A near-to-far learning framework for terrain characterization using an aerial / ground-vehicle tea

In this thesis, a novel framework for adaptive terrain characterization of untraversed far terrain in a natural outdoor setting is presented. The system learns the association between visual appearance of different terrain and the proprioceptive characteristics of that terrain in a self-supervised framework. The proprioceptive characteristics of the terrain are acquired by inertial sensors recording measurements of one second traversals that are mapped into the frequency domain and later through a clustering technique classified into discrete proprioceptive classes. Later, these labels are used as training inputs to the adaptive visual classifier. The visual classifier uses images captured by an aerial vehicle scouting ahead of the ground vehicle and extracts local and global descriptors from image patches. An incremental SVM is utilized on the set of images and training sets as they are grabbed sequentially. The framework proposed in this thesis has been experimentally validated in an outdoor environment. We compare the results of the adaptive approach with the offline a priori classification approach and yield an average 12% increase in accuracy results on outdoor settings. The adaptive classifier gradually learns the association between characteristics and visual features of new terrain interactions and modifies the decision boundaries.