Improved YOLOv5-based for small traffic sign detection under complex weather.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Traffic sign detection is crucial for autonomous driving but challenged by varying object sizes and weather conditions.
• Existing models struggle with detecting small or occluded traffic signs, especially in adverse weather, impacting overall detection accuracy.

Purpose of the Study:

• To develop an improved YOLOv5s algorithm for robust traffic sign detection in complex weather.
• To enhance the detection precision and recall for small and occluded traffic signs.

Main Methods:

• Incorporated coordinate attention (CA) mechanism in the backbone for improved feature extraction.
• Added a prediction head to YOLOv5s to leverage fine-grained features from shallower layers for small object detection.
• Utilized Alpha-IoU to enhance bounding box regression accuracy over CIoU loss.

Main Results:

• The improved YOLOv5s model achieved 88.1% precision and 79.8% recall for small objects on the CCTSDB 2021 dataset.
• Demonstrated significant improvements of 12.5% in precision and 23.9% in recall compared to the original YOLOv5s model.
• Effectively detected small traffic signs under diverse weather conditions with reduced miss rates and high accuracy.

Conclusions:

• The proposed algorithm effectively addresses the challenges of detecting small traffic signs in complex weather conditions.
• The integration of coordinate attention, an additional prediction head, and Alpha-IoU significantly boosts detection performance.
• This work contributes to safer and more reliable autonomous driving systems through improved traffic sign recognition.