Fast-PGMED: Fast and Dense Elevation Determination for Earthwork Using Drone and Deep Learning

Abstract

This paper presents a time- and cost-effective elevation determination method for earthwork operations using ready-to-fly imaging drones and deep learning technologies. The proposed method is named the fast pixel grid/group matching and elevation determination (Fast-PGMED) algorithm. The input data are a pair of approximate 2:1-scale top-view images, and the output is the determined elevation map for the scanned station. Feature matching of the two multiscale images is conducted by calculating correlations between target patch predictions (via DeepMatchNet, a fully convolutional network) and potential target patches (via virtual elevation model). The overall processing time is about 21 s (including 5 s for low-high orthoimage assembly, 3 s for patch feature generation, and 13 s for pixel matching) to process a 2,500-pixel grid, and the generated elevation values are as accurate as photogrammetry (within 5-cm error) but took much less time. Moreover, the developed method has been evaluated with two different drones. Volume measurement was quickly conducted via 2D elevation maps and accurately estimated via dense point clouds and Civil 3D.