Terrain Classification from UAV Flights Using Monocular Vision

With the popularization of small Unmanned Aerial Vehicles (UAVs) and their usage diversification among various fields, such as aerial mapping applications, it is important to develop better terrain following techniques that rely solely on the vehicle's sensing capabilities. The objective of this paper is to evaluate whether is possible to gather information about terrain inclination and elevation from monocular video captured from such aircrafts. Our approach is biologically inspired by trying to reproduce some insects behaviour with the use of optical flow to infer about the terrain. We built an UAV specifically for this research which uses a gimbal stabilized down-facing camera and flew it at a fixed Above Sea Level (ASL) altitude. After performing preliminary analysis on sparse optical flow data and validating the concept, we moved towards a dense optical flow algorithm and created different descriptors to feed multiple decision trees in order to infer about terrain characteristics. We achieved accuracies of 77.34%, 86.75% and 91.85% depending on the evaluated characteristic, showing that our approach is valid.