基于深度图序列和时空占用模式的人类行为识别研究

针对人类行为识别问题,提出一种基于深度图序列和时空占用模式的行为识别方法。首先,使用深度图序列将空间和时间轴分为多个部分,从而定义一个4D网格;然后,将网格中的每个单元与占用值相关联,利用高维特征向量的所有单元占用值组成时空占用模式。最后,利用基于正交级别学习的主成分分析降维获得低维特征向量,并完成最终的行为识别。在公共3D行为数据集MSR上的实验结果表明,提出的离线识别方法比现有方法所获得的识别准确度较高。

The research of human behavior recognition based on depth map sequence and space occupancy patterns

For the issue that human behavior recognition, A model of depth map sequence space occupied and based on behavior recognition method is proposed. First use of depth map sequence space and time axis is divided into multiple parts, so as to define a 4D grid. Then, each cell in the grid and a use value is associated, occupancy values of high dimensional feature vector units of are used to compose space occupied patterns. Finally, dimensionality reduction is done by PCA based on OCL to get low dimensional feature vector so as to finishing behavior recognition. The experimental results on common 3D action data set MSR show that proposed off-line recognition method is better than existing methods to obtain high accuracy.