Abstract: The purpose of 3D human pose estimation is to predict information such as the 3D coordinate position and angle of human joint points, and construct human representations (such as human bones) for further analysis of human posture. With the continuous advancement of deep learning methods, more and more high-performance 3D human pose estimation methods based on deep learning have been proposed. However, due to the human occlusion of the picture and the large demand for training scale, there are still challenges in 3D human pose estimation. The research purpose of this paper is to review a number of research papers in recent years, analyze and compare the reasoning process and core elements of these methods, and comprehensively elaborate the 3D human pose estimation methods based on deep learning in recent years. In addition, this paper also introduces the relevant data- sets and evaluation indicators, compares the experimental data of some models on the Human3.6M dataset, Campus dataset and Shelf dataset, and analyzes and compares the experimental results. Finally, according to the results of this survey, the difficulties and challenges faced by the current 3D human pose estimation are discussed, and the future development of 3D human pose estimation is discussed.

Key words: 3D human pose estimation, deep learning, neural networks, joints detection

摘要： 三维人体姿态估计的目的是预测出人体关节点的三维坐标位置和角度等信息，构建人体表示（如人体骨骼），以便进一步分析人体姿态。随着深度学习方法的不断推进，越来越多的基于深度学习的高性能三维人体姿态估计方法被提出。然而由于图片的人体遮挡、训练规模需求较大等原因，三维人体姿态估计仍然存在挑战。该研究目的是通过对近年来的多篇研究论文进行回顾，分析和比较这些方法的推理过程和核心要素，从不同输入的角度入手，全面阐述近年来基于深度学习的三维人体姿态估计方法。此外，还介绍了相关数据集和评价指标，在Human3.6M、Campus和Shelf数据集上对部分模型进行实验数据比对，分析对比实验结果。最后，根据本次调查的结果，讨论目前三维人体姿态估计所面临的困难和挑战，对三维人体姿态估计的未来发展进行了探讨。

关键词: 三维人体姿态估计, 深度学习, 神经网络, 关键点检测