Abstract: To address the challenge of detecting sheep behavior at night, this study proposes a lightweight infrared image-based behavior recognition method. First, an efficient channel attention module with cross-channel interaction is designed to improve the feature extraction and learning capabilities of the model. Second, depthwise separable convolution is adopted in place of standard convolution within the backbone network, thereby reducing model parameters, increasing computational speed, and enhancing the algorithm's suitability for deployment on mobile devices. Finally, to further improve computational efficiency, the small-target detection head is eliminated. Experimental results demonstrate that the proposed Night-YOLO algorithm improves average detection accuracy by 4% compared with the benchmark model, achieving 94.4%, while reducing the number of parameters by 10% to only 2.7 M. These findings confirm that the algorithm can accurately localize and identify sheep behavior during nighttime conditions.