{
  "paper_id": "51a467a3f22162b44b7872f7998834e4",
  "paper_title": "Convolutional Networks for Semantic Heads Segmentation using Top-View Depth Data in Crowded Environment",
  "paper_authors": [
    "Daniele Liciotti",
    "Marina Paolanti",
    "Rocco Pietrini",
    "Emanuele Frontoni",
    "Primo Zingaretti"
  ],
  "publication_date": "",
  "publication_venue": "2018 24th International Conference on Pattern Recognition (ICPR)",
  "abstract": "Detecting and tracking people is a challenging task in a persistent crowded environment (i.e. retail, airport, station, etc.) for human behaviour analysis of security purposes. This paper introduces an approach to track and detect people in cases of heavy occlusions based on CNNs for semantic segmentation using top-view depth visual data. The purpose is the design of a novel U-Net architecture, U-Net3, that has been modified compared to the previous ones at the end of each layer. In particular, a batch normalization is added after the first ReLU activation function and after each max-pooling and up-sampling functions. The approach was applied and tested on a new and public available dataset, TVHeads Dataset, consisting of depth images of people recorded from an RGB-D camera installed in top-view configuration. Our variant outperforms baseline architectures while remaining computationally efficient at inference time. Results show high accuracy, demonstrating the effectiveness and suitability of our approach.",
  "keywords": [
    "semantic segmentation",
    "top-view depth data",
    "crowded environment",
    "U-Net",
    "CNN",
    "people detection",
    "tracking"
  ],
  "paper_institutions": [
    "Department of Information Engineering (DII), Università Politecnica delle Marche"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:54.815410Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本论文针对拥挤环境（如零售店、机场、车站等）中人员检测与跟踪面临的重遮挡和动态复杂背景挑战，提出了一种基于卷积神经网络（CNN）的语义分割方法，利用顶视深度数据进行人头分割。研究背景包括RGB-D相机在应对严重遮挡和复杂背景中的优势，以及语义分割在场景理解中的重要性。该方法设计了一种新型U-Net架构（U-Net3），在每层末端添加批量归一化（首次ReLU后以及每次池化和上采样后），以改进分割性能。实验使用新采集的TVHeads数据集，包含1815张顶视RGB-D深度图像，由6位标注员手工标注真实标签，场景密度超过3人/平方米。数据集提供16位和8位两种深度图像，用于比较不同输入。实现设置包括学习率0.001、Adam优化器、200轮训练，并与SegNet、ResNet、FractalNet、U-Net、U-Net2等基线进行对比。评估指标包括Jaccard指数、Dice系数、准确率、精确率、召回率和F1分数。结果显示U-Net3（8位）在Jaccard（0.9625）和准确率（0.9946）上最优，FractalNet16位在精确率、召回率和F1上略高，但U-Net3整体表现均衡且推理效率高（单核CPU上10fps）。研究结果表明该方法的有效性和适用性，未来将扩大数据集并与其他CNN比较，收集更复杂场景。核心机制为U-Net3的架构改进，即在编码-解码路径中通过批量归一化增强特征归一化和训练稳定性。\n\n### 关键要点\n- 问题背景：拥挤环境中多人检测与跟踪面临持续遮挡、外观变化和复杂背景等挑战，传统监控技术难以处理。\n- 传感器选择：RGB-D相机因可靠性、可用性和低成本被广泛采用，深度信息有助于应对遮挡。\n- 方法概述：提出基于CNN的语义分割方法，利用顶视深度数据进行人头分割，属于首次将顶视配置RGB-D数据用于拥挤环境人员检测。\n- U-Net3架构：在U-Net基础上修改，在首个ReLU后及每次最大池化和上采样后添加批量归一化，改进特征学习。\n- 数据集：TVHeads数据集包含1815张16位深度图（320×240），并转换为8位以突出头部轮廓，由6位标注者人工标注，采用多数投票策略。\n- 实验设置：学习率0.001、Adam优化器、训练200轮，输入为16位和8位深度图像，并与SegNet、ResNet、FractalNet、U-Net、U-Net2对比。\n- 结果指标：使用Jaccard、Dice、准确率、精确率、召回率、F1分数评估；U-Net3 8位在Jaccard（0.9625）和准确率（0.9946）上最佳。\n- 性能对比：FractalNet 16位在精确率（0.9927）、召回率（0.9933）和F1（0.9930）上略优，但U-Net3整体表现均衡。\n- 实时性：系统在单核CPU上达到10fps，满足实时应用需求。\n- 适用约束：方法不要求人员初始孤立，只需头部可见；假设相机垂直向下且地面已知，无阴影反射干扰。\n- 定性结果：U-Net3产生的分割形状更精确，而U-Net和SegNet产生平滑但可能失真的头部形状。\n- 未来方向：计划扩大数据集、与其他CNN架构比较，并收集更复杂场景以增强挑战性。"
}