Computer Vision–Based Estimation of Flood Depth in Flooded-Vehicle Images

Abstract

This study proposes a vision-based method for flood depth estimation using flooded-vehicle images with a ground-level view. The proposed method is comprised of three main processes: segmentation of vehicle objects, cross-domain image retrieval, and estimation of flood depth. First, Mask region-based convolution neural network (R-CNN) is used to detect flooded vehicles in flooding images. Second, on the basis of feature maps from VGGNets, dynamic feature space selection is employed to select a three-dimensional (3D) rendered car image most similar to the flooded object using the metric of cosine distance. Finally, the flood depth is calculated through a comparison of the flooded object and the 3D rendered image. The feature maps from Pooling layer 4 of VGG19, under the condition of a cosine distance of <0.55, produces an average error of 7.51 pixels, corresponding to 9.40% of the tire height. A total of 500 flooding images are used to validate the method.