A geodesic distance regression-based semantic keypoints detection method for pig point clouds and body size measurement

Abstract

Pig body size reflects its physical shape and growth development, making accurate non-contact body size measurement crucial for practical farming production. The point cloud-based non-contact body size measurement method provides an effective alternative to traditional manual measurement, with the key challenge being the accurate identification of measurement keypoints. Many recent studies have focused solely on point cloud slicing or segmentation to indirectly locate body size keypoints, while research on directly predicting keypoints from livestock point clouds remains scarce. Therefore, we propose a method for directly detecting point clouds keypoints based on geodesic distance regression, which enables efficient measurement of pig body size through these keypoints. This approach transforms the detection of semantic keypoints in point clouds into a regression problem of geodesic distances between points and keypoints through heatmaps. The improved PointNet++ encoder-decoder architecture is utilized to learn distances on the manifold, enabling efficient keypoint detection. The model can be viewed as outputting probability values for each point corresponding to various keypoints, with the point having the highest probability selected as the predicted keypoint. Experimental results demonstrate an average root mean square error (RMSE) of 4.115 cm across eight keypoint types. The derived pig body size parameters achieve mean absolute percentage errors (MAPE) of 2.83 % for body length, 5.33 % for body width, 2.84 % for body height, 3.73 % for rump circumference, 4.83 % for thoracic circumference, and 3.83 % for abdominal circumference. The proposed geodesic distance regression-based semantic keypoints detection method for pig point clouds enables automated, accurate, and robust body size measurements, demonstrating significant potential for widespread application.