Deep Learning-based Vehicle Image Matching for Flooding Damage Estimation

Abstract

Images representing flooding damages can provide valuable information, such as the damage location and severity. Automated and quantifiable analyses of those images allow asset managers to accurately understand the vulnerability of the infrastructure. To this end, this paper proposes a methodology to match a vehicle in a flooding image to a 3D vehicle image. The proposed method is a part of a framework for flooding depth estimation. As the initial step of the framework, the proposed method uses Mask R-CNN and VGG network to extract the vehicle object and its features, respectively. The features of the vehicle images are compared with those of 3D vehicle image, to find a good match. A total of 87 vehicle objects were used to validate the proposed method, and promising levels of matching accuracy were obtained. Once the framework is completed, the proposed method is expected to automatically analyze flooding images for its damage assessment.