Computationally inexpensive labeling of appearance based navigable terrain for autonomous rovers

In this paper we describe a monocular vision based method to learn navigable terrain for autonomous rover navigation. A self-supervised learning mechanism adjusts the surface appearance model using monocular image sequences. We propose a computationally inexpensive approach to labeling of the ground plane pixels using a reactive pre-filter. An active window, centered at the current position of the robot, implements the pre-filter based on majority voting. The selection criterion employs only a comparison, addition, and bit shifts using integer arithmetic. Hence, the scoring mechanism is directly implementable using an integer data path, associated with a reduced area overhead for resource constrained rover applications. The labeled navigable pixels are used as training data. The learning algorithm uses a mixture of Gaussians to model the terrain. We present empirical results on heterogeneous obstacle field configurations and varying terrain types.