EnhancedMulti-Scenario Pig Behavior Recognition Based on YOLOv8n

Abstract

Advances in smart animal husbandry necessitate efficient pig behavior monitoring, yet traditional approaches suffer from operational inefficiency and animal stress. We address these limitations through a lightweight YOLOv8n architecture enhanced with SPD-Conv for feature preservation during downsampling, LSKBlock attention for contextual feature fusion, and a dedicated small-target detection head. Experimental validation demonstrates superior performance: the optimized model achieves a 92.4% mean average precision (mAP@0.5) and 87.4% recall, significantly outperforming baseline YOLOv8n by 3.7% in AP while maintaining minimal parameter growth (3.34M). Controlled illumination tests confirm enhanced robustness under strong and warm lighting conditions, with performance gains of 1.5% and 0.7% in AP, respectively. This high-precision framework enables real-time recognition of standing, prone lying, lateral lying, and feeding behaviors in commercial piggeries, supporting early health anomaly detection through non-invasive monitoring.