Scholarlink

<p>Accurately detecting multilingual, distant, and small traffic signs in real time remains a major challenge for autonomous vehicles operating in diverse and dynamic environments. This study introduces M-Ghost YOLOv8, a lightweight and efficient model designed for multilingual traffic sign recognition. The model integrates GhostConv and C3Ghost modules to reduce parameters and computation while preserving detection accuracy. A Selective Pyramid Pooling and Efficient Lateral Aggregation Network (SPPELAN) is incorporated to enhance multi-scale feature fusion, improving recognition of small and partially occluded signs. Furthermore, an additional detection head is introduced to enhance performance on extra-small objects, which are commonly encountered in real-world driving scenarios. To support multilingual adaptability, we present TT100k-G, a bilingual version of the TT100k dataset that combines Chinese and globally interpretable traffic sign annotations. Experimental results on the TT100k, GTSDB, and TSDY datasets show that M-Ghost YOLOv8 achieves mean average precision mAP(0.50) scores of 90.5%, 98.5%, and 97.1%, respectively, outperforming recent models such as YOLOv12. Compared with YOLOv8, our model reduces parameters by 76.5%, GFLOPs by 57.5%, and doubles inference speed (up to 60 FPS). Real-world tests on the Wheeltec autonomous platform confirm reliable performance under multilingual and multi-scale conditions. M-Ghost YOLOv8 offers a scalable, edge-ready solution for intelligent transportation systems, balancing efficiency, accuracy, and language adaptability to enable safer, real-time decision-making in autonomous driving applications.</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

M-Ghost YOLOv8: a novel lightweight architecture for real-time multilingual traffic sign detection and classification

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Mohammed Hussein Osman Alnageeb,
  • Wen-Xing Zhu

摘要

Accurately detecting multilingual, distant, and small traffic signs in real time remains a major challenge for autonomous vehicles operating in diverse and dynamic environments. This study introduces M-Ghost YOLOv8, a lightweight and efficient model designed for multilingual traffic sign recognition. The model integrates GhostConv and C3Ghost modules to reduce parameters and computation while preserving detection accuracy. A Selective Pyramid Pooling and Efficient Lateral Aggregation Network (SPPELAN) is incorporated to enhance multi-scale feature fusion, improving recognition of small and partially occluded signs. Furthermore, an additional detection head is introduced to enhance performance on extra-small objects, which are commonly encountered in real-world driving scenarios. To support multilingual adaptability, we present TT100k-G, a bilingual version of the TT100k dataset that combines Chinese and globally interpretable traffic sign annotations. Experimental results on the TT100k, GTSDB, and TSDY datasets show that M-Ghost YOLOv8 achieves mean average precision mAP(0.50) scores of 90.5%, 98.5%, and 97.1%, respectively, outperforming recent models such as YOLOv12. Compared with YOLOv8, our model reduces parameters by 76.5%, GFLOPs by 57.5%, and doubles inference speed (up to 60 FPS). Real-world tests on the Wheeltec autonomous platform confirm reliable performance under multilingual and multi-scale conditions. M-Ghost YOLOv8 offers a scalable, edge-ready solution for intelligent transportation systems, balancing efficiency, accuracy, and language adaptability to enable safer, real-time decision-making in autonomous driving applications.