural networks (CNN) are applied using the YOLOv8 framework to develop an automatic crack detection model, which was trained and validated using two image datasets. The second stage involves applying image processing techniques to extract and measure the geometric characteristics of the detected cracks, such as length and propagation angles. The crack detection stage achieved its best performance with a recall of 77.5% and a mean average precision (mAP) of 79.7%. In the second stage, the crack’s length and predominant angle are calculated from a straight line segment connecting the pixels representing its endpoints. To allow these values to be compared with real measurements, system calibration is required for the correct conversion from pixels to length units. The results obtained from the computational experiments indicate the potential of the proposed method for application in automated inspections