An Off-Road Terrain Dataset Including Images Labeled With Measures Of Terrain Roughness

Abstract

This paper describes the structure and functionality of a dataset designed to enable autonomous vehicles to learn about off-road terrain using a single monocular image. This dataset includes over 12,000 images of off-road terrain and the corresponding sensor data from a global positioning system (GPS), inertial measurement units (IMUs), and a wheel rotation speed sensor. The paper also describes and empirically evaluates eight roughness labeling schemas derived from IMU z-axis acceleration for labeling the images in our dataset. These roughness labels can be used for training deep learning models to detect terrain roughness.