Absolute localization of mobile robots in forest environments by correlating ground LiDAR to overhead imagery

Abstract

A method for the autonomous geolocation of ground vehicles in forest environments is presented. The method provides an estimate of the global horizontal position of a vehicle strictly based on finding a geometric match between a map of observed tree stems, scanned in 3D by Light Detection and Ranging (LiDAR) sensors onboard the vehicle, to another stem map generated from the structure of tree crowns analyzed from aerial orthoimagery of the forest canopy. Identification of stems from 3D data is achieved by using Support Vector Machine (SVM) classifiers and height above ground filters that separate ground points from vertical stem features. Identifications of stems from overhead imagery is achieved by calculating the centroids of tree crowns extracted using a watershed segmentation algorithm. Matching of the two stem maps is achieved by using a robust Iterative Closest Point (ICP) algorithm that determines the rotation and translation required to align both datasets. The final alignment is used to calculate the absolute location of the vehicle. The localization algorithm has been tested with real-world data and has been able to estimate vehicle geoposition with errors of less than 2.1 m RMS. It is noted that the algorithm's accuracy is currently limited by the accuracy and resolution of the aerial orthoimagery used. This method can be used in real-time as a complement to the Global Positioning System (GPS) and Simultaneous Localization and Mapping (SLAM) in areas where signal coverage is inadequate due to attenuation by the forest canopy, or due to intentional denied access. The method has two key properties that are significant: i) It does not require a priori knowledge of the area surrounding the robot. ii) Based on estimated vehicle state, it uses the geometry of detected tree stems as the only input to determine horizontal geoposition.