目的 随着监控摄像头的增多,基于多摄像头的智能分析显得很重要.基于此,提出一种新的基于特征变换和数据集分块的行人比对方法.方法 首先提出一种新的基于变换矩阵来消除特征差异的算法.这种算法能够在高维空间中,通过变换矩阵,让某特征向量逼近另一特征向量,从而消除了同一行人特征间的差异.与此同时,还提出一种新的将行人数据集中特征分块的算法,使每个分块中的行人特征具有相似的性质,从而能够共用某个变换矩阵,从而能更好地消除同一行人在不同镜头下的特征差异.结果 基于VIPeR数据集和复杂街道场景数据集设计了行人比对实验.实验结果表明,本文的比对方法具有较高的比对准确率,VIPeR数据集(50%训练,50%检测)的Rank1的比对准确率为22%.同时本文设计了特征变换和数据集分块这2个模块的对照实验.实验结果表明,特征变换和数据集分块模块对结果都有提升的效果.结论 本文新的行人比对方法通过恰当的特征变换让同一行人在多镜头下的特征互相接近.实验结果表明该方法能够较好地在多镜头下匹配行人.

Objective In recent years, a number of surveillance cameras have been placed. The number of surveillance videos that need to be observed and analyzed is rising rapidly because of increased traffic road monitoring and indoor surveillance needs. Single cameras are insufficient to meet the need of certain monitoring tasks because of the significant increase in the number of outdoor cameras. The multi-camera-based intelligent analysis of human behavior is increasingly valuable in the field of video surveillance, and multi-camera-based pedestrian re-identification is a significant research direction. This paper proposes a new method of pedestrian re-identification based on feature transformation and dataset classification. Method In this method, a new pedestrian re-identification algorithm is proposed based on the transformation matrix that can massively eliminate the differences of the features. This algorithm can be used in high-dimensional space,in which a vector approaches another vector through the transformation matrix.Therefore, this algorithm is capable of eliminating the feature differences of the same pedestrian. This paper also proposes a new algorithm based on the pedestrian feature classification of datasets; the pedestrian features of each class have similar properties.Thus, these features can share the same transformation matrix, and the algorithm can eliminate the feature differences of a pedestrian under multi-camera. In detail, blocks are built by clustering based on block features. The transformation matrix in each block is trained based on the multi-channel features under different cameras. Each block has a corresponding transformation matrix.The correct block for each testing pedestrian can be determined by comparing that block to that of the training pedestrians. Thus,the transformation matrix of the chosen block is the appropriate matrix. The chosen transformation matrix then eliminates the differences in the features of the testing pedestrian under multi-camera. Result Experimental results prove that the proposed method can improve the accuracy of finding the same pedestrian under multi-camera. Specifically, the Rank1 matching rate of the method in VIPeR dataset (50% training, 50% testing) is 22%, which is better than the results of other existing methods. The feature transformation module and the dataset classification module can match the same pedestrian under multi-camera. The matching rate of pedestrian can be reduced significantly if one of the two modules is eliminated. Moreover, the experiments are designed to verify the robustness characteristics of the chosen feature. The experimental results based on real street scene dataset show that the chosen feature is more out standing and robust than some of the existing features. Conclusion This paper proposes a new method of pedestrian re-identification based on feature transformation and dataset classification. An appropriate transformation matrix can eliminate the influences of illumination and pedestrian poses under multi-camera. Thus, the method can make the features of the same pedestrian close to each other. The method can also significantly improve the accuracy of matching pedestrians under multi-camera.