E-YOLOv4-tiny: a traffic sign detection algorithm for urban road scenarios.

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Driving Systems

Background:

• Traffic sign detection is crucial for autonomous driving but challenged by small object sizes and complex urban environments.
• Existing methods struggle with accuracy and efficiency in real-world road scenarios.

Purpose of the Study:

• To develop a more accurate and efficient traffic sign detection model for unmanned driving systems.
• To overcome limitations of current YOLOv4-tiny architectures in feature extraction and multi-scale integration.

Main Methods:

• An improved E-YOLOv4-tiny architecture was proposed, featuring an efficient layer aggregation lightweight block for enhanced feature extraction.
• A feature fusion refinement module with efficient coordinate attention was introduced to integrate multi-scale features and refine interference.
• An improved Spatial-Region of Interest Fusion (S-RFB) module was incorporated to add contextual information.

Main Results:

• The improved model achieved a 3.76% and 7.37% increase in mean Average Precision (mAP) on the CCTSDB and Tsinghua-Tencent 100K datasets, respectively.
• The number of model parameters was reduced by 21% compared to the original YOLOv4-tiny.
• Demonstrated superior performance over other advanced methods in balancing accuracy, real-time capability, and model size.

Conclusions:

• The proposed E-YOLOv4-tiny model offers significant improvements in traffic sign detection accuracy and efficiency for autonomous driving.
• The method presents a practical solution with high application value due to its balanced performance metrics.