A Stable Long-Term Tracking Method for Group-Housed Pigs

Shibo Gao, Jinmeng Gong, Peipei Yang, Chao Liang, Linlin Huang

Abstract: In recent years, computer vision technologies have been increasingly applied to livestock farming for improving efficiency and reducing the labor force in surveillance. Tracking of group-housed pigs is an important task for monitoring the daily behaviors of pigs, which can be used to preliminarily evaluate the health status of pigs. Most researchers directly apply existing multi-object tracking algorithms to this task, but often suffer from tracking failures due to false detection, stacking, occlusion, video jamming, etc. It usually produces a lot of incorrect ID switches that are disastrous for follow-up tasks. In this paper, we propose a group-housed pigs tracking method that can achieve stable long-term tracking. As the identity and number of monitored pigs remain unchanged during a feeding period, we introduce a new object matching mechanism with a classifier, which avoids most incorrect ID switches and effectively improves the matching accuracy. Thus, our tracking method is more robust to complex posture variations of the pig and achieves stable long-term tracking. The experimental results on real videos captured in a pigs farm prove the effectiveness of our method.