@article{WANG2026, 
author = {Yanchao WANG and Zhixin HUA and Ming GAO and Kaiying WANG},
title = {Approach for detecting and tracking estrous ewes based on improved YOLOv11n and OSTrack},
year = {2026},
journal = {Transactions of the Chinese Society of Agricultural Engineering},
volume = {42},
number = {11},
pages = {79-88},
keywords = {smart livestock farming, deep learning, estrus detection, object tracking, automated monitoring},
url = {https://www.sciopen.com/article/10.11975/j.issn.1002-6819.202601270},
doi = {10.11975/j.issn.1002-6819.202601270},
abstract = {Large-scale and intensive sheep farming is ever-increasing in modern animal husbandry, particularly with the high demand for sheep products, such as meat, milk, and wool. It is often required to accurately identify estrous ewes for effective reproductive management. The optimal breeding window can be expected to capture for the high conception rates and overall reproductive efficiency. However, manual observation cannot fully meet the need of large-scale sheep farming in recent years, due to its labor-intensive and time-consuming nature, which is prone to human errors. There are the high miss rates, low efficiency, and difficulty in sustained surveillance with manual estrus detection. In this study, an optimal detection and tracking method was proposed for estrous ewes using an improved YOLOv11n and OSTrack framework. The vision solution was also developed under complex scenarios, such as high animal density, strong visual similarity among individuals, and frequent occlusions in real-world farms. Two key components were integrated in a complementary manner. (1) Standard convolutions in the YOLOv11n network were replaced with space-to-depth convolution (SPDConv). Complexity and computational load were reduced to preserve critical spatial information after downsampling, thus detecting fine-grained behavioral cues, such as mounting postures or tail raising. Unlike conventional pooling or strided convolutions, which reduced spatial resolution. Spatial pixel information of the SPDConv model was reorganized into channel dimensions, thereby maintaining structural fidelity without increasing parameters. (2) A Triplet Attention mechanism was incorporated into the neck of the network to jointly capture dependencies among channel, height, and width dimensions. Subtle postural differences were extracted and then distinguished, particularly in crowded scenes with the common occlusion and appearance ambiguity. Informative regions were obtained in three orthogonal views. The network was more sensitive to context and alignment, which was crucial for identifying estrus in visually homogeneous flocks. The optimal YOLOv11n detector was combined with the OSTrack single object tracker to form the complete YOLO-OSTrack framework. The initial high-confidence bounding box was provided for an ‘Estrus’ labeled ewe, which was then used to initialize the tracker. Subsequent frames were processed by OSTrack alone, enabling continuous and efficient localization without repeated detection calls. This tracking decoupling reduced inference latency for the temporal smoothness and identity consistency over long sequences. Experimental results also demonstrated the effectiveness of the approach. The improved YOLOv11n was achieved in an F1-score of 93.0%, an average precision (AP) of 98.0% for mounting behavior, and 93.4% for estrous ewes, thus representing gains of 1.1, 0.5, and 2.0 percentage points over the baseline model. A lightweight architecture exhibited with only 2.2 M parameters and a model size of 4.5 MB, with a computational cost of 5.6 G. There were reductions of 15.4%, 13.5%, and 12.5%, respectively, compared with the baseline. The model was highly suitable for deployment on the edge devices with limited memory and processing power. In tracking, OSTrack achieved a success rate (area under curve, AUC) of 85.1%, precision (P) of 87.0%, and norm precision of 96.1%, indicating its robustness against challenges, such as high-density sheep flocks, pose variations, high visual similarity among individuals, and partial occlusion. Overall, the YOLO-OSTRack framework can be expected to accurately detect and stably track estrous ewes over the long term in practical farming environments. Reliable technical support was offered for real-time monitoring, early warning, individualized precision management, and reproductive efficiency in modern intelligent livestock systems, thereby contributing to more sustainable and data-driven sheep farming.}
}