Automatic measurement method of sheep body size based on 3D reconstruction and point cloud segmentation

Sheep body measurement data can comprehensively reflect body size, structural characteristics, growth status, and developmental relationships between different body parts. Automatically obtaining sheep body size parameters represents a critical technological advancement towards digital intelligent livestock farming. Currently, computer vision-based livestock body measurement techniques have garnered significant research interest due to their non-contact and efficient nature. However, sheep’s collective behavior and highly flexible body postures present challenges to vision-based measurement methodologies. Addressing these challenges, this paper proposes an automatic body size measurement method for sheep utilizing 3D reconstruction and point cloud segmentation technologies. Three KinectV2 cameras were strategically positioned in the passageway leading to the activity field to capture multi-view sheep images. Through multi-view image alignment, a 3D reconstruction was achieved to reproduce the sheep’s spatial morphology. The PointNet++ deep learning model was trained to develop an automatic sheep body segmentation model. Based on local pose normalization, keypoints were automatically detected by utilizing morphometric features, then body size parameters were computed. Farm experiment results demonstrated measurement accuracy: average relative errors for wither height, chest width, rump height, rump width, body length, and chest girth were 1.67 %, 3.63 %, 1.14, 2.71 %, 3.57 %, and 3.71 %, respectively. Experimental validation demonstrates the proposed approach reduces hardware requirements for automated body‑size measurement while maintaining the accuracy of body measurements. This enhancement facilitates its broader adoption and improves animal welfare. Meanwhile, experimental outcomes revealed that irregular posture challenges the efficiency of automated body size measurements, highlighting the necessity of integrating pose estimation techniques to further enhance measurement precision in future research.