基于机器视觉的舍饲羊只行为识别方法

Housed Sheep Behavior Recognition Method Based on Machine Vision

摘要:

在规模化绵羊养殖场中，畜禽的行为特征能够有效反映其健康状况及环境适应能力。针对传统舍饲羊只行为识别方法在羊群密度变化条件下存在的监测效率低、识别精度不足等问题，提出了一种基于改进YOLO 11n模型的舍饲羊只行为识别方法。在羊圈斜上方安装2D摄像机，采集羊群的视频数据，并构建包含站立、进食、饮水和休息4种行为的舍饲羊行为数据集。在YOLO 11n模型基础上，结合CARAFE（Content-aware reassembly of features）上采样结构，并引入高效多尺度注意力机制EMA（Efficient multi-scale attention）与动态检测头DyHead（Dynamic feature learning for head detection）构成YOLO-CFED模型，提升羊只行为检测的特征提取与识别能力。结果表明，相较原YOLO 11n模型，改进YOLO-CFED模型在自建数据集上的性能提升显著：识别精确率（Precision）为95.6%（提升1.2个百分点）、召回率（Recall）为93%（提升0.4个百分点）、mAP@0.5为94%（提升0.3个百分点）、mAP@0.5:0.95为82.4%（提升1.3个百分点）、F1值为93.4%（提升0.9个百分点）。该方法能够有效识别羊只4种主要行为，为实现羊只行为智能化监测与健康管理提供了有力技术支持。

Abstract:

In large-scale sheep farming operations, the behavioral characteristics of livestock serve as effective indicators of their health status and environmental adaptability. Aiming to address the challenges of low monitoring efficiency and insufficient recognition accuracy inherent in traditional methods for housed sheep behavior recognition, particularly under varying flock densities, a novel behavior recognition method was proposed based on an improved YOLO 11n model. Initially, 2D cameras were installed diagonally above the sheep enclosures to capture video data of the flock, from which a housed sheep behavior dataset was constructed, comprising four distinct behaviors: standing, feeding, drinking, and resting. Subsequently, the YOLO 11n model was enhanced by integrating the content-aware reassembly of features (CARAFE) upsampling structure, the efficient multi-scale attention (EMA) mechanism, and the dynamic head (DyHead) framework, resulting in the proposed YOLO-CFED model. This was designed to improve the feature extraction and recognition capabilities for sheep behavior detection. The experimental results showed that the improved YOLO-CFED model significantly improved its performance on self built datasets: the recognition precision reached 95.6%, the recall reached 93%, an mAP@0.5 reached 94%, an mAP@0.5:0.95 reached 82.4% and an F1 score of 93.4%, all indicators were superior to that of the original YOLO 11n model. The proposed method effectively identified the four primary behaviors of sheep, thereby offering robust technical support for the implementation of intelligent behavioral monitoring and health management in sheep farming.