YOLO-SDD: An Effective Single-Class Detection Method for Dense Livestock Production

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Single-class object detection is crucial for optimizing farm operations through animal identification, counting, and tracking.
• Dense occlusion in group animal activities presents a significant challenge for accurate detection.

Purpose of the Study:

• To develop an effective object detection network, YOLO-SDD, specifically for single-class, densely populated scenarios.
• To improve the recognition of occluded targets in livestock group settings.

Main Methods:

• Introduced Wavelet-Enhanced Convolution (WEConv) for improved feature extraction under occlusion.
• Proposed an occlusion perception attention mechanism (OPAM) to leverage low-level and high-level features for better occluded target recognition.
• Incorporated a Lightweight Shared Head (LS Head) optimized for single-class dense detection tasks.

Main Results:

• YOLO-SDD variants (n, s, m) showed significant AP50:95 improvements over YOLOv8 on the ChickenFlow dataset.
• Outperformed the latest real-time detector, YOLOv11, in detection performance.
• Achieved state-of-the-art results on GooseDetect and SheepCounter datasets for crowded livestock detection.

Conclusions:

• YOLO-SDD provides a robust solution for automated livestock tracking and counting in dense conditions.
• The model's efficiency and accuracy support advancements in precision livestock farming.
• Demonstrated superior performance in handling dense occlusion scenarios in animal detection.