EBR-YOLO: A Lightweight Detection Method for Non-Motorized Vehicles Based on Drone Aerial Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Smart Transportation Systems

Background:

• Smart city initiatives require robust recognition of diverse traffic elements, including non-motorized vehicles.
• Drones offer flexible aerial data acquisition superior to fixed systems for urban monitoring.
• Existing object detection models may not be optimized for small object detection from drone imagery.

Purpose of the Study:

• To propose a novel, lightweight object detection architecture (EBR-YOLO) for identifying small non-motorized vehicles using drone-captured aerial images.
• To enhance the model's capability in detecting small objects by enriching detail information and improving learning.
• To reduce computational complexity and improve detection speed without significant accuracy loss.

Main Methods:

• Developed a lightweight YOLO-based architecture (EBR-YOLO) by reducing Backbone layers.
• Introduced the AZML module to enhance detail information and model learning for small targets.
• Reconstructed the C2f module using PConv to decrease computational load and increase detection speed.
• Reshaped downsampling operations with the ADown module to further reduce computational cost.

Main Results:

• EBR-YOLO achieved 98.9% mAP and 89.8 FPS on a custom dataset.
• The model demonstrated a 70% reduction in parameters compared to the baseline YOLOv8.
• Performance on UCAS-AOD and CARPK datasets showed comparable results to the baseline, indicating good generalization.

Conclusions:

• EBR-YOLO effectively balances high accuracy and computational speed while significantly reducing model size.
• The proposed lightweight architecture is suitable for real-time detection of small non-motorized vehicles in smart city environments.
• The model's efficiency and generalization capabilities make it a promising solution for urban intelligent transportation systems.