面向复杂遮挡环境的双目单光子人体检测算法研究
DOI:
CSTR:
作者:
作者单位:

国防科技大学电子科学学院

作者简介:

通讯作者:

中图分类号:

基金项目:

国防科技大学自主科研基金项目


Study on Binocular Single-Photon Human Detection Algorithm for Complex Occlusion Environments
Author:
Affiliation:

College of Electronic Science and Technology National University of Defense Technology

Fund Project:

Innovation Reserch Foundation of National University of Defense Technology (Innovation Reserch Foundation of NUDT

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    面向城市反恐等高风险场景,亟需在非接触条件下对窗后人体目标实现快速、准确的姿态识别与精确定位,而现有依赖主动激光照射成像后进行图象识别算法的透窗系统难以处理复杂遮挡下的低信噪比回波数据。为此,基于双目单光子激光透窗系统,提出一种适用于目标回波分类的一维时序直方图的R-UNet网络结构。首先在前端基于目标回波的噪声特性使用轻量化的残差网络自适应地抑制背景噪声,同时保留高频信号细节。随后在骨干部分设计两级UNet结构捕捉回波信号中的多尺度时间依赖关系,确保微弱的目标信号特征得到增强。最后通过末端的全连接头将高层特征映射至标签空间完成分类任务。为验证所提方法的有效性,我们构建了一个采集自真实场景的综合数据集,包括了十种不同的目标类别。实验结果表明,提出的网络相较于基线网络在保持了96%以上的分类准确率的同时,减少了60.25%推理时间以及91.59%的参数量,更适合在边缘端进行部署。

    Abstract:

    For high-risk scenarios such as urban counter-terrorism, rapid and accurate pose recognition and precise localization of human targets behind windows under non-contact conditions are urgently required. However, existing through-window systems, which typically rely on post-imaging recognition algorithms following active laser illumination, struggle to process low signal-to-noise ratio (SNR) echo data amidst complex occlusions. To address this challenge, we propose an R-UNet architecture designed for target echo classification using one-dimensional (1D) temporal histograms, implemented within a binocular single-photon laser through-window system. First, a lightweight residual network is employed at the front end to adaptively suppress background noise based on the noise characteristics of target echoes, while simultaneously preserving high-frequency signal details. Subsequently, a two-stage UNet structure is designed within the backbone to capture multi-scale temporal dependencies in the echo signals, ensuring the enhancement of weak target signal features. Finally, a fully connected head maps high-level features to the label space to complete the classification task. To validate the effectiveness of the proposed method, we constructed a comprehensive dataset collected from real-world scenarios, comprising ten distinct target categories. Experimental results demonstrate that the proposed network achieves a classification accuracy exceeding 96%. Furthermore, compared to baseline networks, it reduces inference time by 60.25% and parameter count by 91.59%, making it highly suitable for edge deployment.

    参考文献
    相似文献
    引证文献
引用本文
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-04-17
  • 最后修改日期:2026-04-17
  • 录用日期:2026-04-17
  • 在线发布日期:
  • 出版日期:
文章二维码