A Training Strategy–Driven YOLOv11s Method for Traffic Sign Detection

Abstract

Traffic sign detection is a fundamental component of intelligent transportation systems, where both detection accuracy and real-time performance are critical for practical deployment. Although deep learning–based object detectors, particularly single-stage YOLO-based methods, have achieved promising results in traffic sign detection, their performance remains limited when dealing with visually similar categories, small-sized targets, and complex traffic environments. Moreover, many existing approaches rely on architectural modifications or additional modules, which may increase model complexity and hinder real-time applicability. To address these challenges, this paper proposes a training-strategy-driven traffic sign detection method based on YOLOv11s, termed TS-YOLOv11s. Without modifying the original network architecture or loss function formulation, the proposed method improves detection performance by optimizing the training process. Specifically, loss weight adjustment is employed to enhance fine-grained category discrimination, label smoothing is introduced to mitigate overconfidence caused by limited samples, and data augmentation strategies tailored for small objects and complex scenes are applied to improve robustness and generalization. Extensive experiments conducted on the Chinese Traffic Sign Dataset demonstrate that the proposed method achieves high detection accuracy while maintaining real-time inference efficiency. The results indicate that TS-YOLOv11s provides an effective balance between performance and computational cost, highlighting the potential of training strategy optimization as a general and practical approach for traffic sign detection in real-world intelligent transportation systems.