Abstract: In response to issue of livestock crossing boundaries in pastures and limitations of traditional electronic fences in preventing such occurrences, a virtual electronic fence was designed based on improved version of YOLOv5．Initially, YOLOv5s model was used as foundation, and transfer learning and addition of an ECA attention module were performed．Subsequently, the model was trained using PyTorch framework and evaluated．Compared to original YOLOv5s, improved YOLOv5s achieved 0.2, 1.3, and 0.7 percentage points increase in precision, recall, and mAP for cattle detection, respectively, while reducing single-frame inference total time by 0.5 ms．Finally, improved model was converted to RKNN format and deployed on an NPU-equipped RK3588 development board to accelerate model's inference speed．Results showed that, application of deep learning technology and ROI delineation technology has successfully designed a virtual electronic fence for livestock, optimized management system of smart ranches, improved management efficiency, reduced management costs, and has certain practical value．