Assessment of off-road traversability using linear discriminant analysis in dense disparity space

Abstract

The topographic variations and unstructured nature of cross-country environments pose a great deal of challenges to vision based autonomous navigation. Algorithms that search for geometric descriptions of such undulating terrains, usually suffer from the lack of viable assumptions. This paper presents an innovative obstacle-ground discrimination method for cross-country navigation using stereo vision. We describe well known methods such as planar ground approximation and "v-disparity" analysis using a common mathematical framework, and investigate their failure modes under off-road conditions. In order to mitigate these deficiencies we define a classifier, which is uniquely described by a family of loosely coupled linear decision boundaries, with optimized parameters under a set of soft constraints. Results for simulated and real world test data are presented, and compared with pre-existing obstacle detection algorithms.