Dense-RefineDet for Traffic Sign Detection and Classification.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Detecting small traffic signs in large images is difficult due to their limited pixel representation.
• Existing object detection models struggle with accuracy-speed trade-offs for small object detection.

Purpose of the Study:

• To develop a deep learning model for accurate and efficient detection and classification of small traffic signs.
• To improve the utilization of feature layers for enhanced contextual information in object detection.

Main Methods:

• Proposed Dense-RefineDet, a deep learning model based on the RefineDet single-shot object detection framework.
• Introduced a dense connection-related transfer-connection block to integrate high-level and low-level features.
• Developed a specialized anchor-design method tailored for small traffic sign detection.

Main Results:

• Dense-RefineDet achieved high-speed detection (0.13 s/frame) with competitive accuracy across small, medium, and large traffic signs.
• Demonstrated recall rates of 84.3% (small), 95.2% (medium), and 92.6% (large), with precision rates of 83.9%, 95.6%, and 94.0%, respectively.
• Achieved a miss rate of 54.03% for pedestrians (height > 20 pixels) on the Caltech dataset, surpassing state-of-the-art methods.

Conclusions:

• Dense-RefineDet effectively addresses the challenge of detecting small traffic signs while maintaining high accuracy and speed.
• The model's architecture and anchor design contribute to superior performance in small object detection.
• The findings suggest significant advancements in computer vision for traffic sign recognition and pedestrian detection.