An advanced real-time traffic sign detection method based on improved YOLOv11n

Traffic Sign Recognition (TSR) is an indispensable component of autonomous driving technology, which is evolving rapidly. To address the challenge of low small-object detection accuracy in generalized object detectors for TSR, this paper proposes OCA-YOLOv11n, an improved object detection algorithm based on the latest YOLOv11n. We introduce the C3k2_OREPA module to replace the C3k2 module in the original network, enhancing feature extraction capability while boosting inference speed and reducing computational costs. Meanwhile, an additional small-object detection layer is incorporated, along with the introduction of the Cross-Scale Feature Fusion Module (CCFM), collectively boosting the model's small target detection capability. Additionally, the original Detect Head is replaced with the Four-head Adaptive Feature Fusion Detection Head (FASFFHead), performing secondary extraction focusing on small targets, which further enhances the model's capability to recognize minor traffic signs. Experiments conducted on the NVIDIA Jetson Xavier NX platform using the TT100K traffic sign dataset demonstrate that the OCA-YOLOv11n model achieves significant performance improvements over the original YOLOv11n model. The Precision, Recall, mAP50, and MAP50-95 increased by 7.8%, 13.3%, 12.8%, and 10.6%, respectively. The OCA-YOLOv11n model provides high precision and is available in embedded devices, meeting the practical requirements of robust traffic sign recognition.