YOLO-UTD: A Domain-Specific Detection Framework for Small Objects in UAV Traffic Surveillance.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Detecting small, densely packed objects in drone aerial imagery presents significant challenges.
• Existing object detection models often struggle with the scale and distribution of targets in UAV surveillance.

Purpose of the Study:

• To develop a specialized small object detector, YOLO-UTD (YOLO-UAV Traffic Detection), for drone traffic surveillance.
• To enhance the YOLOv8 framework for improved detection of small objects in aerial imagery.

Main Methods:

• Introduced a specialized small-object detection head to replace the large-object head.
• Integrated a shallow-augmented feature pyramid network (SFPN) for enriched semantic content in high-resolution features.
• Incorporated a C2fA layer for adaptive fusion of spatial and semantic information using cross-attention.

Main Results:

• YOLO-UTD achieved a 3.6% higher mean average precision (mAP) compared to YOLOv8 on the VisDrone2019 dataset.
• Demonstrated a significant 5.3% increase in vehicle detection accuracy.
• Maintained a low parameter footprint while enhancing detection capabilities.

Conclusions:

• YOLO-UTD effectively addresses the challenges of small object detection in drone surveillance.
• The model shows strong potential for smart city applications and advanced drone traffic monitoring.
• The proposed enhancements significantly boost performance on fine-grained features and small targets.