Streamlining Livestock Surveillance: YOLOv8 Approach to Cattle Body Segmentation

This research paper introduces an innovative approach to optimize livestock surveillance by employing YOLOv8 for cattle body segmentation. The study attained a mAP(50), mean Average Precision of 0.598 and 0.464 and mAP(50-95) on a dataset comprising 704 images. Preceding model training, extensive preprocessing and image augmentation techniques were implemented, resulting in a dataset of 1476 training, 171 validation, and 71 testing images. The developed model demonstrates considerable promise in enabling real-time mon-itoring of cattle and their health status. Additionally, to enhance its suitability for edge computing environments, quantization-aware training (QAT) abd weight compression techniques were employed, ensuring efficient deployment on resource-constrained devices. The results highlight the effectiveness and practicality of employing YOLOv8-based segmentation for livestock surveillance, with notable implications for precision agriculture and animal welfare. However, further enhancements are warranted, particularly in achieving higher mAP scores across a broader range of Intersection over Union (IoU) thresholds, to fully address the segmentation challenges in this domain.