Terrain Classification and Segmentation using Non-Semantic Range Data

Terrain classification is an important component for path planning and control of Autonomous Ground Vehicles that operate in unstructured and uneven terrain. Current terrain classification has been predominantly achieved through the use of visual and semantic spatial data. This paper presents a method, for terrain type classification and scene segmentation, as well as a confidence measure in classification, based on non-semantic range data. The features used in the inference process are calculated at the pixel level, from the highfrequency components of the terrain depth images. The extracted features are then transformed using Multiple Discriminant Analysis (MDA), with the Bhattacharyya distance used in terrain classification for the different typical terrain types; grass, artificial turf, gravel, tile and concrete. The effects of relevant parameters, such as the size of the region of interest, on the performance of the approach (i.e. classification, boundary detection and computation time) are also investigated. A classification accuracy of 85.1% was achieved using non-semantic range data with a confidence threshold of 40%.