Learning Long-range Terrain Perception for Autonomous Mobile Robots

Long-range terrain perception has a high value in performing efficient autonomous navigation andrisky intervention tasks for field robots, such as earlier recognition of hazards, better path planning, and higherspeeds. However, Stereo-based navigation systems can only perceive near-field terrain due to the nearsightednessof stereo vision. Many near-to-far learning methods, based on regions’ appearance features, are proposed topredict the far-field terrain. We proposed a statistical prediction framework to enhance long-range terrainperception for autonomous mobile robots. The main difference between our solution and other existing methods isthat our framework not only includes appearance features as its prediction basis, but also incorporates spatialrelationships between terrain regions in a principled way. The experiment results show that our frameworkoutperforms other existing approaches in terms of accuracy, robustness and adaptability to dynamicunstructured outdoor environments