Traffic Sign Detection and Recognition Using Gradient Training with an Improved YOLO Network

Automatic detection and recognition of traffic signs with utmost precision remains a central and unresolved hurdle in advancing high-tier assistance systems for vehicles. Current methods struggle with accurately discerning de-faced and obstructed signs and are susceptible to overfitting due to limitations in extensive training data. To overcome these deficiencies, we pioneer a training strategy harnessing tri-modal gradient information, enhancing the model’s capacity to discern signs amidst interference and occlusion without sacrificing performance in pristine, unobscured settings. Capitalizing on this innovation, we propose the IYOLO-TS model, specifically engineered to meet heightened robustness and real-time detection requisites. Derived from YOLOv8, the IYOLO-TS model introduces structural modifications, including the replacement of Detect-linked C2f architectures in the neck layer with VoVGSCSPC modules, and adoption of GSConv in lieu of standard convolutional layers. Comparative analysis shows that IYOLO-TS outperforms YOLOv8n, delivering enhanced robustness and real-time detection speeds concurrently with reduced computational complexity. Empirical evaluations demonstrate marked enhancements: the upgraded model, operating at a mere 7.3 GFLOPs, achieves 0.902 accuracy and 232.61 FPS on the third gradient dataset. Remarkably, it economizes computational resources and enhances real-time detection capabilities without compromising accuracy. The model’s demonstrated adaptability across varied conditions and high training efficiency renders it particularly suitable for practical vehicle deployments with stringent performance, size, and cost constraints.