LLSA‐YOLO: Lightweight Day‐and‐Night Cattle Detection for Edge Deployment in Real‐World Pasture Monitoring

ABSTRACT

To address the challenges of cattle detection in real‐world pasture environments, such as pronounced day‐and‐night illumination variations, distant small targets, and limited computing power of edge devices, this paper proposes a lightweight cattle detection method, LLSA‐YOLO, for day‐and‐night pasture monitoring. Building upon existing object detection frameworks, this method makes targeted improvements to three key stages: downsampling, feature fusion, and target prediction. These improvements aim to reduce information loss under low‐light conditions, suppress interference from complex backgrounds, and enhance the perception of distant small targets. Furthermore, considering that real‐world monitoring systems are typically deployed on resource‐constrained edge devices, the model design strikes a balance between detection accuracy and computational cost. Based on real pasture monitoring videos, this study constructs a cattle data set encompassing both daytime and night‐time scenes. A pure‐dark test set (PD‐Test) is further established to evaluate the model's detection capabilities in low‐light environments. Experimental results show that LLSA‐YOLO achieves an mAP@0.5 of 0.9837 on the standard test set and 0.8980 on the PD‐Test set, outperforming several mainstream object detectors on both benchmarks. Meanwhile, the model has a complexity of only 3.3 GFLOPs, a size of 2.8 MB, and can achieve real‐time detection of 28.40 FPS on edge devices. The results indicate that the proposed method achieves high detection accuracy while maintaining a lightweight design, adaptability to night‐time scenes, and deployment practicality.