YOLOv5DA: an improved YOLOv5 model for pig posture detection in a herd environment

Abstract

Abstract Animal behavior is an important indicator for diagnosing diseases, and accurate posture detection is the foundation for analyzing animal behavior and emotional states, which can promote animal welfare. However, current methods for pig posture detection often suffer from problems of missed or false detection due to the aggregation, occlusion, and adhesion of pigs in a herd environment. This study proposed a new object detection model (YOLOv5DA) for pig posture detection based on YOLOv5s, which incorporates Mosaic9 data augmentation, deformable convolution, and adaptively spatial feature fusion, to effectively deal with missed or false detection in the case of mutual occlusion and bonding of pig bodies. The images of pigs were first collected and annotated, and a dataset was established. Then, by using the established dataset, an object detection model YOLOv5DA based on YOLOv5s was trained. Finally, the test results showed that YOLOv5DA could accurately identify the three postures of standing, prone lying, and side lying with an average precision (AP) of 99.4%, 99.1%, and 99.1%, respectively, and the performance is superior to that of mainstream object detection algorithms including Faster-RCNN, YOLOv4, YOLOv5, FCOS, and CenterNet. Compared with YOLOv5s, YOLOv5DA could effectively handle occlusion while increasing the mean precision (mAP) by 1.7% in complex scenarios, which reached about 86.8%. Overall, YOLOv5DA provides a highly accurate, effective, low-cost, and stress-free strategy for pig posture detection in the herd environment, which can elevate the application of intelligent technique in the pig industry.