Using Binary Classifiers to Augment Stereo Vision for Enhanced Autonomous Robot Navigation

Autonomous robot navigation in unstructured outdoor environments is a challenging area of active research. At the core of this navigation task lies the concept of identifying safe, traversable paths which allow the robot to progress toward a goal. Stereo vision is frequently exploited for autonomous navigation, but has limitations in terms of its density and accuracy in the far field. This paper describes image classification techniques which augment near field stereo to identify safe terrain and obstacles in the far field. Machine Learning classification techniques using appearance-based features appear well suited to the task of far-field obstacle detection, where stereo vision fails. In particular, binary classifiers are appropriate for this task and have performance characteristics suitable for real-time navigation systems. In this paper, we examine the use of stereo vision to identify obstacles and safe terrain in the near field, then using the appearance of these identified regions from the image to classify the remaining far field regions. We rigorously evaluate five binary classifiers as applied to the problem for logged image and navigation data and report on their performance. We also perform live experiments on a DARPA LAGR robot and show that the use of image classification techniques to augment stereo vision results in an enhanced navigational capability in the far field.