UDRT-DETR: a small UAV detection method based on infrared imaging and RT-DETR

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Illegal unmanned aerial vehicle (UAV) flights pose significant public safety risks, necessitating effective detection methods.
• Existing UAV detection systems struggle with small targets due to low resolution, clutter, and dense distribution.
• Transformer-based detectors offer improved accuracy but are computationally intensive, limiting real-time application.

Purpose of the Study:

• To develop an efficient and accurate infrared-based detection method for small UAVs.
• To overcome the computational and latency limitations of existing transformer detectors for UAV surveillance.
• To enhance the detection of small, densely packed UAVs in complex environments.

Main Methods:

• Proposed UDRT-DETR, integrating infrared imaging with the Real-Time Detection Transformer (RT-DETR).
• Introduced a cascaded inverted residual backbone (CIRB) to reduce computational cost and enhance feature representation.
• Developed a super token attention-based intra-scale feature interaction (STA-IFI) module for efficient global context and dense target detection.
• Designed a slim-neck-ASF for precise multi-scale feature fusion and an inner-MPDIoU loss for improved bounding-box regression.

Main Results:

• UDRT-DETR achieved 90.71% precision on a self-built infrared UAV dataset, outperforming RT-DETR by 4.66%.
• The method reduced computational complexity by 17.84% (GFLOPs).
• Demonstrated state-of-the-art accuracy and enabled real-time surveillance capabilities.

Conclusions:

• UDRT-DETR offers a significant advancement in real-time small UAV detection using infrared imaging.
• The proposed architectural modifications effectively balance accuracy and computational efficiency.
• This method provides a viable solution for enhancing public safety against unauthorized UAV activities.