YOLIP: An Enhanced Framework for UAV-Assisted Wildlife Monitoring Based on YOLO Integrated with the CLIP Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Ecology

Background:

• Wildlife monitoring using Unmanned Aerial Vehicles (UAVs) faces challenges like low target visibility and scale variation in aerial images.
• Existing detection models struggle with accuracy and robustness in complex environmental conditions for wildlife observation.

Purpose of the Study:

• To present YOLIP, a novel fusion framework designed to improve the accuracy and feature representation for UAV-based wildlife detection.
• To address the challenges of complex environments and scale variations in aerial wildlife monitoring.

Main Methods:

• Developed YOLIP, integrating a detection head with semantic perception and an implicit feature adjustment module.
• Redesigned the detection head for simultaneous spatial and semantic feature learning.
• Implemented a dual-path fusion mechanism for geometric-semantic feature fusion and an asynchronous scheduling strategy for computational optimization.

Main Results:

• YOLIP achieved an 11.6% improvement in mAP@0.5 compared to existing models on self-built and public datasets.
• The framework demonstrated efficient inference speed alongside enhanced detection performance.
• Cross-dataset evaluation confirmed the method's stable performance and generalization capabilities.

Conclusions:

• YOLIP offers a significant advancement in UAV-based wildlife monitoring systems.
• The proposed fusion framework effectively enhances detection accuracy, robustness, and computational efficiency for aerial wildlife detection.