Real-time reconstruction of heightmaps from images taken with an UAV

Abstract

In this paper, an approach for a real-time three dimensional reconstruction of a dense point cloud, from images obtained with an unmanned aerial vehicle, is presented. The approach also estimates the model scale, using a known marker as reference, obtaining a fairly precise heightmap of the environment. The proposed image analysis framework is also capable of an off-line 3D reconstruction from monocular images, without known markers in the environment. The proposed framework is an adaptation of the Patch-based Multiview Stereo (PMVS) algorithm, in order to obtain a faster point cloud generation, suitable for real-time environment reconstruction. For the obtained heightmap, the error estimation ranges between 1–1.5 % for manual detection of marker and 4–5 % for automatic marker detection, which we consider is accurate enough for autonomous navigation and path planning.