{
  "paper_id": "edc85ec572b590ff72497c30962aa8e7",
  "paper_title": "The Progress of Human Pose Estimation: A Survey and Taxonomy of Models Applied in 2D Human Pose Estimation",
  "paper_authors": [
    "Tewodros Legesse Munea",
    "Yalew Zelalem Jembre",
    "Halefom Tekle Weldegebriel",
    "Longbiao Chen",
    "Chenxi Huang",
    "Chenhui Yang"
  ],
  "publication_date": "July 20, 2020",
  "publication_venue": "IEEE Access",
  "abstract": "Human pose estimation localizes body keypoints to accurately recognizing the postures of individuals given an image. This step is a crucial prerequisite to multiple tasks of computer vision which include human action recognition, human tracking, human-computer interaction, gaming, sign languages, and video surveillance. Therefore, we present this survey article to fill the knowledge gap and shed light on the researches of 2D human pose estimation. A brief introduction is followed by classifying it as a single or multi-person pose estimation based on the number of people needed to be tracked. Then gradually the approaches used in human pose estimation are described before listing some applications and also flaws facing in pose estimation. Following that, a center of attention is given on briefly discussing researches with a significant effect on human pose estimation and examine the novelty, motivation, architecture, the procedures (working principles) of each model together with its practical application and drawbacks, datasets implemented, as well as the evaluation metrics used to evaluate the model. This review is presented as a baseline for newcomers and guides researchers to discover new models by observing the procedure and architecture flaws of existing researches.",
  "keywords": [
    "Human pose estimation",
    "pose estimation and action recognition",
    "pose estimation survey",
    "single and multi-person pose estimation"
  ],
  "paper_institutions": [
    "School of Informatics, Xiamen University",
    "Department of Electronic Engineering, Keimyung University"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:47:06.900151Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对二维人体姿态估计这一计算机视觉基础问题进行了系统综述。该问题旨在从图像中定位人体关键点（如头、肩、肘、腕等），其准确性直接影响动作识别、人机交互、视频监控等应用。早期方法如 pictorial structures 和手工特征（HOG、颜色直方图等）受限于表达能力，泛化性能差。随着深度学习和大规模公开数据集（如 MPII、COCO、LSP、FLIC）的出现，姿态估计取得了显著进展。本文按追踪人数将姿态估计分为单人与多人两类，并详细梳理了自 DeepPose 以来的代表性模型（如 ConvNet Pose、CPM、Stacked Hourglass、DeeperCut、IEF、PAF、CPN、Simple Baselines、HRNet、CFA、OccNet、DarkPose），从骨干网络、损失函数、数据集、评价指标（如 PCK、PCKh、OKS、AP/mAP）等方面进行了比较分析。文章指出，当前主流模型多采用 ResNet 等残差结构以缓解梯度消失，且多使用 L2 损失；自上而下方法通常优于自下而上，但计算成本随人数增加；PCKh@0.5 是最常用的评估指标。最后，文章总结了遮挡、人员交互、身体结构约束等开放性挑战，并展望了未来研究方向。核心机制包括级联回归、热图预测、多尺度特征融合及坐标表示优化等，其中 DarkPose 通过泰勒展开实现亚像素精度的关节定位，显著提升了坐标解码精度。\n\n### 关键要点\n- 人体姿态估计的核心步骤是先定位关键点，再将其组合为有效的姿态配置，不同数据集（如 COCO 17 关节、MPII 14 关节）定义的关键点格式各异。\n- 早期方法采用树状图模型（pictorial structures）和手工特征（HOG、边缘等），但泛化能力差；深度学习（如 AlexNet、VGG、ResNet）显著提升了准确率。\n- DeepPose 是首个基于 CNN 的姿态估计模型，使用 AlexNet 进行关节坐标回归，但在局部区域表现不佳，后续研究转向热图预测。\n- ConvNet Pose 使用多分辨率 CNN 生成离散热图，并与图模型结合，证明了热图优于直接回归。\n- CPM（Convolutional Pose Machines）通过多阶段序列预测和中间监督学习空间关系，在 MPII 上踝关节 PCKh@0.5 比之前 SOTA 提高 10.76%。\n- Stacked Hourglass 采用重复的池化和上采样捕捉多尺度信息，并通过跳跃连接保留空间细节，在 MPII 和 FLIC 上超越先前 SOTA。\n- DeeperCut 基于 ResNet 改进 DeepCut，使用图像条件成对项和增量优化，但每图耗时约 4 分钟且不稳定。\n- PAF（Part Affinity Fields）提出自下而上的方法，同时预测关节置信图和向量场，实现实时多人姿态估计，在 COCO 和 MPII 上取得高性能。\n- CPN 利用 GlobalNet 和 RefineNet 两级结构，在线难关键点挖掘，有效处理遮挡和复杂背景，在 COCO 上超越先前方法。\n- Simple Baselines 在 ResNet 后加多个反卷积层生成热图，结构简单且达到竞争性能，验证了反卷积层数越多热图分辨率越高。\n- HRNet 全程保持高分辨率表示，通过并行多尺度子网络反复融合，避免低分辨率恢复，无需中间热图监督，在 COCO 和 MPII 上表现优异。\n- DarkPose 针对热图坐标编解码问题，提出分布感知解码和泰勒展开亚像素定位，在 COCO 和 MPII 上显著提升精度。"
}