A real-time and efficient detector for small object in UAV aerial images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Small object detection in Unmanned Aerial Vehicle (UAV) images is challenging due to high-resolution requirements and the need for real-time processing on resource-constrained platforms.
• Existing methods often struggle to balance detection accuracy and inference speed for aerial imagery.

Purpose of the Study:

• To propose a real-time UAV aerial image small object detection framework, RT-UAV-SOD, based on the Transformer architecture.
• To enhance the detection performance of small objects in UAV aerial images while maintaining real-time processing capabilities.

Main Methods:

• Developed a backbone network using Cascade Group Attention-inverted Residual Mobile Blocks (CGA-iRMB) to improve feature expression and multi-scale extraction.
• Incorporated a cross-stage fusion module to optimize multi-scale feature fusion, balancing accuracy and inference speed.
• Ensured model compatibility with edge devices for robust aerial image processing.

Main Results:

• On the VisDrone2019-DET dataset, RT-UAV-SOD achieved a 3.3% increase in precision, a 4.5% improvement in mAP50, and a 2.5% enhancement in mAP50:95.
• On the DOTA dataset, the model showed a 4.3% increase in precision, a 1.1% improvement in mAP50, and a 2.6% enhancement in mAP50:95.
• Demonstrated efficient real-time object detection capabilities for UAV applications.

Conclusions:

• RT-UAV-SOD provides an efficient and robust solution for small object detection in UAV aerial images.
• The proposed framework effectively addresses the challenges of real-time processing and accuracy on resource-constrained platforms.
• The integration of CGA-iRMB and cross-stage fusion significantly boosts detection performance in aerial imagery.