Optimized Small Waterbird Detection Method Using Surveillance Videos Based on YOLOv7.

Area of Science:

• Ecology
• Computer Science
• Artificial Intelligence

Background:

• Waterbird monitoring is crucial for wetland conservation and management.
• Advancements in wetland infrastructure generate large volumes of wildlife data.
• Deep learning offers solutions for big data analysis but needs validation for real-time wildlife monitoring.

Purpose of the Study:

• To develop and evaluate an improved deep learning method for real-time waterbird detection in surveillance videos.
• To address the challenge of accurately identifying small waterbirds in dynamic environments.

Main Methods:

• An enhanced YOLOv7 model, named YOLOv7-waterbird, was developed by incorporating an additional prediction head, SimAM attention module, and sequential frame analysis.
• The model was trained and evaluated using the Waterbird Dataset.

Main Results:

• YOLOv7-waterbird achieved a mean average precision (mAP) of 67.3%, a 5% improvement over the baseline model.
• The model demonstrated superior performance in detecting small waterbirds (pixels < 40x40), with a recall of 87.9% and precision of 85%.
• The algorithm showed a 79.1% detection rate for small waterbirds.

Conclusions:

• The YOLOv7-waterbird algorithm significantly improves real-time waterbird detection, particularly for small species.
• This method offers a valuable tool for wildlife conservation, enabling more accurate monitoring using existing surveillance infrastructure.