Efficient traffic sign recognition using YOLO for intelligent transport systems

Abstract

Abstract Accurate traffic sign recognition (TSR) is critical for enhancing the safety and reliability of autonomous driving systems. This study proposes an optimized YOLOv5-based framework to address challenges such as small-scale detection, environmental variability, and real-time processing constraints. Three key innovations are introduced: (1) k-means++ clustering for anchor box optimization, achieving a 77.55% average IoU (vs. 75.95% for traditional k-means) to enhance small-target detection; (2) comprehensive comparative analysis of YOLOv5 variants (s/m/x), revealing precision-speed trade-offs (99.3–99.5% mAP@0.5 vs. 32–45 ms inference time) for deployment flexibility; and (3) systematic hyperparameter tuning to maximize robustness across diverse scenarios. Leveraging the CCTSDB dataset (13,830 annotated images), experiments demonstrate the framework’s superiority: it attains 98.1% mean average precision (mAP), 98.6% recall, and 99.3% precision, outperforming Faster-RCNN and SSD by 5–8% in mAP while maintaining 45 FPS throughput. The YOLOv5s variant achieves optimal balance with 99.3% mAP@0.5 and 32 ms per-image inference, validated through rigorous statistical analysis (Tukey HSD). Robust performance in challenging conditions (e.g., small sample, backlit sample, foggy scenes) is evidenced by detection confidence exceeding 0.90. These results highlight the framework’s applicability in latency-sensitive intelligent transportation systems.