Wutke Martin, Debiasi Damiano, Tomar Shobhana, Probst Jeanette, Kemper Nicole, Gevers Kai, Lieboldt Marc-Alexander, Traulsen Imke
Institute of Animal Breeding and Husbandry, Faculty of Agricultural and Nutritional Sciences, Christian-Albrechts-University Kiel, Kiel, 24118, Germany.
Institute for Animal Hygiene, Animal Welfare and Farm Animal Behavior (ITTN), University of Veterinary Medicine Hannover, Foundation, Hannover, Germany.
Sci Rep. 2025 Jun 20;15(1):20153. doi: 10.1038/s41598-025-05283-8.
Reliable animal identification in livestock husbandry is essential for various applications, including behavioral monitoring, welfare assessment, and the analysis of social structures. Although recent advancements in deep learning models have improved animal identification using biometric markers, their applicability remains limited for species without distinctive traits like pigs. Consequently, synthetic features such as ear tags have become widely adopted. However, challenges such as poor lighting conditions and the complexity of ear tag coding continue to restrict the effectiveness of Computer Vision and Deep Learning techniques. In this study, we introduce a robust, lighting-invariant method for individual pig identification that leverages commercially available ear tags within a sequential detection pipeline. Our approach employs four object detection models in succession to detect pigs, localize ear tags, perform rotation correction via pin detection, and recognize digits, ultimately generating a reliable ID proposal. In a first evaluation stage, we assessed the performance of each model independently, achieving a mAP0.95 value of 0.970, 0.979, 0.974 and 0.979 for the pig detection, ear tag detection, pin detection and ID classification model, respectively. In addition, our method was further evaluated in two different camera environments to assess its performance in both familiar and unfamiliar conditions. The results demonstrate that the proposed approach achieves a very high precision of 0.996 in a familiar top-down camera scenario and maintained a strong generalization performance in an unfamiliar, close-up setup with a precision of 0.913 and a recall of 0.903. Furthermore, by publicly proposing three custom datasets for ear tag, pin, and digit detection, we aim to support reproducibility and further research in automated animal identification for precision livestock farming. The findings of this study demonstrate the effectiveness of ID-based animal identification and the proposed method could be integrated within advanced multi-object tracking systems to enable continuous animal observation and for monitoring specific target areas, thereby significantly enhancing overall livestock management systems.
在畜牧业中，可靠的动物识别对于各种应用至关重要，包括行为监测、福利评估和社会结构分析。尽管深度学习模型的最新进展改进了使用生物特征标记的动物识别，但对于像猪这样没有独特特征的物种，其适用性仍然有限。因此，诸如耳标的合成特征已被广泛采用。然而，诸如光照条件差和耳标编码复杂性等挑战继续限制计算机视觉和深度学习技术的有效性。在本研究中，我们引入了一种强大的、光照不变的个体猪识别方法，该方法在顺序检测管道中利用市售耳标。我们的方法依次使用四个目标检测模型来检测猪、定位耳标、通过引脚检测进行旋转校正以及识别数字，最终生成可靠的ID提议。在第一个评估阶段，我们独立评估了每个模型的性能，猪检测、耳标检测、引脚检测和ID分类模型的mAP0.95值分别为0.970、0.979、0.974和0.979。此外，我们的方法在两种不同的相机环境中进一步评估，以评估其在熟悉和不熟悉条件下的性能。结果表明，所提出的方法在熟悉的自上而下相机场景中实现了0.996的非常高的精度，并且在不熟悉的特写设置中保持了强大的泛化性能，精度为0.913，召回率为0.903。此外，通过公开提出三个用于耳标、引脚和数字检测的自定义数据集，我们旨在支持精确畜牧业中自动动物识别的可重复性和进一步研究。本研究的结果证明了基于ID的动物识别的有效性，并且所提出的方法可以集成到先进的多目标跟踪系统中，以实现连续的动物观察和监测特定目标区域，从而显著增强整体畜牧业管理系统。
