UAV-DETR: An Enhanced RT-DETR Architecture for Efficient Small Object Detection in UAV Imagery

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Aerial imagery from unmanned aerial vehicles (UAVs) presents unique challenges for object detection, including feature degradation and spatial-contextual misalignment.
• Low resolution, complex backgrounds, and dynamic shooting conditions in UAV data hinder accurate detection of small objects.

Purpose of the Study:

• To propose UAV-DETR, an enhanced Transformer-based object detection model specifically designed to address the challenges of aerial imagery.
• To improve feature perception, semantic representation, and spatial alignment for more robust small-object detection in UAV-acquired data.

Main Methods:

• UAV-DETR extends the RT-DETR framework with three novel modules: Channel-Aware Sensing (CAS), Scale-Optimized Enhancement Pyramid (SOEP), and Context-Spatial Alignment (CSAM).
• CAS refines the backbone for better multi-scale feature perception.
• SOEP enhances shallow layer semantic richness via channel-weighted fusion.
• CSAM optimizes the hybrid encoder for contextual and spatial calibration, improving cross-scale integration.

Main Results:

• UAV-DETR demonstrates superior small-object detection performance on the VisDrone2019 dataset, achieving an mAP@0.5 of 51.6%.
• The model outperforms the baseline RT-DETR by 3.5% in mAP@0.5.
• UAV-DETR reduces model parameters from 19.8 million to 16.8 million compared to RT-DETR, indicating improved efficiency.

Conclusions:

• The proposed UAV-DETR effectively mitigates technical challenges in UAV-based object detection, particularly for small objects.
• The novel modules (CAS, SOEP, CSAM) significantly enhance feature perception, semantic representation, and spatial-contextual alignment.
• UAV-DETR offers a promising solution for accurate and efficient object detection in complex aerial scenarios.