Small traffic sign recognition method based on improved YOLOv7.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traffic sign recognition is crucial for autonomous and assisted driving systems.
• Current algorithms struggle with small traffic signs, complex backgrounds, and poor lighting, leading to detection errors.

Purpose of the Study:

• To introduce an enhanced method for small traffic sign recognition using an improved YOLOv7 architecture.
• To improve the accuracy and robustness of traffic sign detection in challenging environmental conditions.

Main Methods:

• Utilized Spatial Pyramid Pooling Fast and Cross-Stage Partial Connection (SPPFCSPC) for enhanced small target feature extraction.
• Developed a Shuffle Attention-CARAFE (S-CARAFE) up-sampling operator to refine feature details and recombination.
• Introduced a Normalized Wasserstein Distance (NWD) method to address limitations of traditional IoU measurements for small targets.

Main Results:

• Achieved a 3.48% increase in mAP@0.5 and 2.29% in mAP@0.5:0.9 on the TT100K dataset.
• Demonstrated improvements of 2.61% in mAP@50 and 2.12% in mAP@50:95 compared to similar algorithms.
• Validated effectiveness on CCTSDB and foreign traffic sign datasets, confirming enhanced small traffic sign recognition.

Conclusions:

• The proposed enhanced YOLOv7 method significantly improves small traffic sign recognition accuracy.
• The method is effective across diverse datasets and challenging conditions, advancing autonomous driving capabilities.
• This work contributes to more reliable and safer autonomous driving systems through superior traffic sign detection.