Authors

Abstract

As the pig farming industry evolves towards intelligence and large-scale operations, multi-object tracking technology plays a pivotal role in enhancing farming efficiency and optimizing health management. Addressing the issues of trajectory interruption and jitter caused by frequent occlusion of targets, significant scale variations, rapid motion, and sudden turns in large-scale breeding scenarios, an improved ByteTrack method is presented. This method employs YOLOv8 as the detector and enhanced the ByteTrack framework. By incorporating an Exponential Moving Average (EMA) trajectory smoothing module, it optimizes the trajectory jitter issues encountered by traditional Kalman filtering in fast-moving or sudden turning scenarios, resulting in smoother and more continuous tracking trajectories. Experimental results on a real-world livestock farm video dataset demonstrate that the proposed method achieves improvements across several key evaluation metrics. Specifically, the Multi-Object Tracking Accuracy (MOTA) increased by 1.61%, Identification F1 Score (IDF1) improved by 2.55%, and the Recall rose by 9.68%. The results demonstrate that the proposed method can effectively track pigs in complex farming environments with frequent occlusions, large scale variations, and abrupt motion, providing reliable underlying data support for downstream applications, such as abnormal behavior identification and automated physical activity statistics in intelligent pig farming systems.

Abstract in Chinese

随着养猪业向智能化与规模化发展，多目标跟踪技术对提升养殖效率与优化健康管理具有关键作用。针对规模化养殖场景中目标频繁遮挡、尺度变化显著、快速运动及突然转向导致的轨迹中断与抖动问题，本文提出一种基于改进ByteTrack算法的生猪多目标追踪方法，旨在应对规模化养殖场景中目标频繁遮挡、尺度变化显著及轨迹中断等挑战，以提升养殖智能化管理效率。该方法以YOLOv8作为目标检测器，并在ByteTrack多目标追踪框架基础上，引入指数移动平均（EMA）轨迹平滑模块，有效缓解了传统卡尔曼滤波在目标快速运动或突然转向时出现的轨迹抖动问题，从而获得更为平滑、连续的追踪轨迹。在实际养殖场视频数据集上的实验结果表明，本方法在多项关键评价指标上均有显著提升，其中多目标跟踪准确率（MOTA）提升1.61%，身份一致性（IDF1）提升2.55%，召回率提升9.68%。该方法能够有效应对复杂场景下的目标追踪任务，通过缓解遮挡与尺度变化对追踪精度的影响，能够为后续生猪异常行为识别、运动量自动化统计等实际生产应用提供稳定的底层数据支撑。