Objective This study was designed to enable accurate recognition of multiple pig behaviors in complex environments, facilitate the development of precision livestock farming, and support pig status and disease monitoring.

Method A lightweight object detection model named LAD-YOLO, based on YOLOv8n, was proposed to address the challenges in behavior detection in scenarios characterized by background similarity, illumination variations, and inter-individual occlusion. An alterable kernel convolution (AKConv) module was incorporated into the backbone network, where its flexible parameterization and variable sampling-shape kernels were used to enhance multi-scale feature extraction. The adaptive downsampling (ADown) module was embedded to reduce model parameters through shortened strides and average pooling, while preserving critical information. In the neck, the large separable kernel attention (LSKA) mechanism was integrated into the C2f module, leveraging its feature extraction capability to enhance the global feature representation and key region awareness of C2f.

Result Experimental results demonstrated that LAD-YOLO increased precision and recall by 2.5 and 7.2 percentage points, respectively, achieved an mAP₅₀ of 98.7%, and reduced the model size to 4.9 MB with an 18.9% decrease in parameters. Compared with mainstream detectors and under complex backgrounds and lighting variations, LAD-YOLO exhibited lighter weight, higher detection accuracy, and better robustness.

Conclusion LAD-YOLO performs excellently in recognizing multiple pig behaviors under complex environments and offers an efficient and reliable technical solution for intelligent livestock management.