RT-DETR-EVD: An Emergency Vehicle Detection Method Based on Improved RT-DETR

Area of Science:

• Computer Vision
• Artificial Intelligence
• Urban Systems

Background:

• Urbanization and increased traffic congestion pose challenges for emergency vehicle response times.
• Timely detection of emergency vehicles is critical for efficient rescue operations.

Purpose of the Study:

• To develop a lightweight and high-precision emergency vehicle detection model (RT-DETR-EVD) for enhanced urban emergency response.
• To improve the efficiency and accuracy of emergency vehicle detection in congested environments.

Main Methods:

• Developed a novel RT-DETR-EVD model using a lightweight CSPDarknet backbone and a hybrid C2f-MogaBlock architecture.
• Integrated a multi-order gated aggregation mechanism for dynamic multi-scale feature fusion.
• Introduced an Attention-based Intra-scale Feature Interaction Dynamic Position Bias (AIDPB) module to replace fixed positional encoding with learnable dynamic position bias (DPB).

Main Results:

• RT-DETR-EVD reduced parameter count by 27.1% (to 14.5 M) and FLOPs by 13.2% (to 49.5 G) compared to the baseline RT-DETR-r18.
• Achieved a 0.5% increase in precision, with recall and mAP50% increasing by 0.6% to reach 88.3%.
• Demonstrated superior performance in emergency vehicle detection with improved accuracy, efficiency, and scene adaptability.

Conclusions:

• The RT-DETR-EVD model offers a breakthrough balance between accuracy, efficiency, and adaptability for emergency vehicle detection.
• Its lightweight design accelerates inference and reduces model size, making it suitable for real-world smart city applications.
• Provides an efficient and reliable solution for improving urban emergency response systems.