<p>Traffic signs are critical for road safety and autonomous driving. However, their detection remains challenging due to hardware limitations and the small size of many signs. Existing methods often suffer from low accuracy and poor adaptability. To address these issues, we propose the hybrid channel feature fusion based you only look once framework (HCF-YOLO), a novel traffic sign detection model. HCF-YOLO incorporates a hybrid channel feature fusion module (HCFF) that efficiently combines channel, spatial, local and global information to enhance feature representation with low computational cost. An auxiliary-enhanced minimum point distance IoU (auxiliary-enhanced MPD-IoU) to improve the bounding box regression function based on auxiliary boxes and minimum point distance error is introduced to improve spatial alignment and reduce scale sensitivity. Additionally, an attention scale sequence feature fusion mechanism (ASFF) and hybrid channel feature fusion module are added to the P2 detection layer, improving multi-scale feature extraction and reducing parameter redundancy. We evaluate HCF-YOLO on the Tsinghua-Tencent 100K (TT100K), Chinese Traffic sign detection benchmark (CCTSDB), and German traffic sign detection benchmark (GTSDB) datasets. The model achieves mAP@50 scores of 79.3<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, 88.7<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 87.3<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, outperforming the attentional scale sequence fusion YOLO (ASF-YOLO) baseline by 5.0<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, 1.3<InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 3.1<InlineEquation ID="IEq6"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>. For mAP@50:95, HCF-YOLO improves performance by 3.5<InlineEquation ID="IEq7"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, 1.0<InlineEquation ID="IEq8"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 1.6<InlineEquation ID="IEq9"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1767_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, respectively. Meanwhile, the FPS remains at 153. In general, HCF-YOLO delivers higher accuracy and efficiency with fewer parameters, showing strong potential for real-world deployment.</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

HCF-YOLO: a high-performance traffic sign detection model with hybrid channel fusion and auxiliary box regression

  • Hongge Ren,
  • Hairui Song,
  • Haiqiang Liu,
  • Anni Fan,
  • Yingying Tan

摘要

Traffic signs are critical for road safety and autonomous driving. However, their detection remains challenging due to hardware limitations and the small size of many signs. Existing methods often suffer from low accuracy and poor adaptability. To address these issues, we propose the hybrid channel feature fusion based you only look once framework (HCF-YOLO), a novel traffic sign detection model. HCF-YOLO incorporates a hybrid channel feature fusion module (HCFF) that efficiently combines channel, spatial, local and global information to enhance feature representation with low computational cost. An auxiliary-enhanced minimum point distance IoU (auxiliary-enhanced MPD-IoU) to improve the bounding box regression function based on auxiliary boxes and minimum point distance error is introduced to improve spatial alignment and reduce scale sensitivity. Additionally, an attention scale sequence feature fusion mechanism (ASFF) and hybrid channel feature fusion module are added to the P2 detection layer, improving multi-scale feature extraction and reducing parameter redundancy. We evaluate HCF-YOLO on the Tsinghua-Tencent 100K (TT100K), Chinese Traffic sign detection benchmark (CCTSDB), and German traffic sign detection benchmark (GTSDB) datasets. The model achieves mAP@50 scores of 79.3 \(\%\) % , 88.7 \(\%\) % and 87.3 \(\%\) % , outperforming the attentional scale sequence fusion YOLO (ASF-YOLO) baseline by 5.0 \(\%\) % , 1.3 \(\%\) % and 3.1 \(\%\) % . For mAP@50:95, HCF-YOLO improves performance by 3.5 \(\%\) % , 1.0 \(\%\) % and 1.6 \(\%\) % , respectively. Meanwhile, the FPS remains at 153. In general, HCF-YOLO delivers higher accuracy and efficiency with fewer parameters, showing strong potential for real-world deployment.