KT-YOLO: A multi-convolution kernel collaboration model for dense Hu sheep behavior detection

Computer vision has been extensively applied to sheep behavior detection in recent years. However, the dense distribution of Hu sheep poses detection challenges, while imbalanced behavioral categories in datasets affect classification accuracy for detection tasks in intensive farming scenarios, resulting in high misclassification rates. Current models often rely on over-parameterization to achieve satisfactory detection performance, which increases computational burden and limits practical deployment. To address these challenges, this study introduces the Hu Sheep Behavior Dataset (HSBD), specifically designed for intensive farming environments. The dataset comprises 280 images capturing four behaviors across 6,766 Hu sheep: standing, lying, eating, and drinking. Building upon this foundation, we developed the KT-YOLO model, which utilizes a novel Kernel-Team Fusion (KTF) method to enhance the YOLOv8n detection framework. By employing four different convolution kernel sizes, this method effectively captures multi-scale features and addresses Hu sheep occlusion challenges. To mitigate accuracy degradation caused by dataset imbalance, KT-YOLO incorporates a SlideLoss function during classification, effectively addressing this challenge. Comparative experiments demonstrate that KT-YOLO achieved a mean Average Precision (mAP50) of 86.4%, representing a 6.3 percentage point improvement over YOLOv8n, with SlideLoss contributing an additional 1 percentage point improvement. Further comparison with YOLOv13n demonstrates KT-YOLO's superior performance in dense Hu sheep behavior detection. By introducing HSBD and developing the innovative KT-YOLO, this study significantly enhances both accuracy and efficiency of dense Hu sheep behavior detection, demonstrating the potential and practical value of deep learning technologies in intensive farming environments.