Vision-Based Module for Herding with a Sheepdog Robot.

Area of Science:

• Agricultural Technology
• Computer Vision
• Wildlife Management

Background:

• Livestock farming faces challenges in predator control, especially in remote areas.
• Technological solutions, including robotics and AI, are increasingly integrated into modern agriculture.
• Predation by animals like the Iberian wolf poses a significant threat to livestock in regions such as the Iberian Peninsula.

Purpose of the Study:

• To develop an automated system for generating animal image datasets.
• To create a vision-based module for real-time predator detection and differentiation from livestock.
• To identify the most efficient object detection model for livestock protection applications.

Main Methods:

• Utilized the iNaturalist API to automatically generate diverse animal image benchmarks.
• Developed a vision-based module for automated detection and classification of predators and other animals.
• Evaluated multiple object detection models for speed and accuracy in real-time environments.

Main Results:

• The YOLOv5m model demonstrated superior performance, processing at 64 frames per second (FPS).
• Achieved a mean Average Precision (mAP) of 99.49% (with IoU of 50%) in detecting and distinguishing wolves from dogs.
• The developed system meets the performance requirements for pasture-based livestock farms.

Conclusions:

• The proposed system offers an effective solution for automated predator detection in livestock farming.
• YOLOv5m is identified as a highly efficient model for real-time wildlife monitoring applications.
• This technology enhances the safety and management of grazing animals against predators.