基于3D点云的母猪体质量多尺度特征提取与估算模型

Multi-scale Feature Extraction and Estimation Model for Sow Mass Based on 3D Point Cloud

摘要:

生猪体质量是养殖管理中监测生长发育的关键指标，传统接触式称量存在效率低、应激风险高等缺陷。现有非接触式测量方法的三维特征利用率不足、体质量估算精度受限，针对此问题，本文提出基于3D点云的母猪体质量多尺度特征提取与估算方法。首先，构建融合位置编码的Point Transformer分割模型，实现复杂背景下猪体背部点云分割;以分割后猪体点云为输入构建Point-PN4体质量估算模型，创新性提出四分支并行特征提取架构，通过差异化卷积核同步捕获多尺度几何特征，并采用跨分支特征融合策略增强局部与全局特征互补性。试验结果表明，Point Transformer猪体点云分割模型准确率达到98.89%，Point-PN4模型RMSE、MAPE和MAE分别为4.36 kg、1.89%和3.99 kg，较Point-PN模型低0.59 kg、0.39%和0.21 kg，模型特征提取能力显著增强，体质量估算精度显著提升。模型内存占用量为4.6 MB，单点云处理时间9.7 ms，实现了轻量化和高效化。本文提出的母猪体质量估算方法，为生猪养殖管理智能化、精准化提供了切实可行的技术支撑。

Abstract:

Live pig body mass is a key indicator for monitoring growth and development in breeding management. Traditional contact-based weighing methods suffer from low efficiency and high stress risks. Existing non-contact measurement methods inadequately utilize three-dimensional features, limiting the accuracy of body mass estimation. To address this issue, a multi-scale feature extraction and estimation method for sow body mass was proposed based on 3D point clouds. Firstly, a Point Transformer segmentation model incorporating positional encoding was constructed to achieve accurate pig back point cloud segmentation in complex backgrounds. Using the segmented point cloud as input, a Point-PN4 body mass estimation model was developed, which innovatively introduced a four-branch parallel feature extraction architecture. This architecture employed differentiated convolutional kernels to simultaneously capture multi-scale geometric features and adopted a cross-branch feature fusion strategy to enhance the complementary relationship between local and global features. Experimental results demonstrated that the point transformer segmentation model achieved an accuracy of 98.89%. The Point-PN4 model exhibited RMSE, MAPE, and MAE values of 4.36 kg, 1.89%, and 3.99 kg, respectively, outperforming the Point-PN model by reductions of 0.59 kg, 0.39%, and 0.21 kg. The model significantly enhanced feature extraction capability and improved estimation accuracy. With a model size of 4.6 MB and a processing time of 9.7 ms per point cloud, the method achieved lightweight and efficient performance. The proposed sow body mass estimation method provided a practical and effective technical solution for intelligent and precise pig breeding management.