Scholarlink

<p>Automated traffic sign inspection is a key component of intelligent transportation systems and sustainable infrastructure management. This paper presents a Jordan-specific benchmark for traffic sign detection that compares three single-stage detectors (YOLOv8, YOLOv9, YOLOv10) trained and evaluated under identical conditions on a new annotated dataset. The dataset contains 1,500 high-resolution dashcam images with 1,158 annotated sign instances collected from 22 urban corridors in Amman across seven sign classes: speed limit, informational, obligatory, warning, regulatory, priority, and traffic light. The dataset was partitioned at the corridor level to prevent spatial leakage between training and validation. An ablation study isolates the contributions of pretrained weights and data augmentation across four configurations, showing that the pretrained-plus-augmentation configuration reaches a validation mAP@0.5 of approximately 0.78 by epoch 100, compared with 0.65 for scratch-plus-augmentation, 0.59 for pretrained-without-augmentation, and 0.50 for scratch-without-augmentation. On the held-out validation set, YOLOv9 achieves the highest mean class mAP@0.5 (0.797), YOLOv8 the highest mean F1-score (0.767), and YOLOv10 the highest inference throughput (42.3 FPS) at the smallest model size (31.9&#xa0;MB). The weakest class across all three architectures is the priority sign (mAP@0.5 between 0.18 and 0.51), attributable to the small number of annotated priority instances (14) and visual overlap with the regulatory and warning categories. The paper closes with a proposed detection output schema and an asset-management integration workflow that links detections to corridor inventory and maintenance prioritisation. The deployment scope is restricted to vehicle-mounted dashcam imagery; CCTV or fixed-camera viewpoints would require a separate dataset and model-tuning pass. The annotated dataset, trained model weights and per-class results, and the training and inference code are openly available at the URLs listed in the Data and Code Availability section.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning for smart roadway asset management: a multi-architecture YOLO comparison for traffic sign recognition in Jordan

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Taqwa Alhadidi,
  • Ibrahim Asi,
  • Ahmad Alomari,
  • Mohammed Elhenawy

摘要

Automated traffic sign inspection is a key component of intelligent transportation systems and sustainable infrastructure management. This paper presents a Jordan-specific benchmark for traffic sign detection that compares three single-stage detectors (YOLOv8, YOLOv9, YOLOv10) trained and evaluated under identical conditions on a new annotated dataset. The dataset contains 1,500 high-resolution dashcam images with 1,158 annotated sign instances collected from 22 urban corridors in Amman across seven sign classes: speed limit, informational, obligatory, warning, regulatory, priority, and traffic light. The dataset was partitioned at the corridor level to prevent spatial leakage between training and validation. An ablation study isolates the contributions of pretrained weights and data augmentation across four configurations, showing that the pretrained-plus-augmentation configuration reaches a validation mAP@0.5 of approximately 0.78 by epoch 100, compared with 0.65 for scratch-plus-augmentation, 0.59 for pretrained-without-augmentation, and 0.50 for scratch-without-augmentation. On the held-out validation set, YOLOv9 achieves the highest mean class mAP@0.5 (0.797), YOLOv8 the highest mean F1-score (0.767), and YOLOv10 the highest inference throughput (42.3 FPS) at the smallest model size (31.9 MB). The weakest class across all three architectures is the priority sign (mAP@0.5 between 0.18 and 0.51), attributable to the small number of annotated priority instances (14) and visual overlap with the regulatory and warning categories. The paper closes with a proposed detection output schema and an asset-management integration workflow that links detections to corridor inventory and maintenance prioritisation. The deployment scope is restricted to vehicle-mounted dashcam imagery; CCTV or fixed-camera viewpoints would require a separate dataset and model-tuning pass. The annotated dataset, trained model weights and per-class results, and the training and inference code are openly available at the URLs listed in the Data and Code Availability section.