Real-Time Concrete Damage Detection Using Deep Learning for High Rise Structures

Abstract

Nowadays, the number of aging high-rise civil structures is growing throughout the world, and most of them use concrete as a building material. Concrete may lose its strength due to continuous loading and environmental impacts. As a result, damage (crack and spall) may occur on the exterior surface of the structure. Whenever these deformities are left unexplored and untreated, the integrity of the structure may be compromised. Therefore, regular maintenance of the structure is needed. Some previous vision-based studies have used a drone as a vehicle to capture and record the current state of the structure. Later, captured videos and images were analyzed to determine damage using object classification, localization, and segmentation methods. Also, in a few studies, drones relay the collected data to the server using a wireless medium. However, the developed systems are very complex, time-intensive, and require high bandwidth. To address these problems, this paper uses the edge computing principle to propose a real-time multi-drone damage detection system using one of the advance deep learning models called You Look Only Once-version3 (YOLO-v3) for high-rise civil structures. The proposed system uses Jetson-TX2 as a hardware platform to run YOLO-v3 and is deployed on Pixhawk’s hardware standards-based open source hexacopter drone. During a coordinated survey, if the damage is observed, the Jetson-TX2 on-board, after processing, sends only the damage-related information to the server that resides on the ground through the Wi-Fi channel. Our proposed system is evaluated on a dataset containing 800 ( $480\times480$ pixels) images of different types of damage, collected from multiple structures of CSIR-CEERI, Pilani. Manually annotated images are used for the training and testing of the modified YOLO-v3 classifier. As a result, the proposed approach offers reliable performance with an accuracy of 94.24% and can process an image in 0.033 seconds.