Level-of-detail assessment of structural surface damage using spatially sequential stereo images and deep learning methods

In this paper, we report an innovative framework for automating structural surface damage assessment in engineering practice. Assessment of structural surface damage has been heavily relied on human-based inspection, which incurs significant cost to stakeholders of civil structures and infrastructure and often severe risk to the inspectors. Recognizing the promise of aerial robotics that can access dangerous locations and envisaging a future of structural inspection that ought to be fully autonomous, we have developed a framework, termed level-of-detail assessment of structural surface damage, that is geared towards real-time implementation for use in practice. The level-of-detail assessment is enabled by a remote sensing approach based on a small Unmanned Aerial Vehicle (UAV or drone) platform with an integrated payload of a low-cost stereo camera and a compact embedded computer. To achieve real-time detection, we propose the use of the faster region-based Convolution Neural Network (faster RCNN) as an artificial intelligence (AI) utility at different imaging depths. The stereo-camera based geometric reconstruction provides the basis of achieving level-of-detail quantitative damage assessment. In this paper, we further propose a novel data preparation method to accommodate the RCNN training. In the end, we will showcase some of these results based on our current implementation and experimental evaluation.