A Lightweight Traffic Sign Detection Method With Improved YOLOv7-Tiny

Abstract

As intelligent transportation systems continue to evolve, effective traffic sign detection is crucial for enhancing road safety and managing traffic congestion efficiently. This paper introduces the YOLOv7-tiny-RCA model, a novel lightweight approach designed specifically to address the dual challenges of high model complexity and diminished detection accuracy in long-distance scenarios commonly faced in traffic sign recognition. By integrating an improved ELAN-REP module into the YOLOv7-tiny backbone, our model significantly reduces computational complexity during the inference stage. Additionally, the incorporation of the CBAM attention mechanism enhances the model’s capability to extract and fuse relevant information from traffic scenes more effectively. To further optimize performance, we replaced the traditional feature fusion network with an Asymptotic Feature Pyramid Network, which facilitates better interaction between different layers and reduces the overall computational burden. Our experimental results demonstrate that the YOLOv7-tiny-RCA model achieves a mean Average Precision of 81.03% and a parameter reduction from 6.13M to 4.99M, highlighting its efficiency and potential for deployment on edge devices. These significant improvements indicate that our model not only advances traffic sign detection technologies but also offers practical applications for modern intelligent transportation systems.

Keywords

• Asymptotic feature pyramid network
• attention mechanism
• object detection
• RepConv
• traffic sign detection