Place recognition using regional point descriptors for 3D mapping

Abstract

In order to operate in unstructured outdoor environments, globally consistent 3D maps are often required. In the absence of a absolute position sensor such as GPS or modifications to the environment, the ability to recognize previously observed locations is necessary to identify loop closures. Regional point or keypoint descriptors are a way to encode the structure within a small local region as a fixedsized vector, though individually do not include enough context to fully identify a previously seen place. Multiple queries to a database of descriptor vectors can quickly identify similar features, and places can be recognized from a consistent set of descriptor matches.We investigate the problem of designing informative keypoint descriptors for 3D laser maps. Several models are considered and evaluated, with a particular focus on the optimal descriptor scale and keypoint sampling density. The approach is evaluated on 3D laser point cloud data collected from a vehicle driving in unstructured off-road environments. Consistent 3D maps constructed from this data without assistance from any other sensor (such as wheel encoders, GPS, or IMU) demonstrate the effectiveness of our approach.