YOLO-CATTLE: an improved YOLO11-based model for ground cattle detection

Abstract

Ground cattle detection remains a challenging task due to dense object distribution, significant scale variation, interference from complicated backgrounds and challenges in identifying tiny objects. This research suggests an enhanced detection model to overcome these problems, termed YOLO-CATTLE, based on a YOLO framework. Specifically, several enhancements are introduced. First, a lightweight backbone network based on the MBConv structure is incorporated to reduce model parameters and computational complexity while improving feature extraction capability. Second, a Bi-Level Routing Attention (BRA) mechanism is integrated to enhance the representation of small-scale and occluded cattle, especially in long-distance scenarios. Third, an improved Wise-IoU loss function is adopted to refine bounding box regression accuracy and improve training stability through dynamic weighting. The suggested approach outperforms the baseline YOLO model, achieving 94.3% mAP@0.5 and 69.5% mAP@0.5:0.95, according to experimental results on a self-constructed cow dataset.