An Improved Point Cloud Processing Framework for 3D Body Measurement of Beef Cattle

ABSTRACT Accurate and contactless measurement of cattle body dimensions is essential for phenotyping, breeding evaluation, and herd management. Traditional manual methods are labor‐intensive and stressful to animals, while 2D vision approaches cannot fully capture three‐dimensional morphology. This study proposes an improved 3D point cloud processing framework for beef cattle body measurement using multiview RGB‐D data acquired from three Kinect V2 sensors. The framework integrates an Adaptive Threshold RANSAC (AT‐RANSAC) method for robust ground plane segmentation, a multiscale curvature‐based feature‐preserving sampling strategy with adaptive thresholding, and a combined NDT–ICP registration pipeline for multiview point cloud fusion. Based on the reconstructed 3D model, anatomical landmarks are identified through curvature mutation analysis of dorsal contours, and a slicing‐based method is used to extract key body parameters, including body height, withers height, chest girth, abdominal girth, and body oblique length. Experiments conducted on 137 Chinese Yellow cattle (approximately 170,000 points per scan) show that the proposed sampling strategy significantly reduces geometric feature loss compared with random and voxel grid sampling, while maintaining measurement accuracy. Comparison with manual measurements demonstrates good agreement, with mean relative errors of 1.86% for body height, 2.0% for withers height, 3.78% for chest girth, 3.89% for abdominal girth, and 6.06% for body oblique length. The proposed framework provides an efficient and automated solution for large‐scale, low‐stress cattle phenotyping and body size measurement.