Document type: scientific article published in Smart Agricultural Technology
Authors: Guo, Clémence A.E.M. Orsini, Patrick P.J.H. Langenhuizen, Yue Sun, Shoujun Huo, Lisette E. van der Zande, Inonge Reimert, J. Elizabeth Bolhuis, Piter Bijma, Peter H.N. de With
Preview: Damaging behaviors in pigs, such as tail biting and ear biting, compromise animal welfare and farm productivity. Continuous monitoring of these behaviors is essential to intervene before escalation, gain insights into underlying causes, and develop breeding programs to select pigs with lower genetic propensity for such behaviors. However, manual observations are impractical at a large scale. To address this challenge, we propose a video-based behavior recognition model that facilitates the automated monitoring of individual pigs. Two state-of-the-art video-based methods are investigated: SlowFast and Improved Multiscale Vision Transformers (MViTv2) for recognizing tail and ear biting in pigs, by exploiting spatiotemporal domain features. Data are collected on a commercial pig farm. In total, 532 tail-biting events (63,815 frames) and 750 ear-biting events (78,132 frames) are annotated across seven pens of tail-docked pigs. Tail biting and ear biting are defined as nibbling, sucking, chewing, or biting the tail or the ear of a pen mate. The best-performing method is based on the MViTv2-S model, which enables efficient spatiotemporal modeling. The detection accuracies obtained for tail and ear biting are 72.22% and 72.37%, respectively. An important and novel aspect to our knowledge is that for the first time, behavior detection is developed without a posture requirement on the biter or the victim. The conducted experiments demonstrate the feasibility of computer-vision-based models for the recognition of damaging behaviors on commercial pig farms. This study is a crucial step towards the development of an automated early-warning approach and breeding programs to reduce tail biting and ear biting.
