Uncrewed Ground Vehicle (UGV) LiDAR Data Enhancement for High-Precision and High-Throughput Phenotyping in Mechanized Fields

Over the past few decades, advances in proximal sensing uncrewed ground vehicles (UGVs) have provided researchers with vast opportunities to harness high-fidelity phenotyping information in seed breeding trials. These platforms feature a variety of sensors that are often directly georeferenced using a global navigation satellite system/inertial navigation system (GNSS/INS) for high absolute accuracy. However, one of the biggest challenges with using GNSS is the limited availability of satellite signals under plant canopy. The poor georeferencing accuracy arising from GNSS signal outage together with sensor noise leads to discrepancies within single/multi-mission data, hindering the high-precision aspect of the high-throughput (HT) phenotyping. This paper presents a UGV mapping system for proximal scanning of agricultural crops and proposes a technique to enhance the georeferencing accuracy degraded by under-canopy data acquisitions. This platform is equipped with a light detection and ranging (LiDAR) sensor and a fisheye camera that are georeferenced using an integrated GNSS/INS unit. The proposed technique begins by identifying geometric primitives pertaining to mechanized fields popular for seed breeding trials. These primitives originate from ground patches, plant rows, and individual plant stalks, extracted from one or more multi-track UGV datasets. The extracted features are then incorporated in an optimization framework that improves the precision of UGV point clouds. This data enhancing strategy simultaneously yields individual stalk information, which can be valuable for deriving plant's inner attributes such as leaf count and angle. The proposed workflow is demonstrated using two maize field datasets separated by a period of several weeks. The enhancement results in a reduction of feature fitting error for linear features from as much as 44 cm to 2 cm. Furthermore, the approach also improves the absolute accuracy of the resulting point clouds after the inclusion of a reference UAV point cloud. The improvement is shown through the reduction of the overall feature fitting error to within the sensor noise range of under 3 cm for the combined multi-platform dataset.