Tunnel cracks detection via Geographical Object-based Image Analysis

Abstract

The inspection of tunnel lining conditions is have been carried out with many technologies among which LiDAR is the most efficiency method for crack detection However it collects an immense amount of point cloud data which could be handled manually Automatic data processing didn’t provide enough function in determining the existence of cracks Moreover a majority of current tunnel crack detection software use image processing or deep learning to detect abnormalities These methods require extensive training and time-consuming detection for accuracy of only approximately 80% Therefore it is an important issue to ameliorate the problems faced by LiDAR in tunnel crack detection To improve the efficiency of auto-detection via point clouds form at tunnel linings a combined routine is established The point cloud data are then filtered and cropped for visibly discernible cracks to translate into image format Image recognition for crack detection was then performed by combining image processing with three detection methods: Principal Component Analysis (PCA) Object-Based Image Analysis (OBIA) and Geographic Object-Based Image Analysis (GEOBIA) Crack measurements were produced from these image recognition results and then compared with the actual measurements to determine the level of error The types of tunnel lining examined in this study are concrete tunnel linings and brick tunnel linings which were evaluated using the accuracy rate error rate false-positive rate false-negative rate and Kappa coefficient after recognition via the three detection methods From the three applied methods GEOBIA produced the best results This is because of its automated selection of the scale of image segmentation along with its image recognition capability that outperforms pixel-based classification due to its implementation of elements in addition to pixels The identification results of the two types of lining have consistent accuracy rates of over 95% and Kappa coefficients of over 0 85 In terms of measurements the maximum error of length in the recognition results was between 4–34% and the maximum error of width was between 23–88%; the minimum width detected was 1 1 mm

Date of Award	2019
Original language	English
Supervisor	Ting-To Yu (Supervisor)