Abstract
In developing Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS), traffic sign detection is a critical task. However, existing methods face limitations in real-time performance and small-object detection accuracy, partly due to deployment platform constraints. Our proposed Neck-Heavy YOLOv5 (NH-YOLOv5) introduces significant improvements. It features a hybrid data augmentation approach that enhances the detection of small objects. Moreover, a unique neck-heavy network architecture is designed to optimize detection performance without increasing computational complexity. Notably, NH-YOLOv5 is integrated into the digital twin framework of traffic systems. By leveraging the digital twin’s ability to simulate and monitor traffic scenarios in real time, the model can continuously adapt and improve its detection capabilities. Experimental results based on the Tsinghua - Tencent 100K (TT100K) dataset demonstrate that our method achieves a 53% reduction in computation while enhancing accuracy and small object detection ability, making it highly suitable for real-time traffic sign detection in the context of digital twins.