基于人体骨架的视频分析已有许多研究，主要是利用姿态估计算法或深度摄像机获取人体骨架信息，然后利用人体骨架信息对视频中的人类动作进行处理和分析。然而在实践中，由于实际应用场景下可能会出现遮挡或人群拥挤的现象，姿态估计算法或深度摄像机将不能够完全识别到视频中每个人的完整骨架，导致最终得到的人体骨架图出现关节点缺失。现有的姿态估计算法或动作识别模型并没有专门针对这种关节点缺失问题进行处理，而这种带有缺失关节点的人体骨架会对后续的人体动作识别等工作产生不良影响。因此，有必要对不完整人体骨架图的缺失关节点进行补全。围绕人体骨架关节点缺失的问题，本文展开了讨论，得到了如下主要成果： ?基于模糊c均值的补全预处理算法 该算法利用模糊c均值聚类对不完整人体骨架图的缺失关节点进行补全预处理。模糊c均值聚类能够赋予缺失关节点较为接近真实坐标值的初始值，使得图卷积网络能够利用初始值学习到更准确的关节点坐标值。 ?改进的图卷积网络 通过对原始图卷积公式中添加偏置项矩阵，扩展了参数空间，从而提高整个图卷积网络的学习能力。 ?基于模糊聚类和改进图卷积的人体骨架关节点补全模型 为了取得较为准确的不完整人体骨架图的缺失关节点坐标值补全结果，本文利用模糊c均值聚类和改进图卷积构建了一个人体骨架关节点补全模型。利用模糊c均值对不完整人体骨架图的缺失关节点进行补全预处理，得到缺失关节点坐标的初始值；然后将经过补全预处理后的人体骨架图输入改进图卷积网络中得到缺失关节点坐标值的补全结果。该模型通过补全预处理和网络学习两个部分来实现较为准确的不完整人体骨架图的缺失关节点坐标值补全。

There have been many researches on video analysis based on human skeleton. It mainly uses pose estimation algorithm or depth camera to obtain human skeleton information, and then uses human skeleton information to process and analyze human actions in video. However, in practice, due to the occlusion or crowding in the actual application scene, the pose estimate algorithm or depth camera will not be able to fully recognize the complete skeleton of each person in the video, resulting in the loss of joint coordinates in the final human skeleton graph result. The existing pose estimate algorithms or action recognition models do not specifically deal with the problem of missing joint coordinates, and the human skeleton with missing joint coordinates will have a negative impact on the subsequent work of human action recognition. Therefore, it is necessary to complete the missing joint coordinates in the incomplete human skeleton graph. Focusing on the problem of missing joint coordinates of human skeleton, this paper obtains the following main results:
?A complete preprocessing algorithm based on fuzzy c-means
The algorithm uses fuzzy c-means clustering to complete the missing joint coordinates of incomplete human skeleton graph. Fuzzy c-means clustering can give the initial value of missing joint coordinates closer to the real coordinates, so that the graph convolution networks can use the initial coordinates to learn more accurate joint coordinates.
?Improved graph convolution networks
By adding the bias matrix to the original graph convolution formula, the size of the parameter space is expanded, so as to improve the learning ability of the whole graph convolution networks.
?Human skeleton completion model based on fuzzy clustering and improved graph convolution
In order to obtain more accurate complement results of missing joint coordinates in incomplete human skeleton graph, a human skeleton completion model is constructed by using fuzzy c-means clustering and improved graph convolution. Fuzzy c-means is used to complete the missing joint coordinates in the incomplete human skeleton graph, and the initial values of the coordinates of the missing joint are obtained; Then, the human skeleton graph after completion preprocessing is feed to the improved graph convolution networks to obtain the completion result of the coordinates of the missing joints. The model completes the coordinates of missing joints of incomplete human skeleton graph accurately through two parts: completion preprocessing and network learning.