VISION-BASED TERRAIN CLASSIFICATION AND LEARNING TO IMPROVE AUTONOMOUS GROUND VEHICLE NAVIGATION IN OUTDOOR ENVIRONMENTS

Abstract

Terrain is an important factor for autonomous ground vehicles (AGV), potentially ruining a mission or the platform itself. The purpose of this thesis is to develop a method for an AGV to identify and avoid hazardous terrain. This work builds on a previously developed system that uses artificial potential fields to avoid obstacles and navigate to a goal. Terrain was identified by developing a random forest machine-learning algorithm, classifying terrain as hazardous or traversable. The random forest was grown using data from images collected during this work. The classification of hazardous terrain was used to generate a repulsive force for use with artificial potential fields. The system was designed to avoid known areas of hazardous terrain using path planning, developing paths using approximate cell decomposition and the A* search algorithm. Tests of the developed random forest revealed accurate classification capabilities for all terrain types, but a tendency to misclassify certain terrain types. Portions of the navigation solution were simulated and confirmed the path planning capability. Trials conducted in a real-world environment revealed the solution stopped the AGV from entering hazardous terrain, and successfully planned routes around hazardous terrain. Improvements to the localization solution will allow the AGV to perform more consistently and over longer ranges.