Scholarlink

<p>Crowd counting in dense scenarios poses significant challenges due to the overlap of multiple Gaussian patches in density maps. This paper introduces a Two-Stage Crowd Counting Network with Fusion of Location Probability and Density (TSCCN-LPD) to address this issue. The first stage, the Position Probability Analytic Network(PPANet), transforms crowd images into position probability maps using a feature extraction network and a residual attention module. The second stage, the Probabilistic-Density Dual-Information Fusion Counting Network (PDFNet), combines detailed location information from the position probability map with overall crowd density information to improve counting accuracy. PPANet and PDFNet employ distinct loss functions: a Blended Error Loss for PPANet and a Mean Squared Error Loss for PDFNet. Extensive experiments on three public datasets—ShanghaiTech, UCF_CC_50 and UCF-QNRF demonstrate that TSCCN-LPD achieves state-of-the-art performance, outperforming existing methods, particularly in high-density scenarios. The introduction of the position probability map as an intermediate representation effectively mitigates the impact of Gaussian blob overlap, enhancing the robustness and accuracy of crowd counting. The codes can be available at <a href="https://github.com/YeningChen/CrowdCountingTwoStage.git">https://github.com/YeningChen/CrowdCountingTwoStage.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

Two-stage crowd counting network integrating positional probabilities and densities

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Jianwei Li,
  • Yening Chen,
  • Shaozhuo Xie,
  • Yao Dai,
  • Jingyu Lai

摘要

Crowd counting in dense scenarios poses significant challenges due to the overlap of multiple Gaussian patches in density maps. This paper introduces a Two-Stage Crowd Counting Network with Fusion of Location Probability and Density (TSCCN-LPD) to address this issue. The first stage, the Position Probability Analytic Network(PPANet), transforms crowd images into position probability maps using a feature extraction network and a residual attention module. The second stage, the Probabilistic-Density Dual-Information Fusion Counting Network (PDFNet), combines detailed location information from the position probability map with overall crowd density information to improve counting accuracy. PPANet and PDFNet employ distinct loss functions: a Blended Error Loss for PPANet and a Mean Squared Error Loss for PDFNet. Extensive experiments on three public datasets—ShanghaiTech, UCF_CC_50 and UCF-QNRF demonstrate that TSCCN-LPD achieves state-of-the-art performance, outperforming existing methods, particularly in high-density scenarios. The introduction of the position probability map as an intermediate representation effectively mitigates the impact of Gaussian blob overlap, enhancing the robustness and accuracy of crowd counting. The codes can be available at https://github.com/YeningChen/CrowdCountingTwoStage.git.