Rapid Terrain/Map Generation for Robotic and Autonomous Vehicle Simulations

The objective of this project is to build 3D terrain models from satellite imagery and Digital Elevation Maps (DEM) using the Unreal Engine platform. Terrain and other objects are added to the 3D model by classifying objects from RGB satellite images. The classification of objects is typically the most time-consuming step within the entire process, which is currently done manually by a user. In Phase I, a prototype for automated classification will be demonstrated using a fast approach known as random forests. This classifier will be trained, requiring only 10-20 training images.  It will also be designed to label objects using a continuous class system to deal with mixed terrain, e.g. soil and grass. A simple and fast inpainting technique is proposed to label occluded regions under trees and shadows. The technique will pinpoint regions that should be inspected by the user such as unclassified objects. This will allow the user to overwrite, modify, and inspect the model. In Phase I, Lickenbrock will develop a working prototype and demonstrate the user intervention with the software. In Phase II, Lickenbrock will extend this work and also develop the plugin software to export data to Unreal Engine.