TY - JOUR TI - Classifying ground terrain using multiveiw methods DO - https://doi.org/doi:10.7282/T3028TD4 PY - 2015 AB - In this thesis, we acquire outdoor and indoor ground terrain photos at multiple viewing angles in natural light as a means to discover fundamental differences between terrains due to reflectance variation in natural illumination. Ground terrain affects robot control algorithms but little attention has been given to automatically determining the ground terrain composition using vision-based recognition techniques. For example, autonomous cars that can understand the presence of ice and wet road surfaces can automatically determine potential hazards, this in turn allows the car to adjust its driving accordingly to keep passengers safe. We examine an existing classification method, Texton Boost, to set the basis for comparison to a new proposed method based on angular gradient differential illumination histograms for automatically determining and classifying ground terrain surfaces. A database of 50 ground covers, each imaged from 33 viewing angles including finely quantized angular directions for computing gradients, was acquired by a mobile robot platform for this work. A selection of terrains in the database include gravel, carpet, tile, snow, ice, pebbles, grass, leaves, concrete, and asphalt. Multiple outdoor samples also include wet and dry varieties. The database of 1,650 ground terrain images is made publicly available for research use. KW - Electrical and Computer Engineering KW - Terrain study KW - Ground cover plants LA - eng ER -