Traffic sign detection is a key part of self-driving cars and driver assistance systems. In real life, signs can be very small, blocked by other objects, or hard to see in bad weather. Many recent works try to solve this by changing the model – they add attention, extra detection heads, or special loss functions. These changes make the model large and slow, and hard to use on small devices. In this paper, we take a different path. We keep the YOLOv8 model exactly as it is. We only improve the data through a strong augmentation pipeline (Mosaic, MixUp, rotation, cropping, brightness changes). We test on a public traffic sign dataset of 10,000 images with 29 classes. After removing two classes with wrong labels, our best model (YOLOv8l + full augmentation) gets 94.18% mAP@0.5 and 97.70% precision. The small YOLOv8n model with the same augmentation reaches 94.10% mAP , only 0.08% behind the large model, but with 18 times fewer parameters. A hard class (“50 mph speed limit”) improves by +45.30%. Our results prove that a good data strategy can match complex model changes and is much better for real-world deployment.

Traffic Sign Detection,Autonomous Driving,Data Centric AI,Deep Learning,Driver Assistance System