An Improved Bird Detection Method Using Surveillance Videos from Poyang Lake Based on YOLOv8.

Area of Science:

• Environmental Science
• Computer Vision
• Artificial Intelligence

Background:

• Poyang Lake, China's largest freshwater lake, is ecologically significant.
• Deep-learning video surveillance aids biodiversity preservation by monitoring bird species.
• Challenges include multi-scale detection, high density, and occlusion in complex environments.

Purpose of the Study:

• To develop an advanced deep-learning model for accurate bird detection and counting.
• To address challenges in detecting multi-scale, dense, and occluded bird species.
• To support biodiversity monitoring in Poyang Lake's complex ecosystem.

Main Methods:

• Proposed a novel YOLOv8-bird model incorporating Receptive-Field Attention convolution.
• Redesigned a feature fusion network (DyASF-P2) for enhanced small object feature capture.
• Implemented a lightweight detection head and Inner-ShapeIoU loss function for improved localization.

Main Results:

• The YOLOv8-bird model achieved high performance metrics: 94.6% precision, 89.4% recall, 94.8% mAP@0.5, and 70.4% mAP@0.5:0.95.
• Demonstrated superior accuracy compared to other mainstream object detection models.
• Validated on the PYL-5-2023 dataset for bird detection tasks.

Conclusions:

• The YOLOv8-bird model is effective for bird detection and counting in complex environments.
• The model's design overcomes challenges of scale, density, and occlusion.
• This technology supports crucial biodiversity monitoring efforts in Poyang Lake.