Global Map Generation Using Local Feature Grid Maps for Autonomous Vehicles in Frequently Changing Off-road Environments

Robust navigation in unstructured off-road environments is challenging for autonomous vehicles, as no paths or landmarks often exist, and classification algorithms are prone to frequent environmental changes such as light and reflections or foliage, dust, and vegetation. Humans, however, can navigate in such environments, mainly when they drive the same route frequently. They remember their environment and rely on their stored knowledge. State-of-the-art global maps like OSM or gmaps are helpful for planning but often only contain qualitative data for structured environments like cities. The proposed system follows a similar approach by mapping the perceived environment and the robot’s position on a global map using local grid maps generated from sensor readings. It is shown that the global map can be used for local navigation by transforming memorized data back to a local view for path planning purposes, compensating for disturbances, and increasing the system’s robustness. The approach was tested on a landfill environment with the off-road capable truck Unimog U5023.