A self-supervised near-to-far approach for terrain-adaptive off-road autonomous driving

Loading...
Thumbnail Image
Date
2021-07-05
Authors
Mayuku, Orighomisan
Surgenor, Brian W.
Marshall, Joshua A.
Keyword
robotics , visualization , image colour analysis , convolutional neural networks , autonomous vehicles
Abstract
For use in off-road autonomous driving applications, we propose and study the use of multi-resolution local binary pattern texture descriptors to improve overall semantic segmentation performance and reduce class imbalance effects in off-road visual datasets. Our experiments, using a challenging publicly available off-road dataset as well as our own off-road dataset, show that texture features provide added flexibility towards reducing class imbalance effects, and that fusing color and texture features can improve segmentation performance. Finally, we demonstrate domain adaptation limitations in nominally similar off-road environments by cross-comparing the segmentation performance of convolutional neural networks trained on both datasets.
External DOI
Collections