Scholarlink

<p>Accurate and efficient traffic light detection is a critical component for safe decision-making in autonomous driving systems. However, existing methods often struggle to balance detection accuracy and inference speed, and most focus only on the red, yellow, and green states while neglecting abnormal conditions. To address the challenges of traffic light detection in complex traffic environments, this paper proposes a novel model, chroma star DETR (CSDETR), built upon the RT-DETR framework. The proposed model integrates three structural innovations. First, it adopts the Chroma-StarNet backbone, which combines the chromatic gate and Multiscale-Star block to enable efficient feature extraction under lightweight settings. In addition, a color-light attentive downsampler is designed to replace conventional downsampling operations, enhancing the extraction of color features that are crucial for traffic light classification. Furthermore, by introducing the Attention-based Intra-scale Feature Interaction CloFormer (AIFI-CloFormer) module, the model significantly improves small object detection in complex traffic scenes without increasing the number of parameters. Experiments on the Bosch small traffic lights dataset (BSTLD) show that, compared with the baseline model, CSDETR improves mAP50 by 2.4%, reduces parameters by 13.1%, and achieves 345.9 FPS, fully meeting real-time detection requirements. The codes are available at <a href="https://github.com/NuChen0602/CSDETR.git">https://github.com/NuChen0602/CSDETR.git</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

CSDETR: a lightweight and efficient real-time traffic light detection model based on improved RT-DETR

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Nu Chen,
  • Shan Liu,
  • Xi Shen,
  • Ya Zhang

摘要

Accurate and efficient traffic light detection is a critical component for safe decision-making in autonomous driving systems. However, existing methods often struggle to balance detection accuracy and inference speed, and most focus only on the red, yellow, and green states while neglecting abnormal conditions. To address the challenges of traffic light detection in complex traffic environments, this paper proposes a novel model, chroma star DETR (CSDETR), built upon the RT-DETR framework. The proposed model integrates three structural innovations. First, it adopts the Chroma-StarNet backbone, which combines the chromatic gate and Multiscale-Star block to enable efficient feature extraction under lightweight settings. In addition, a color-light attentive downsampler is designed to replace conventional downsampling operations, enhancing the extraction of color features that are crucial for traffic light classification. Furthermore, by introducing the Attention-based Intra-scale Feature Interaction CloFormer (AIFI-CloFormer) module, the model significantly improves small object detection in complex traffic scenes without increasing the number of parameters. Experiments on the Bosch small traffic lights dataset (BSTLD) show that, compared with the baseline model, CSDETR improves mAP50 by 2.4%, reduces parameters by 13.1%, and achieves 345.9 FPS, fully meeting real-time detection requirements. The codes are available at https://github.com/NuChen0602/CSDETR.git.