Scholarlink

<p>We propose RTNet, a lightweight and efficient dual-branch network with cross-attention for RGB-T semantic segmentation in urban traffic scenes. We achieve a competitive trade-off between inference speed and accuracy on MFNet and FMB datasets. Specifically, RGB-T semantic segmentation technology in urban traffic scenes helps vehicles better perceive their surrounding scene information to assist decision-making. Most existing works focus on complex cross-modal feature extraction and fusion, which to some extent sacrifice the real-time performance of the network. Considering the importance of real-time capability in practical applications, we propose an efficient encoder–decoder semantic segmentation network. In the encoder, a dual-branch encoder consisting of spatial feature extraction branch (SFEB) and a global context integration branch (GCIB) is designed to extract features from four-channel RGB-T images, with Efficient Cross-Attention Transformer Modules (ECATMs) integrated to enable cross-branch interactive fusion. In the decoder, a multi-scale feature fusion module based on Efficient Multi-scale Convolutional Attention Module (EMCAM) is designed to output fused high-resolution features. Finally, an Efficient Kernel Update Segmentation Head (EKUSH) is designed to refine the final segmentation results with minimal additional computational overhead. Our code and experimental results will be released publicly at <a href="https://github.com/jayden-li05/RTNet">https://github.com/jayden-li05/RTNet</a>.</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

RTNet: a real-time dual-branch network with cross-attention for urban RGB-T semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Chenming Li,
  • Wentao Zhou,
  • Renjie Qiao,
  • Chengtao Cai,
  • Jian Zhao

摘要

We propose RTNet, a lightweight and efficient dual-branch network with cross-attention for RGB-T semantic segmentation in urban traffic scenes. We achieve a competitive trade-off between inference speed and accuracy on MFNet and FMB datasets. Specifically, RGB-T semantic segmentation technology in urban traffic scenes helps vehicles better perceive their surrounding scene information to assist decision-making. Most existing works focus on complex cross-modal feature extraction and fusion, which to some extent sacrifice the real-time performance of the network. Considering the importance of real-time capability in practical applications, we propose an efficient encoder–decoder semantic segmentation network. In the encoder, a dual-branch encoder consisting of spatial feature extraction branch (SFEB) and a global context integration branch (GCIB) is designed to extract features from four-channel RGB-T images, with Efficient Cross-Attention Transformer Modules (ECATMs) integrated to enable cross-branch interactive fusion. In the decoder, a multi-scale feature fusion module based on Efficient Multi-scale Convolutional Attention Module (EMCAM) is designed to output fused high-resolution features. Finally, an Efficient Kernel Update Segmentation Head (EKUSH) is designed to refine the final segmentation results with minimal additional computational overhead. Our code and experimental results will be released publicly at https://github.com/jayden-li05/RTNet.