Abstract: n this study, an improved FairMOT model was proposed to tackle the challenges encountered in multi‑object tracking of group‑raised pigs, such as issues related to similar appearances, mutual occlusions, and dynamic interactions, which led to errors in identity recognition, missed detections and false detections. An EMA attention mechanism was embedded within the backbone network to optimize the feature maps obtained during the down‑sampling stage and to enhance the representation of pig position features. Furthermore, a discrimination feature learning network was introduced, aiming to strengthen the fine‑grained differences in appearance features among different pigs, thereby improving individual discrimination. Additionally, the model adopted a three‑phase strategy for feature matching, IoU matching, and occlusion recovery matching, which enhanced the tracking accuracy. The test results in self‑made datasets demonstrated that the improved FairMOT model excelled in key metrics such as HOTA, MOTA, MOTP, and IDF1, with average scores reaching 85.87%、96.53%、96.07% and 94.82% respectively. These scores significantly outperformed those of the original FairMOT model and other five trackers. The model also exhibited high accuracy and stability under various lighting and occlusion conditions, proving its effectiveness and practicality in complex environments.

Key words: group?raised pigs, animal detection, pig multi?object tracking, behavior recognition, occlusion matching

摘要： 为解决群养生猪多目标跟踪中的挑战，如猪只外观相似、互相遮挡和动态交互等导致的身份识别错误、跟踪不连续问题，提出一种改进型FairMOT模型。该模型在主干网络中嵌入EMA注意力机制，优化下采样阶段的特征图，增强猪只位置特征的表达能力。同时，引入区分特征学习网络，通过加强不同猪只之间外观特征的细粒度差异，提高个体间的区分度。此外，模型采用特征匹配、IoU匹配和遮挡恢复匹配的三阶段策略，以增强跟踪的准确性。测试结果显示，改进FairMOT在高阶跟踪精度HOTA、多目标跟踪准确率MOTA、多目标定位精度MOTP、识别F1得分等关键指标上表现卓越，在自制数据集的平均得分分别达到85.87%、96.53%、96.07%和94.82%，明显优于原始FairMOT和其他五种跟踪器。且在不同光照和遮挡条件下，其展现出高准确性和稳定性，证明在复杂环境中的有效性和实用性。

关键词: 群养生猪, 动物检测, 猪只多目标跟踪, 行为识别, 遮挡匹配