Wildlife and Livestock Dataset with Emphasis on Predators for Vision-Based Herding Systems

This dataset supports real-time detection of herding-related entities including sheep, wolves, dogs, bears, coyotes, foxes, red foxes, wild dogs, and humans for autonomous robotic shepherds. It has been reformatted from raw YOLOv5-style datasets using a custom pipeline that ensures class balancing, strong augmentations via Albumentations, and source diversity based on video origin. It combines three sources: a sheep dataset (Sheep_dataset_link), the COCO dataset (Coco_dataset_link) for person and dog classes, and predator datasets (predator_dataset_link). Rare classes (e.g., wolf, cow, dog, wild dog) have been augmented using a recycling strategy that combines oversampling with high-variance augmentations to increase robustness (Python_Code). Splits (train/validation/test) are carefully balanced to maintain class distribution and diversity. The dataset is compatible with YOLOv5–YOLOv12, RT-DETR, and custom models such as SHEPHERD, and has been validated in real-time on ROS 2 (Humble) deployed on quadruped robots (e.g., Unitree Go1) for flock monitoring, predator deterrence, and shepherd alerting. The archive contains images, YOLO-format labels, metadata CSVs, and summary plots for class balance and entropy, providing a ready-to-use resource for object detection, robotics, and intelligent livestock management.