An unmanned aerial vehicle implemented network for real-time crack detection in stone cladding

Abstract

Crack damage to stone cladding occurs widely owing to the impact of disasters and degradation over its service life. However, current crack detection measures are not suitable for automatically identifying and measuring cracks in tall and large buildings with cladding. An unmanned aerial vehicle (UAV)-implemented network for real-time crack inspection in stone cladding, CladdingNet, was developed to evaluate the structural safety of stone cladding. The CladdingNet system consists of a binocular camera, a graphics processing unit (GPU), and a UAV. The damage states of cracks were defined for the convenience of classification and identification of the damage level of the stone cladding. A route map was constructed automatically based on the 3D model of the target building to identify the location of each cladding panel, particularly the cracked panels. The 3D coordinates of the cracked panels can be conveniently determined using the node number on the map. The frame rate of CladdingNet ranges from 1.0 to 10.0 fps, which is upgradable by improving the performances of the binocular camera, GPU, and deep learning algorithm. The trained and validated loss functions agreed with each other. However, the magnitude of the focal loss was approximately 20% that of the cross-entropy loss. MobileMamba, ResNet50, DeepLab-v3, PSPNet, and U-Net were employed to compare the accuracies of CladdingNet, which is the highest among the five models. The CladdingNet-based UAV platform was beneficial for defect inspection of building cladding and other structures.