基于深度学习的二维人体姿态估计综述

Review of 2D Human Pose Estimation Based on Deep Learning

人体姿态估计是近年来计算机视觉问题中的一个热门话题,它在改善人类生活方面具有巨大的益处和潜在的应用。近年来深度神经网络得到快速发展,相较于传统方法而言,采用深度学习的方法更能提取图像表征信息。综合分析近年来人体姿态估计的进展,根据检测人数分为单人和多人人体姿态估计。针对单人姿态估计,介绍了基于直接预测人体坐标点的坐标回归方法及基于预测人体关键点高斯分布的热图检测方法;针对多人姿态估计,采用解决多人到解决单人过程的自顶向下方法和直接处理多人关键点的自底向上方法。总结了各方法网络结构的特点和优缺点,并阐述当前面临的问题及未来发展趋势。

Human pose estimation was a hot topic in computer vision.It was of great benefit and potential in improving human life.In recent years,deep neural network has developed rapidly.Compared to traditional methods,deep learning could be used to improve extraction information from the image representation.The studies of human posture estimation were comprehensively analyzed in recent years,which could be divided into single-person and multi-person human pose estimation according to the number of people tested.For single-person pose estimation,a coordinate regression method based on direct prediction of human coordinate,and a heat map detection method based on prediction of Gaussian distribution of hu-man key points were introduced.For multi-person pose estimation,a top-down approach from solving multi-person to solving single-person process and a bottom-up approach directly dealing with multi-person key points were adopted.Finally,the characteristics,advantages and disadvantages of each method network structure were summarized,and the current problems and future development trend were expounded.