Toward Precision Biodiversity Detection: An Edge-Deployable Framework for Mitigating Data Redundancy.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Wildlife Ecology

Background:

• Infrared cameras in remote areas generate excessive false positives, leading to data overload and wasted power.
• Current methods for wildlife monitoring struggle with environmental interference and efficient data processing.

Purpose of the Study:

• To develop a lightweight, edge-optimized wildlife detection model (WS-YOLO) for field applications.
• To integrate false-trigger filtering and accurate animal detection into a single, efficient model.
• To reduce data transmission and power consumption in remote wildlife monitoring systems.

Main Methods:

• Proposed WS-YOLO, an improved YOLOv11n model incorporating Wavelet Convolution (WTConv) for multi-scale feature extraction.
• Integrated Spatial-Channel Synergistic Attention (SCSA) for enhanced focus on wildlife targets.
• Developed a SlideLoss function to improve handling of ambiguous detection instances.

Main Results:

• WS-YOLO achieved high mAP@0.50 performance: 93.0% (African Wildlife), 98.0% (Amur Tiger), 98.5% (NTLNP infrared), and 79.4% (PASCAL VOC).
• Demonstrated robustness across diverse wildlife datasets and scenarios.
• Validated real-time, energy-efficient field deployment on a Jetson Nano edge device.

Conclusions:

• WS-YOLO effectively filters false triggers and accurately detects wildlife at the edge.
• The model significantly enhances wildlife monitoring efficiency by reducing data load and power usage.
• Enables direct, on-device animal recognition and classification in remote environments.