Low-overlap Crop Point Cloud Registration Method based on Improved GeoTransformer with Normal Vectors and Gaussian Curvature

Crop point cloud registration holds significant importance in measuring and monitoring crop phenotypic features. However, the 3D representation of crops differs from that of indoor or urban scenes due to the pronounced curvature of their trunks, branches, and other components, resulting in lower overlap rates between point clouds. To address this issue, we leverage the discernible normal vectors and curvature features of crops and propose a Gaussian curvature-based enhancement to GeoTransformer for aligning plant point clouds with low overlap rates. This enhancement encodes the geometric attributes of the point cloud. To assess the method's effectiveness, we conducted experiments using real-world data. The proposed method achieves RMSE errors ranging from only 1.7% to 3.8%, 6% to 28.7%, and 28.2% to 51% of those of ICP, Predator, and GeoTransformer, respectively, for low overlap rates (20%) and average values.