单相机多人体无标记运动捕捉基于视觉信号实现多人体运动信息的重构，是国际上本领域前沿研究的热点。本项目致力于发展多人体无标记运动捕捉的理论与方法，揭示多人体无标记运动捕捉系统复杂度与运动捕捉稳定性之间的内在矛盾，突破多视角采集系统复杂以及单深度相机采集范围有限的缺陷，构建单彩色相机多人体无标记运动捕捉的平台与系统。具体的研究内容包括：分析人体姿态节点指向性的内部结构特征，提出金字塔全卷积神经网络模型，提高二维人体姿态估计的精度；建立多人体二维轮廓的图像数据集，提出多人体二维轮廓与姿态联合估计的级联全卷积网络框架，并分析人体姿态与轮廓估计之间的相互影响；探索单相机重构三维信息的不适定问题，提出时空一致性多人体三维模型优化方法，实现欠信息采集条件下的高精度多人体无标记运动捕捉。

Multi-body markerless motion capture from a single RGB camera is an attractive scheme for reconstructing the human motions with visual signals. It is one of the hottest topics in the area of computer vision and computer graphics. This project aims to develop the theory and methods of multi-body markerless motion capture, and to reveal the contradictions between the acquisition stabilities and hardware complexities. The goal of this project is to overcome the drawbacks of the complex multi-view camera array as well as the acquisition-limited depth camera, and finally to build a multi-body markerless motion capture platform with only one single RGB camera. The main contents of this project includes: Analyze the structures of the inherent human skeleton-based pose, propose a pyramid fully convolutional neural network and improve the precision of human 2D pose estimation; Build the multi-person image database with labelled 2D silhouettes, propose a cascaded fully convolutional neural network by jointly considering the silhouette and human 2D pose information and discuss the influences between these two factors; Explore the difficulties of estimating the 3D information from single image, propose a uniform time-spatial optimization method for reconstructing multi-body 3D models and accomplish the highly precise multi-body markerless motion capture results without sufficient input information.