Drone-DETR: Efficient Small Object Detection for Remote Sensing Image Using Enhanced RT-DETR Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicles (UAVs) require efficient object detection for tasks like remote sensing.
• Current embedded UAV systems struggle to balance speed and accuracy, especially with small, overlapping, or complexly-backgrounded objects.
• High-precision remote sensing image analysis presents significant computational challenges for onboard UAV processing.

Purpose of the Study:

• To develop an advanced object detection model for UAVs that excels in accuracy and efficiency.
• To specifically address the challenges of detecting numerous small objects in complex, ultra-wide-angle remote sensing images.
• To create a lightweight yet powerful model suitable for resource-constrained embedded UAV systems.

Main Methods:

• Introduced the Drone-DETR model, an adaptation of RT-DETR, incorporating the Effective Small Object Detection Network (ESDNet).
• Developed the Enhanced Dual-Path Feature Fusion Attention Module (EDF-FAM) to improve multi-scale object detection capabilities.
• Integrated a dynamic competitive learning strategy and the P2 shallow feature layer for enhanced feature fusion, particularly for small objects.

Main Results:

• The Drone-DETR model achieved a mean Average Precision (mAP) of 53.9% (mAP^50) on the VisDrone2019 dataset.
• The model utilizes only 28.7 million parameters, demonstrating a lightweight architecture.
• Achieved an 8.1% performance enhancement compared to the RT-DETR-R18 baseline.

Conclusions:

• Drone-DETR offers a significant advancement in UAV-based object detection, particularly for small and challenging targets.
• The proposed ESDNet and EDF-FAM modules effectively improve the handling of multi-scale and small objects in complex imagery.
• The model provides a robust and efficient solution for high-precision remote sensing applications on UAVs.