Scholarlink

<p>With the swift progress of rail transport, it is essential to have stable and efficient obstacle detection networks for rail transit to guarantee the secure operation of trains. To address the deployment challenges on mobile devices arising from the high complexity of the obstacle detection model, a lightweight obstacle detection network named LDCD-Net is proposed in this paper. It exhibits high performance for obstacle detection with a lower quantity of parameters and computational burden by using parallel spatial channel convolution, which adopts a strategy of locally parallel concatenation. This paper also designs a loss function for bounding box intersection over union regression, which resolves the challenge of accurately articulating the relative positional angle information between the predicted bounding boxes and the ground truth bounding boxes, consequently enhancing the training effectiveness of the network. The experimental results indicate that LDCD-Net achieves a mAP of 93.2% with a model complexity of 9.406 million parameters and 23.8 GFLOPs. This performance surpasses that of mainstream obstacle detection models while satisfying the requirements for deployment on mobile devices.</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

LDCD-Net: lightweight dual-channel detection network for obstacles detection on rail transit

  • 期刊论文

  • 发表日期：

    2025年2月13日

查看全文
  • Mingchao Zhang,
  • Deqiang He,
  • Qi Liu,
  • Jinxin Wu,
  • Yuliang Qin,
  • Zhenzhen Jin,
  • Chonghui Ren

摘要

With the swift progress of rail transport, it is essential to have stable and efficient obstacle detection networks for rail transit to guarantee the secure operation of trains. To address the deployment challenges on mobile devices arising from the high complexity of the obstacle detection model, a lightweight obstacle detection network named LDCD-Net is proposed in this paper. It exhibits high performance for obstacle detection with a lower quantity of parameters and computational burden by using parallel spatial channel convolution, which adopts a strategy of locally parallel concatenation. This paper also designs a loss function for bounding box intersection over union regression, which resolves the challenge of accurately articulating the relative positional angle information between the predicted bounding boxes and the ground truth bounding boxes, consequently enhancing the training effectiveness of the network. The experimental results indicate that LDCD-Net achieves a mAP of 93.2% with a model complexity of 9.406 million parameters and 23.8 GFLOPs. This performance surpasses that of mainstream obstacle detection models while satisfying the requirements for deployment on mobile devices.