Automatic tracking is an important basis for abnormal behavior management and disease prediction of livestock. In commercial farms, the use of surveillance video to track and monitor dairy goats is conducive to improving production efficiency and commercial welfare. In this paper, an algorithm-based on Siamese strategy is presented for the automated tracking of a single dairy goat in the surveillance video. The Dairy Goat Dataset (DG-dataset) containing 200 dairy goat motion videos with a total of 161,000 frames of images randomly collected from the farm was created, and the ImageNet VID, Youtube-BB, and GOT-10 k were used for training. First, the proposed tracker named SiamBNAN employs an effective and highly modular backbone network constructed by the Multi-Convolution Residual Blocks (MCRBs) and Down-sampling Multi-Convolution Residual Blocks (D-MCRBs). The MCRBs and D-MCRBs replace the original square convolution kernel with a kernel augmented by asymmetric convolution kernels and perform a set of group convolutions. Finally, the Regional Proposal Network (RPN) is used for foreground-background classification and proposal regression. The experimental results show that this algorithm outperforms SiamFC, SiamRPN, and SiamRPN + in terms of both Expected Average Overlap (EAO), Robustness (R), Success Rate (Succ), and Precision (Prec) on the DG-dataset. The SiamBNAN runs at 70 fps with low computing space requirements showing that it is effective and could be used for monitoring the behavior of dairy goats in the real farming scene.

自动跟踪是牲畜异常行为管理和疾病预测的重要依据。在商品农场，利用监控视频对奶山羊进行跟踪监控，有利于提高生产效率和商业福利。在本文中，提出了一种基于 Siamese 策略的算法，用于在监控视频中自动跟踪单个奶山羊。创建了奶山羊数据集（DG-dataset），其中包含 200 个奶山羊运动视频，共有 161,000 帧从农场随机收集的图像，并使用 ImageNet VID、Youtube-BB 和 GOT-10 k 进行训练。第一的，提议的名为 SiamBNAN 的跟踪器采用由多卷积残差块 (MCRB) 和下采样多卷积残差块 (D-MCRB) 构建的有效且高度模块化的骨干网络。MCRBs 和 D-MCRBs 用非对称卷积核增强的核替换原来的方形卷积核，并执行一组组卷积。最后，区域提案网络（RPN）用于前景-背景分类和提案回归。实验结果表明，该算法在 DG 数据集上的预期平均重叠 (EAO)、鲁棒性 (R)、成功率 (Succ) 和精度 (Prec) 方面均优于 SiamFC、SiamRPN 和 SiamRPN+。SiamBNAN 运行在 70 MCRBs 和 D-MCRBs 用非对称卷积核增强的核替换原来的方形卷积核，并执行一组组卷积。最后，区域提案网络（RPN）用于前景-背景分类和提案回归。实验结果表明，该算法在 DG 数据集上的预期平均重叠 (EAO)、鲁棒性 (R)、成功率 (Succ) 和精度 (Prec) 方面均优于 SiamFC、SiamRPN 和 SiamRPN+。SiamBNAN 运行在 70 MCRBs 和 D-MCRBs 用非对称卷积核增强的核替换原来的方形卷积核，并执行一组组卷积。最后，区域提案网络（RPN）用于前景-背景分类和提案回归。实验结果表明，该算法在 DG 数据集上的预期平均重叠 (EAO)、鲁棒性 (R)、成功率 (Succ) 和精度 (Prec) 方面均优于 SiamFC、SiamRPN 和 SiamRPN+。SiamBNAN 运行在 70 DG 数据集上的稳健性 (R)、成功率 (Succ) 和精度 (Prec)。SiamBNAN 运行在 70 DG 数据集上的稳健性 (R)、成功率 (Succ) 和精度 (Prec)。SiamBNAN 运行在 70fps计算空间要求低，表明它是有效的，可用于监测真实养殖场景中奶山羊的行为。