Autonomous Off-Road Navigation Using Near-Feature-Based World Knowledge Incorporation on the Example of Forest Path Detection

Abstract

This paper presents a novel approach for robust off-road navigation based on deep convolutional neural networks which are combined with OpenStreetMap data to perform a forest path-based local localization approach. Corresponding near features are used to integrate navigation relevant world knowledge into a local multi-feature map. A behavior-based controller adapts the robot’s trajectory based on available features and its detection quality. The approach was tested in the Rhineland-Palatinate forest. Different forest way detection setups were evaluated and are discussed in detail. Additionally, the autonomous mobile robot GatorX855D followed a forest trail using the resulting multi-feature map.