Terrain Classification using Depth Texture Features

For Autonomous Ground Vehicles, modelling of the terrain is important for predicting future behaviours and control options. Current terrain modelling is able to identify geometric hazards but has a limited ability to identify complex terrain characteristics. This paper presents a method for segmenting and classifying terrain types based on range data to establish a geometric model of the terrain that maps terrain types to friction coefficients. Segmentation of the depth data is performed by a Piece-Wise Multi-Linear surface approximation, this enables the separation of the dominant surface geometry from the surface material characteristics. Depth texture features that are extracted in this process are used to classify the terrain types, specifically grass, road, dirt, rocks, gravel and sand. Using a Weighted Majority Voting with Dominance classifier, a depth texture only classification accuracy rate of 94.2% was achieved for the different terrain types, with 98.9% for depth and colour descriptors and 92.3% for colour only descriptors.