Enhancing Supervised Terrain Classification with Predictive Unsupervised Learning

摘要: This paper describes a method for classifying the traversability of terrain by combining unsupervised learning color models that predict scene geometry with supervised relationship between geometric features and traversability. A neural network is trained offline on hand-labeled computed from stereo data. An online process learns association geometry, enabling robot to assess regions which there little range information estimating passing this network. continuous extremely rapid, allows quick adaptations different lighting conditions changes. The sensitivity judgment further adjusted feedback robot’s bumper. Terrain assessments classifier are merged pure classifications in an occupancy grid computing intersection ray associated pixel ground plane We present results DARPA-conducted tests demonstrate its effectiveness variety outdoor environments.