GLE-YOLOv11n: A lightweight network for real-time recognition of sheep behaviours in edge computing environments

Real-time monitoring of sheep behaviour in large-scale sheep farms is crucial for precision livestock farming and can provide useful information for group-level welfare-related monitoring. Traditional deep learning-based behaviour recognition models often suffer from high computational complexity and large parameter sizes, making it difficult to efficiently deploy on resource-constrained edge computing devices. To overcome this limitation, this paper proposes a lightweight sheep behaviour recognition method based on an improved YOLOv11n. First, standard convolutions in backbone and neck networks are replaced with the ghost-shuffle convolution (GSConv) module, effectively reducing model parameters and computational cost while preserving feature extraction capability. Second, the original cross-stage partial spatial attention (C2PSA) module is replaced with a lightweight gated attention-lite (LWGA-Lite) mechanism, enhancing the representation of behaviour-critical features. The bounding box regression is optimized using the efficient IoU (EIoU) loss function, improving localization accuracy and convergence speed for multi-scale targets, thereby facilitating the recognition of four sheep behaviours: lying, standing, eating, and drinking. Finally, the GLE-YOLOv11n model was deployed on the Jetson Nano B01 platform using the TensorRT framework with FP16 precision optimization. Experimental results indicate the capability of the improved GSConv LWGA-Lite EIoU-YOLOv11n (GLE-YOLOv11n) model to achieve a detection accuracy of 94.79% in terms of mAP@0.5:0.95, while maintaining only 2.20 M parameters and a computational cost of 5.5 GFLOPs. On the Jetson Nano platform with an input resolution of 640 × 640 pixels, the model achieves an inference rate of 15.0 FPS. The proposed method improves model compactness and maintains practical inference speed while achieving high accuracy, demonstrating its suitability for edge-intelligence applications in smart livestock farming.