Far-Field Terrain Perception Using Max-Margin Markov Networks

Abstract

Far-field terrain perception plays an important role in performing outdoor robot navigation, such as earlier recognition of obstacles, efficient path planning. Stereo vision is an effective tool to detect obstacles in the near-field, but it cannot provide reliable information in the far-field, which may lead to suboptimal trajectories. This can be settled through the use of machine learning to accomplish near-to-farlearning, in which near-field terrain appearance features and stereo readings are used to train models able to predict far-field terrain. In this paper, we propose a near-to-far learning method using Max-Margin Markov Networks (M3N) to enhance long-range terrain perception for autonomous mobile robots. The method not only includes appearance features as its prediction basis, but also uses spatial relationships between adjacent parts. The experiment results show that our method outperforms other existing approaches.