Real-Time and Efficient Multi-Scale Traffic Sign Detection Method for Driverless Cars.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Traffic sign detection is critical for autonomous driving but challenging due to small object sizes.
• Existing object detection methods often struggle with the accuracy required for small traffic sign recognition.

Purpose of the Study:

• To improve the accuracy of small traffic sign detection for driverless cars.
• To enhance the YOLOv4 network's capability in identifying small traffic signs.

Main Methods:

• An improved triplet attention mechanism was integrated into the YOLOv4 backbone for better feature acquisition.
• A bidirectional feature pyramid network (BiFPN) was employed in the neck for enhanced feature fusion and perception of small objects.

Main Results:

• The enhanced YOLOv4 model achieved 60.4% mAP (Mean Average Precision) on the TT100K-COCO dataset, an 8% improvement over the original YOLOv4.
• A maximum performance of 66.4% mAP was reached with a larger input size, demonstrating superior small object detection capabilities.

Conclusions:

• The proposed optimization strategy significantly enhances traffic sign detection accuracy for autonomous driving systems.
• This research offers a valuable reference for developing more robust and accurate traffic sign recognition systems in driverless vehicles.