MH-YOLO: A Lightweight Traffic Sign Detection Method Based on YOLOv10n with Hybrid Attention Transformer and Multi-Scale Dilated  Attention

In intelligent transportation systems, the accurate and real-time detection and recognition of traffic signs are crucial for autonomous and assisted driving. Despite improvements in efficiency and accuracy of existing deep learning object detection algorithms, challenges remain in detecting small objects, handling multi-scale targets, and achieving real-time detection in low computational resource environments. To address these challenges, we propose a lightweight YOLOv10n-based method that incorporates a Hybrid Attention Transformer (HAT) to enhance the super-resolution reconstruction of small targets, thereby improving detection accuracy. Additionally, we introduce a Multi-Scale Dilated Attention mechanism (MDA), embedded within the YOLOv10n model to capture multi-scale semantic information through self-attention mechanisms, effectively enhancing the model's detection performance. The proposed method, termed MH-YOLO, significantly improves detection effectiveness while maintaining low computational complexity. Furthermore, to address the lack of nighttime traffic sign data and region-specific sign data in existing datasets, we constructed the Chinese Traffic Sign Dataset (CTSDB), covering all time periods and various complex scenarios. Experimental results show that MH-YOLO achieves detection accuracy comparable to more complex models while maintaining model lightweightness.