中文摘要：

针对人体姿势估计算法部分节点效果差的问题,提出降低训练开销并增加用多视角深度图的后处理改进技术。首先生成一个规模减小而包含常见典型动作的训练集合,并用小型集群服务器训练后进行骨架节点的估计,之后对置信度不高的骨架节点,在深度图投影得到侧视图和顶视图中再次计算需修正的骨架节点的后处理方法以提高节点的准确度。实验表明,在使用样本数量少一个数量级的情况下能取得比微软原算法平均误差小9 mm的结果。

英文摘要：

In predicting 3D positions of human skeleton joints,the errors of some joints may be large. This paper proposed a scheme to decrease the cost of training and add post-processing using multi-view depth images to improve the performance.First the algorithm generated small image sets which contained important and typical human poses,then trained the classifier using a small cluster in the laboratory and estimated the human skeleton joints. For the joints with low confidence,it added the post-processing algorithm using the information of side view and top view depth images which could improve the precision of skeleton joints obviously. The experiments using relatively small training sets show that average error of skeleton joints is 9 mm less than Microsoft＇s original algorithm.