<p>Real-time traffic sign detection is critical for autonomous driving and intelligent transportation system safety. A key challenge in this field is balancing real-time performance enhancement, model lightweighting, and detection accuracy. To address this, we propose HAS-DETR, a real-time lightweight traffic sign detection model based on RT-DETR, with novel improvements to the backbone network, feature fusion mechanism, and small-target detection capability. Specifically, we design a High-Frequency Enhanced CSP Backbone (HFCSP-Backbone) to resolve the issues of insufficient high-frequency detail extraction and excessive parameters in traditional backbones; we introduce an Attention Scale Sequence Fusion (ASSF) module to dynamically model contextual correlations across multi-scale features for better feature representation; and we add a Small-Target Enhancement (STE) detection head embedded in the Transformer decoder (via a S2 small-scale feature layer) to mitigate small-target missed detection. Experiments on the TT100K dataset show that HAS-DETR reduces parameters by 58.99<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and computational load by 20.49<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, while improving detection performance: mAP@0.5 increases from 84.8 (RT-DETR-R18 baseline) to 86.6<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, and mAP@0.5:0.95 from 63.1 to 66.4<InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, with precision and recall reaching 90<InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 81.7<InlineEquation ID="IEq6"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>, respectively. Compared with existing methods, HAS-DETR achieves a superior balance between lightweighting and accuracy, offering an efficient solution for real-time traffic sign detection in complex scenarios.</p>

HAS-DETR: a real-time lightweight traffic sign detection model

期刊论文
发表日期：

2025年12月10日

查看全文

Junhao Dong,
Hongxiang Liao,
Xiaohui Ji

摘要

Real-time traffic sign detection is critical for autonomous driving and intelligent transportation system safety. A key challenge in this field is balancing real-time performance enhancement, model lightweighting, and detection accuracy. To address this, we propose HAS-DETR, a real-time lightweight traffic sign detection model based on RT-DETR, with novel improvements to the backbone network, feature fusion mechanism, and small-target detection capability. Specifically, we design a High-Frequency Enhanced CSP Backbone (HFCSP-Backbone) to resolve the issues of insufficient high-frequency detail extraction and excessive parameters in traditional backbones; we introduce an Attention Scale Sequence Fusion (ASSF) module to dynamically model contextual correlations across multi-scale features for better feature representation; and we add a Small-Target Enhancement (STE) detection head embedded in the Transformer decoder (via a S2 small-scale feature layer) to mitigate small-target missed detection. Experiments on the TT100K dataset show that HAS-DETR reduces parameters by 58.99 $\%$ $%$ and computational load by 20.49 $\%$ $%$ , while improving detection performance: mAP@0.5 increases from 84.8 (RT-DETR-R18 baseline) to 86.6 $\%$ $%$ , and mAP@0.5:0.95 from 63.1 to 66.4 $\%$ $%$ , with precision and recall reaching 90 $\%$ $%$ and 81.7 $\%$ $%$ , respectively. Compared with existing methods, HAS-DETR achieves a superior balance between lightweighting and accuracy, offering an efficient solution for real-time traffic sign detection in complex scenarios.