Characterization of social cohesion status of pre-weaning piglets based on lightweight pose estimation

The social experiences of piglets during lactation have a lasting impact on their behavior after weaning. However, previous studies evaluating the social cohesion of lactating piglets mainly relied on manual measurements, resulting in a time-consuming and labor-intensive process. Furthermore, these studies primarily focused on assessing social cohesion within piglet groups, often overlooking the analysis of outliers. To overcome these limitations, we propose a novel two-stage method that aims to automate the estimation of social cohesion in pre-weaning piglets while simultaneously detecting outliers. In the first stage, we developed an improved YOLOv8 pose estimation model. This lightweight model enables rapid and accurate determination of individual piglets’ positions and orientations. The subsequent stage involves using enhanced outlier detection and clustering algorithms to detect inlier groups and outlier piglets. Benefitting from both stages, the social cohesion status of piglets can be characterized at different levels based on their spatial proximity and heading direction, including the pen level, inlier group level, and outlier individual level. Regarding the experimental results, our pose estimation model achieves a mean average precision (mAP) of 86.1% and 95.3% for object detection and pose estimation, surpassing the performance of the baseline YOLOv8 and other state-of-the-art models. Additionally, our method has successfully been applied to analyze a three-hour video within a farrowing pen, providing an intuitive understanding and characterization of piglet social status over time. This application demonstrates the potential of our method to comprehensively understand the intricate social dynamics and relationships among pre-weaning piglets. © 2024 Elsevier B.V.