Design and Implementation of a Long Range Visual Terrain Classifier for Legged Robots
In Proceedings of the World Symposium on Computer Applications & Research WSCAR'2014, (WSCAR-2014), 18.1.-20.1.2014, Sousse, IEEE, pages 117-122, Jan/2014. ISBN: 978-1-4799-2805-7.
A key challenge in autonomous legged robots is the extraction of meaningful information from sensor data, which would allow a good interpretation of the nearby terrain, and a reasonable assessment of more distant areas. In order to navigate efficiently, legged robots need to assess the nearby terrain, avoiding obstacles and other hazards such as excessive tilt and roughness. Although the field of automated terrain classification is relatively new, its advances and goals are scattered across different robotic platforms and applications. In this paper, we present an automated terrain classification approach that works with a single camera and maintains high classification rates that are robust to varying lighting conditions. Terrain is classified using keypoint descriptors created from speeded up robust features (SURF) with a support vector machine (SVM) classifier. This technique works by extracting salient features, and matching these to a database of pre-extracted features to perform the classification. We demonstrate the good performance of our approach with the results of the experiments done to classify patches of different terrain types.