Self-Supervised Near-to-Far Learning for Terrain-Adaptive Off-Road Autonomous Driving

Abstract

Autonomous vehicles are an increasingly important tool for exploring unknown off-road environments where navigation decisions must include considerations of terrain traversability, unlike in urban environments that have engineered roads. In this thesis, a new self-supervised method is proposed for prioritizing traversable terrain while autonomously navigating a vehicle to a goal position in an unknown off-road environment. Leveraging the colour discriminant bias of off-road terrain types, and using images from a vehicle-mounted camera, a homography that maintains the real-world planar layout of the terrain is employed to cluster terrain types by colour and autonomously register corresponding traversability characteristics such as roughness and slip for terrain-adaptive navigation. As it navigates the vehicle, the algorithm also generates training images for use in contemporary end-to-end navigation schemes. Compared to the existing expert-guided near-to-far approaches, the test results demonstrate the higher autonomy introduced by the proposed approach for navigating off-road environments with unknown traversability characteristics, and highlight its fit to contemporary supervised semantic segmentation schemes that require foreknowledge of traversability characteristics, are limited by insufficient data, and suffer significant class imbalance and poor cross-domain performance. Finally, the effectiveness of non-discretionary self-supervised image labelling is discussed.