Scholarlink

<p>Urban congestion significantly challenges emergency response times, directly affecting public safety. This paper introduces a novel hybrid deep learning architecture that integrates ConvNeXt and Vision Transformer (ViT) with focal self-attention for real-time emergency vehicle classification. The proposed model combines local feature extraction with global context modeling to overcome limitations of existing single-architecture approaches. Using a balanced dataset of 6,222 images across four vehicle categories (non-emergency, ambulance, police car, and firefighter truck), the architecture incorporates focal self-attention modules that selectively emphasize critical features such as flashing lights, rooftop bars, and vehicle contours while reducing background noise. Key innovations include a strategic 30% layer freezing approach, robust preprocessing, and a data augmentation pipeline tailored for urban traffic scenarios. The model achieved a classification accuracy of 99.26% with 99.7% precision in ambulance detection, significantly outperforming ConvNeXt (99.01%) and DenseNet201 (98.60%). Paired t-test results confirmed that the improvements over ConvNeXt (p = 0.02) and DenseNet201 (p = 0.005) are statistically significant at <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha \)</EquationSource> </InlineEquation> = 0.05. Furthermore, the model maintains computational efficiency with an inference speed of 15.8 ms per image (63 FPS) on RTX 3090, making it suitable for deployment in real-time intelligent transportation systems.</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

Hybrid convNeXt-ViT architecture with focal self-attention for emergency vehicle classification in urban traffic

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Anant Saini,
  • Jyoti Srivastava

摘要

Urban congestion significantly challenges emergency response times, directly affecting public safety. This paper introduces a novel hybrid deep learning architecture that integrates ConvNeXt and Vision Transformer (ViT) with focal self-attention for real-time emergency vehicle classification. The proposed model combines local feature extraction with global context modeling to overcome limitations of existing single-architecture approaches. Using a balanced dataset of 6,222 images across four vehicle categories (non-emergency, ambulance, police car, and firefighter truck), the architecture incorporates focal self-attention modules that selectively emphasize critical features such as flashing lights, rooftop bars, and vehicle contours while reducing background noise. Key innovations include a strategic 30% layer freezing approach, robust preprocessing, and a data augmentation pipeline tailored for urban traffic scenarios. The model achieved a classification accuracy of 99.26% with 99.7% precision in ambulance detection, significantly outperforming ConvNeXt (99.01%) and DenseNet201 (98.60%). Paired t-test results confirmed that the improvements over ConvNeXt (p = 0.02) and DenseNet201 (p = 0.005) are statistically significant at \(\alpha \) = 0.05. Furthermore, the model maintains computational efficiency with an inference speed of 15.8 ms per image (63 FPS) on RTX 3090, making it suitable for deployment in real-time intelligent transportation systems.