2-D-Guided 3-D Girth Analysis for Non-Contact Estimation of Cattle Weight

초록

This paper proposes a non-contact cattle weight estimation system that integrates two-dimensional segmentation with three-dimensional point cloud refinement for application in farm environments. To address challenges associated with traditional manual labor-intensive methods, the proposed system segments top-view RGB images to generate torso masks, which are projected onto the 3D point cloud to prune floor and background points. Residual noise is subsequently filtered through Z-thresholding and density-based spatial clustering of applications with noise (DBSCAN). Interpretable geometric features are extracted from the refined dorsal surface. This approach introduces three new cross-sectional girths (fore, mid, and hind), estimated from partial arcs via RANSAC-based circle fitting. These features, along with straight-line length measurements and overall torso length, form the basis of the cattle physique model. Using an XGBoost regressor, the system achieved a mean absolute percentage error of 6.80% for 932 samples from 109 cattle collected at a farm weighing station. This result is comparable to a baseline trained on manual morphometric data and substantially outperforms a conventional end-to-end model applied to raw point clouds. The pipeline is lightweight and efficient, requiring fewer computational resources by utilizing a small set of key features. It can be implemented using a single overhead RGB-D sensor and provides interpretable predictors intended for practical use in on-farm settings. This study presents a method for integrating smart farming technologies into livestock management, aiming to support farm productivity and animal welfare.

키워드