Access to nest-building material in the preparturient period is beneficial for sows’ welfare. However, on slatted floors, long-stem forage can drop into the slurry and block the drainage system. As a compromise considering the needs of sows for access to adequate nest-building material, farrowing pens with slatted floors are equipped with dispensers (racks) accessible by sows. In this study, we developed a computer vision method to monitor the use of the racks with nest-building material. In total, 12 sows were included in the experiment from 5 days before farrowing to the end of farrowing. Hay rack use behaviors were labeled for all the sows, i.e., pulling hay, nose close to the rack, exploratory behavior, and bar biting. The object detection algorithm RetinaNet was used to extract centroids of parts of the sow’s body and the hay rack. Several feature variables were estimated from the centroids of detected parts of the sow’s body, and random forest was used for the classification of hay rack use behaviors. The model for the detection of pulling hay behavior had the best performance: 83.5% sensitivity, 98.7% specificity, and 98.6% accuracy. The distance between the sows’ nose and the hay rack was the most important feature variable, which indicated the importance of nose location for the recognition of behaviors in which pigs interact with other objects. The developed models could be applied for automated monitoring of the use of nest-building material in preparturient sows. Such monitoring might be especially important in sows housed on slatted floors.
Full Text Finder