基于SGPointNet++模型的奶牛点云分割与表型自动测定系统设计

Design of Automatic Determination System for Point Cloud Segmentation and Morphology of Dairy Cows Based on SGPointNet++ Model

摘要:

针对奶牛体尺人工测量工作量大、容易引起应激反应等问题，利用奶牛点云的三维重建以及点云分割技术，提出改进的点云分割模型并实现奶牛体尺数据的自动计算。本文以中国华西牛为研究对象，通过奶牛三维点云采集系统，采集115头奶牛的212组点云数据；采用Super-4pcs算法配准、进行空间直通滤波、基于邻域的离群点滤波完成奶牛点云的三维重建；基于PointNet++点云分割算法，结合SGE空间分组增强模块，提出改进的SGPointNet++模型，用于奶牛点云分割处理，进一步测量了体高、胸围、腹围、十字部高4个体尺数据。实验结果表明，SGPointNet++模型在测试集上分割平均交并比为81.87%，相较于PointNet、ASSANet、PointNeXt、PointNet++模型分别高27.82、1.55、1.19、1.07个百分点；体尺测量对于体高、胸围、腹围、十字部高平均绝对百分比误差分别为2.38%、3.05%、1.32%、1.69%，表明该方法可用于奶牛体尺测量，在降低工作量的同时保证了计算精度，为动物表型数据连续测定提供方法支撑，为分割和体尺计算模型改进提供技术参考。

Abstract:

Aiming to address the issues of heavy manual workload and the potential for inducing stress responses in dairy cow during traditional body measurement, a three-dimensional (3D) reconstruction and point cloud segmentation approach was proposed. This approach utilized an improved point cloud segmentation model for the automatic calculation of body measurements in cow. The research focused on Chinese Huaxi cow, and 212 sets of point cloud data from 115 dairy cows were collected using a 3D point cloud acquisition system. The Super-4pcs algorithm was used for point cloud registration, followed by spatial pass-through filtering and neighborhood-based outlier filtering to complete the 3D reconstruction of the cow’s point cloud. The PointNet++ point cloud segmentation algorithm, combined with the spatial grouping enhancement (SGE) module, was used to propose the improved SGPointNet++ model for point cloud segmentation. The segmentation results were then used to measure four body parameters: height, chest girth, abdominal girth, and withers height. The experimental results showed that the mean intersection over union (MIoU) for segmentation using the SGPointNet++ model on the test set was 81.87%, which was 27.82 percentage points, 1.55 percentage points, 1.19 percentage points, and 1.07 percentage points higher than that of PointNet, ASSANet, PointNeXt, and PointNet++, respectively. The average absolute percentage errors for body measurements were 2.38%, 3.05%, 1.32%, and 1.69% for body height, chest girth, abdominal girth, and withers height, respectively. These results indicated that this method can be used for dairy cow body measurement, reducing workload while ensuring computational accuracy. It provided a methodological foundation for continuous animal phenotype measurement and offered technical insights for further improvements in segmentation and body measurement calculation models.