{
  "paper_id": "065808cdf179ee56c545a412046d0890",
  "paper_title": "Person Identification Using Top-View Image with Depth Information",
  "paper_authors": [
    "Daichi Kouno",
    "Kazutaka Shimada",
    "Tsutomu Endo"
  ],
  "publication_date": "2013-04",
  "publication_venue": "International Journal of Software Innovation (IJSI)",
  "abstract": "In this paper, the authors describe a novel image-based person identification task. Traditional face-based person identification methods have a low tolerance for occluded situation, such as overlapping of people in an image. The authors focus on an image from an overhead camera. The authors utilize depth information for the identification task. By using depth information, the authors can capture the precise person's area and rich information for the identification task as compared with popular RGB cameras. The authors apply four features extracted from images based on depth information to the identification method; (1) estimated body height, (2) estimated body dimensions, (3) estimated body size and (4) depth histogram. In the experiment, the authors evaluated two situations; (a) standing in front of a door and (b) touching a doorknob. The identification accuracy rates are 94.4% and 91.4% on the two situations. The authors obtained the high accuracy by the proposed method.",
  "keywords": [
    "Computer Vision",
    "Depth Information",
    "Identification Tasks",
    "Person Identification",
    "Top-View Images"
  ],
  "paper_institutions": [
    "Department of Artificial Intelligence, Kyushu Institute of Technology, Iizuka, Fukuoka, Japan"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:52.616549Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文提出了一种基于俯视图像与深度信息的人物识别新方法。传统基于人脸的身份识别在遮挡场景（如人群重叠）中容错率低，而俯视摄像头能有效解决遮挡问题，并降低隐私侵犯风险。作者利用深度信息替代普通RGB图像，更精确地提取人物区域和丰富特征。方法从深度图像中提取四个特征：估计身高、身体尺寸、体型大小和深度直方图，并采用AdaBoost算法（弱分类器为C4.5决策树）进行识别。实验在两种场景下进行：站在门前（8人，每人20张，共160张图像）和触摸门把手（7人，每人15张，共105张图像），采用留一交叉验证。结果表明，结合“身高+深度直方图”的特征在两种场景下分别达到94.4%和91.4%的准确率。单特征中深度直方图表现最佳，而身高特征受人物站位影响，经归一化处理后显著提升。未来工作包括增加数据集规模、引入服饰背景信息、扩展至行走状态，并探索用于行人重识别。核心机制是利用深度值阈值分割人物区域，并提取绝对距离和相对距离特征，结合直方图描述头部与肩部距离等结构化信息。\n\n### 关键要点\n- 俯视摄像头配合深度信息解决了传统RGB图像中人体重叠导致的遮挡问题，同时避免拍摄人脸以保护隐私。\n- 方法利用Kinect获取深度图像，通过深度阈值分割出人物区域，相比RGB图像对光照变化更鲁棒。\n- 提取的四个特征包括：身高（头部最近像素与相机距离）、身体尺寸（面积）、体型大小（x和y方向最大宽度）、深度直方图（包含头部至肩部等距离分布）。\n- 采用AdaBoost作为强分类器，以C4.5决策树（Weka中的J48）作为弱分类器。\n- 站立场景共有8名受试者、每人20张图像，触摸场景共7人、每人15张，图像分辨率均为640×480。\n- 使用留一交叉验证评估，准确率计算公式为正确识别图像数除以总图像数乘以100。\n- 在站立场景中，最优特征组合“身高+深度直方图”达到94.4%准确率，与“全部特征”相同，但明显高于单特征。\n- 单特征中深度直方图效果最好（站立90.6%，触摸89.5%），而身高特征在站立场景仅52.5%，但触摸场景升至88.6%，因触摸时人体位置更稳定。\n- 身高特征受人物站位影响，经归一化（利用人物位置校正距离）后，单特征准确率从52.5%提升至74.4%，验证了归一化的重要性。\n- 体型尺寸和身体尺寸特征对服装变化敏感，而服装信息可作为上下文特征提升识别，是未来研究方向。\n- 混淆矩阵显示误识别主要发生在身高相近的个体之间，未来需引入非身体特征以区分。\n- 当前测试数据集较小，未来需要扩大规模和人数，并考虑将方法扩展至行人重识别等任务。",
  "available_templates": [
    "deep_read",
    "simple"
  ]
}