Urban traffic tiny object detection via attention and multi-scale feature driven in UAV-vision.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicle (UAV) city patrols are crucial for safety and urban operations.
• UAV image detection faces challenges: small objects, complex backgrounds, and speed requirements.

Purpose of the Study:

• To develop an efficient and accurate real-time object detection network for UAV city patrols.
• To address the limitations of current methods in detecting small objects and complex scenes.

Main Methods:

• Introduced a Real-time Small Object Detection network in UAV-vision (RTS-Net).
• Developed a multiscale feature fusion module (MFFM) for enhanced small object detection.
• Integrated a coordinated attention detection module (CADM) for improved background segregation.
• Incorporated a lightweight real-time feature extraction module (RFEM) to increase inference speed.

Main Results:

• Achieved a record detection accuracy of 89.9 mAP on a custom UAV road patrol dataset.
• Demonstrated superior performance over existing methods, especially for small-scale object detection.
• Attained a high inference speed of 163.9 FPS, enabling real-time application.

Conclusions:

• RTS-Net effectively meets the demand for accurate and efficient ground object detection by UAVs.
• The network is suitable for various UAV platforms and complex urban scenarios.
• The proposed method advances the capabilities of UAV-based surveillance systems.