Semi-Supervised Traffic Sign Detection with Dual Confidence Fusion Module and Structured Block-Regularized Neck

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Traffic sign detection is crucial for autonomous driving safety.
• Manual data annotation for training is resource-intensive.
• Semi-supervised learning (SSL) offers a viable alternative but faces challenges with pseudo-label reliability and accuracy.

Purpose of the Study:

• To develop a novel SSL framework to improve traffic sign detection accuracy.
• To address limitations in pseudo-label filtering and feature representation in existing SSL methods.

Main Methods:

• Proposed a framework integrating a Dual Confidence Fusion (DC-Fusion) module for reliable pseudo-labeling.
• Introduced a Structured Block-Regularized Neck (SBR-Neck) for optimized feature representation.
• Incorporated Spatial-Context-Aware Upsampling (SCA-Upsampling) within SBR-Neck to preserve spatial details.

Main Results:

• Achieved mAP50 scores of 10.4% (1% labeled data), 17.8% (2%), 23.7% (5%), and 32.1% (10%).
• Outperformed the 'Efficient Teacher' baseline by 3.07% to 11% across different labeled data percentages.
• Demonstrated robust detection performance in complex traffic scenarios.

Conclusions:

• The proposed DC-Fusion and SBR-Neck framework significantly enhances traffic sign detection using SSL.
• The method effectively improves pseudo-label reliability and feature representation, leading to superior accuracy.
• This framework offers a robust solution for autonomous driving systems requiring accurate traffic sign recognition with minimal labeled data.