UGV based multi-RGBD SLAM framework for high-throughput in-situ 3D phenotyping of field maize

Accurate 3D reconstruction of maize populations under real-world field conditions is a prerequisite for individual-level trait in-situ extraction. Conventional visual mapping techniques frequently encounter challenges in complex field environments, which are characterized by weak textures and significant occlusions. To address these limitations, we propose a novel field maize in-situ phenotyping system utilizing synchronized multi-view RGBD cameras mounted on an Unmanned Ground Vehicle (UGV) platform. Specifically, a fixed multi-camera system is first established to introduce cross-view geometric constraints. Then, a semantic-guided feature distribution regulation strategy is proposed to balance feature sampling between crop foreground regions and background areas to alleviate localization degradation caused by spatially imbalanced feature distributions. Then, multi-view observations are subsequently fused to generate a field-scale dense point cloud with high geometric consistency. Finally, an automated processing pipeline is developed by integrating point cloud density projection and three-dimensional topological analysis. This pipeline enables accurate instance segmentation of individual plants within densely planted maize populations and supports robust in-situ quantification of key morphological traits, including plant height and stem diameter. Field experiments conducted under diverse environmental conditions demonstrated a 94.55% accuracy in plant counting. For morphological traits, the pipeline estimated plant height with an R 2 of 0.85 and an RMSE of 41.29 mm. Stem diameter estimation along two orthogonal axes yields R 2 values of 0.79 and 0.76, with corresponding RMSEs of 1.21 mm and 1.03 mm, respectively. The results validate the feasibility and robustness of the proposed framework for three-dimensional reconstruction in complex agricultural environments, and demonstrate its potential as a reliable solution for high-throughput in-situ crop phenotyping.