Abstract
In pig farming, multi-object tracking (MOT) algorithms are effective tools for identifying individual pigs and monitoring their health, which enhances management efficiency and intelligence. However, due to the considerable variation in breeding environments across different pig farms, existing models often struggle to perform well in unfamiliar settings. To enhance the model's generalization in diverse tracking scenarios, we have innovatively proposed the SDGTrack method. This method improves tracking performance across various farming environments by enhancing the model's adaptability to different domains and integrating an optimized tracking strategy, significantly increasing the generalization of group pig tracking technology across different scenarios. To comprehensively evaluate the potential of the SDGTrack method, we constructed a multi-scenario dataset that includes both public and private data, spanning ten distinct pig farming environments. We only used a portion of the daytime scenes as the training set, while the remaining daytime and nighttime scenes were used as the validation set for evaluation. The experimental results demonstrate that SDGTrack achieved a MOTA score of 80.9%, an IDSW of 24, and an IDF1 score of 85.1% across various scenarios. Compared to the original CSTrack method, SDGTrack improved the MOTA and IDF1 scores by 16.7% and 33.3%, respectively, while significantly reducing the number of ID switches by 94.6%. These findings indicate that SDGTrack offers robust tracking capabilities in previously unseen farming environments, providing a strong technical foundation for monitoring pigs in different settings.