This paper presents a vision system that can be deployed in a ROS2 node that identifies potential predators such as wolves in order to herd autonomously a flock. A pre-trained YOLOv8 model and a pre-trained YOLOv9 model have been fine tunned so as to include more classes that are not supported such as wolves, and the whole models are re-trained with the COCO dataset and the new one, keeping the recognition of the previous learned classes as well as the new ones. To do that a dataset of images of wolves in open spaces has been acquired and labelled for instance segmentation purposes. The obtained vision model is wrapped in a ROS2 node, making it possible to implement it in a robotic device either a 4-legged robot or a rover robot. With this vision-based model, a robot can determine if a potential predator is detected and provide useful information to the herder to avoid area where the flock is in danger. Extensive livestock farming can benefit from the use of this technology to facilitate the tasks required on a daily basis.