视频中运动人体身份识别的认知物理学方法

A Cognitive Physics Method for Identifying Moving Human Bodies in Videos

人脸和步态数据是典型的非线性模型,现有的人脸和步态降维方法主要是线性方法,对信息损耗较大,识别率较低。为此,提出一种人体身份识别的认知物理学方法,将人脸特征和步态特征用数据场进行表征,利用数据间的相互作用和运动实现数据的自组织聚类,以非线性变换的方式实现身份特征数据的降维。用最大势函数值对降维后的样本库进行排序,实现离散点快速检测和样本检测的二分法查找。基于改进的D-S证据论对人脸和步态进行多层融合。实验结果表明,与线性变换的人脸与步态识别方法相比,该方法可提高7%的人体身份识别率,减少40%的识别时间。

Face and gait are typical nonlinear models ,but the existing methods for dimension reduction of face and gait data are movinly linear, which causes information loss and reduces the recognition rate. To sowe this problem, a cognitive physics method for human identification is presented. The facial features and gait features are described by data field. The interaction and movement among data are used to realize self-organizaed cluster which is a nonlinear conversion way for reducing the dimensions of identity feature data. After its dimension is reduced, the sample database is sorted by the maximum potential value so that the rapid detection of discrete point and birary search for sample testing are realized. The facial features and gait features are fused in multiple layers based on the improved D-S evidence theory. Experimental result shows that the presented method can improve the recognition rate by 7% and reduce the recognition time by 40% compared with linear methods.