Computer vision assisted phenotyping, early detection and genetic analysis of abnormal behavior in weaned piglets

The project will focus on developing automated algorithms to recognize damaging behaviors in pigs. Pigs will be continuously video recorded for 10 weeks post-weaning to label behaviors like ear and tail biting. This data will help train computer vision algorithms using skeleton-based behavior recognition, which is advantageous due to its focus on body language and efficiency in detecting specific actions. The study will use advanced algorithms for object detection and behavior recognition, starting with YOLOv10 for identifying pigs based on ear tags. The detected pigs will then be analyzed using ViTPose, a model known for high accuracy in pose estimation, to map their skeletons with 22 key points. The applied method is effective at learning spatiotemporal features and is robust against pose estimation errors. To facilitate the transfer of these algorithms to the Austrian pig industry, photorealistic models of pigs and their environments will be generated using synthetic data techniques, ensuring a robust training dataset. An advanced model will be implemented for tracking individual pigs within a group, using both visual detection and RFID tags for continuous identity updates. Modeling and estimation of genetic parameters for new behavioral traits, calculation of genetic and phenotypic correlations among them and with important production traits will be also done. This comprehensive approach aims to enhance behavior monitoring and improve welfare in pig farming.