Person Re-identification Over Camera Networks

In recent years,as an emerging branch of the artificial intelligence,person re-identification has become a basic task in the video surveillance network,and increasing researchers have been attracted to study it.The main task of person re-identification is to re-identify the target in the different camera networks.It could be used in many works,such as pedestrian tracking.It has a wide range of applicability and broad space for development.However,person re-identification is challenging due to the pedestrian images collected by the cameras will be affected by low resolution,shooting angle,illumination and soon.According to the existing research,we carry out a further study and put forward three kinds of person re-identification methods.The details are as follows:(1)The camera's influencing factors were ignored in most of the existing person re-identification methods,and it would reduce the performance of the model.We propose a distance metric learning method to re-identify person,which takes into account the camera's influencing factors.Firstly,we classify the pedestrian images on the basis of the camera labels,and use the pedestrian images from different cameras or same camera to train the distance metric model respectively.Then,the weights for the corresponding models are given based on the camera's influencing factors.Finally,the final distance metric model is obtained through accumulating and optimizing the product of the model and the corresponding weight.(2)Because of containing a lot of indistinguishable data,it is hard to distinguish pedestrian images.We design a re-identification method named kernel uncorrelated local fisher discriminant analysis.We add the radial basis function and the irrelevant constraint to the original local fisher discriminant analysis method,which makes the model more discriminative.The introduction of the kernel function can deal with the indivisible data effectively while reducing the computational complexity,and the irrelevant constraint can guarantee the low correlation between the pedestrian feature representation,which is helpful to improve the model performance.(3)It is difficult to gather completely independent pedestrian images.In order toavoid overfitting,we propose a method based on regularized joint bayesian for person re-identification.We assume that each pedestrian feature representation is the summation of two independent gaussian latent variables,i.e.,intrinsic variable and intra-personal variable.That is,every feature is a gaussian mixture model.Then,we model the joint form of the two latent variables straightly via using a bayesian frame.At last,the maximum expectation algorithm and the regularization technology are used to learn the model.For our evaluations we use three different publicly available datasets(i.e.,i LIDS,3DPeS and CUHK01).Additionally,we select color and texture features as pedestrian feature representation for model training and testing.Experimental results on three proposed methods demonstrate that they could improve the correct rate of person re-identification and our research is valuable.