Deriving precise orchard maps for unmanned ground vehicles from UAV images

Abstract

Mapping and environment representation are two of the main challenges in agricultural robotics and are vital to navigation tasks like localisation and path planning. In this work, we present a new method that enables the offline creation of orchard maps for unmanned ground vehicles based on unmanned aerial vehicle imagery. We employ photogrammetry to generate high-resolution 3D point clouds from aerial images. A cloth simulation filter is then used to classify ground and off-ground points. In order to obtain detailed probabilistic occupancy grid maps, per cell statistics are evaluated. First results show promising performance when compared to ground truth positions of orchard bushes and manual labelling.