An Efficient Algorithm for Small Livestock Object Detection in Unmanned Aerial Vehicle Imagery

Area of Science:

• Agricultural Technology
• Remote Sensing
• Artificial Intelligence

Background:

• Livestock population surveys are essential for effective grassland management, including disease prevention and resource assessment.
• Unmanned aerial vehicles (UAVs) offer advantages for surveys, but detecting small, densely packed livestock in images remains challenging.
• Existing methods struggle with the accurate identification of small livestock in aerial imagery.

Purpose of the Study:

• To develop an efficient algorithm for accurate livestock population detection using UAV imagery.
• To improve the identification of small and densely packed grazing animals in aerial surveys.
• To enhance livestock management and grassland sustainability through advanced detection techniques.

Main Methods:

• Proposed the Livestock Network (LSNET), a novel YOLOv7-based deep learning algorithm.
• Incorporated a low-level prediction head (P2) for small object detection and removed a deep-level head (P5) to reduce down-sampling effects.
• Introduced the Large Kernel Attentions Spatial Pyramid Pooling (LKASPP) module for high-level semantic feature capture and utilized WIoU v3 loss function.

Main Results:

• The LSNET algorithm demonstrated significantly improved detection accuracy for small livestock objects.
• Achieved a 1.47% increase in mean Average Precision (mAP) compared to the standard YOLOv7.
• Successfully developed a dedicated dataset of grazing livestock from UAV images in Inner Mongolia.

Conclusions:

• The proposed LSNET algorithm offers a practical and effective solution for livestock detection in large-scale farming environments.
• The integration of P2 head, LKASPP module, and WIoU v3 loss function enhances the detection of small livestock.
• This work contributes to advancements in agricultural technology and more efficient livestock management practices.