Kyaw Pyae Phyo, Tin Pyke, Aikawa Masaru, Kobayashi Ikuo, Zin Thi Thi
Graduate School of Engineering, University of Miyazaki, Miyazaki, 889-2192, Japan.
Organization for Learning and Student Development, University of Miyazaki, Miyazaki, 889-2192, Japan.
Sci Rep. 2025 Jul 1;15(1):21938. doi: 10.1038/s41598-025-08277-8.
Cattle health monitoring and management systems are essential for farmers and veterinarians, as traditional manual health checks can be time-consuming and labor-intensive. A critical aspect of such systems is accurate cattle identification, which enables effective health monitoring. Existing 2D vision-based identification methods have demonstrated promising results; however, their performance is often compromised by environmental factors, variations in cattle texture, and noise. Moreover, these approaches require model retraining to recognize newly introduced cattle, limiting their adaptability in dynamic farm environments. To overcome these challenges, this study presents a novel cattle identification system based on color point clouds captured using RGB-D cameras. The proposed approach employs a hybrid detection method that first applies a 2D depth image detection model before converting the detected region into a color point cloud, allowing for robust feature extraction. A customized lightweight tracking approach is implemented, leveraging Intersection over Union (IoU)-based bounding box matching and mask size analysis to consistently track individual cattle across frames. The identification framework is built upon a hybrid PointNet ++ Siamese Network trained with a triplet loss function, ensuring the extraction of discriminative features for accurate cattle identification. By comparing extracted features against a pre-stored database, the system successfully predicts cattle IDs without requiring model retraining. The proposed method was evaluated on a dataset consisting predominantly of Holstein cow along with a few Jersey cows, achieving an average identification accuracy of 99.55% over a 13-day testing period. Notably, the system can successfully detect and identify unknown cattle without requiring model retraining. This cattle identification research aims to integrate the comprehensive cattle health monitoring system, encompassing lameness detection, body condition score evaluation, and weight estimation, all based on point cloud data and deep learning techniques.
牛健康监测与管理系统对农民和兽医来说至关重要，因为传统的手动健康检查既耗时又费力。此类系统的一个关键方面是准确的牛只识别，这有助于进行有效的健康监测。现有的基于二维视觉的识别方法已取得了不错的成果；然而，它们的性能常常受到环境因素、牛只纹理变化和噪声的影响。此外，这些方法需要重新训练模型以识别新引入的牛只，这限制了它们在动态农场环境中的适应性。为了克服这些挑战，本研究提出了一种基于使用RGB-D相机捕获的彩色点云的新型牛只识别系统。所提出的方法采用了一种混合检测方法，该方法首先应用二维深度图像检测模型，然后将检测到的区域转换为彩色点云，以便进行强大的特征提取。实施了一种定制的轻量级跟踪方法，利用基于交并比（IoU）的边界框匹配和掩码大小分析来在各帧之间持续跟踪单个牛只。识别框架基于使用三元组损失函数训练的混合PointNet ++连体网络构建，确保提取用于准确牛只识别的判别特征。通过将提取的特征与预先存储的数据库进行比较，该系统无需重新训练模型就能成功预测牛只ID。所提出的方法在一个主要由荷斯坦奶牛以及少数泽西奶牛组成的数据集上进行了评估，在为期13天的测试期内平均识别准确率达到了99.55%。值得注意的是，该系统无需重新训练模型就能成功检测和识别未知牛只。这项牛只识别研究旨在整合全面的牛健康监测系统，该系统涵盖跛行检测、体况评分评估和体重估计，所有这些都基于点云数据和深度学习技术。
