YOLOv7-TS: A Traffic Sign Detection Model Based on Sub-Pixel Convolution and Feature Fusion

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning has advanced object detection, with traffic sign detection being a key application.
• Existing methods struggle with small object omission and low accuracy in real-world traffic scenes.

Purpose of the Study:

• To develop an improved traffic sign detection model addressing limitations of current approaches.
• To enhance the detection of small traffic signs and overall accuracy in complex environments.

Main Methods:

• Proposed YOLOv7-Traffic Sign (YOLOv7-TS) model utilizing sub-pixel convolution and feature fusion.
• Introduced Feature Map Extraction Module (FMEM) to prevent channel information loss.
• Developed Multi-feature Interactive Fusion Network (MIFNet) for better feature interaction and small object perception.
• Incorporated Deep Feature Enhancement Module (DFEM) to enrich high-level features.

Main Results:

• YOLOv7-TS demonstrated significant improvements over the baseline YOLOv7.
• Achieved 3.63% and 2.68% higher mean Average Precision (mAP) on CCTSDB2021 and TT100K datasets, respectively.
• The model shows enhanced perception of small objects and improved detection accuracy.

Conclusions:

• The proposed YOLOv7-TS model effectively addresses the challenges of small object omission and low accuracy in traffic sign detection.
• The integration of sub-pixel convolution and advanced feature fusion modules leads to superior performance.
• YOLOv7-TS offers a promising solution for real-world traffic sign recognition systems.