Occlusion-Resistant Spatial Analysis of Pig Distribution Pattern in Farrowing Pens Using Center Clustering Network

The spatiotemporal distribution of animals in controlled housing systems is a good indicator of their physiological wellness and welfare. However, to monitor the pigs in farrowing pens for spatial distribution analysis, the existence of farrowing crates brings inevitable visual occlusions, leading to a need for occlusion-resistant computer vision methods. This study aims to use Center Clustering Network (CClusnet) to characterise pig distribution patterns in farrowing pens with closed and half-open crates. Three videos with crates closed and three with crates half-open were collected. A total of 4,600 images were extracted from the videos and labelled to train CClusnet, and the trained model was then used to analyse individual video frames. The model outputs, including centre points of individual piglets and semantic segmentation of the sow and piglets, were accumulated into piglet-position heatmaps (PPH) and bodily-space-usage heatmaps (BSUH) for spatial distribution analysis, respectively. The BSUH of sows revealed differences in sow space usage, e.g., the sow utilized 1.5 times more space in half-open crates than in closed crates in this study. In addition, the BSUH of sows showed different preferences in sow lying sides, e.g., five of the six sows had unbalanced lying side frequency. The BSUH of piglets demonstrated the most frequent area that piglets visited or stayed in, e.g., around the heat pad and the pen border. The PPH supplemented the missing information under occlusions, especially the suckling area of piglets. Our method could be further used for sow lying side preference analysis and thus precaution of lesions due to one-side lying.