Person Re-identification In Video Scenes

Person re-identification refers to the process of matching snapshots of people from non-overlapping camera views at different times and places.Person re-identification has become increasingly popular in computer vision and other related fields due to its widespread application and research significance.It plays an important role in maintaining social public safety.The existing research on person re-identification can be mainly divided into two categories: feature representation and metric learning.An effective feature representation should be robust to illumination and viewpoint changes.At present,the most commonly used feature representation is to combine color features and texture features according to different requirements;metric learning is to learn a distance function that can reflect the internal relationship between data through supervision or non-supervisory methods so that the data samples can maintain good separability and achieve the purpose of improving the matching accuracy.The person re-identification based on metric learning is essentially the dimensionality reduction of high-dimensional spatial features to low-dimensional sub-spaces mapping features so as to make samples have good separability.However,in actual research,it is found that there are different disadvantages in finding suitable low-dimensional sub-spaces through different methods,such as low generalization ability and parameters optimization problems.To solve the specific deficiencies of different algorithms,the main works in this paper are as follows :1.We analyze the problem of over-fitting when top-push distance metric learning model the appearance variations of pedestrians between different cameras.We combine trace-norm regularization with top-push distance learning model and propose top-push distance learning model with trace-norm regularization(TDL-T)method.Firstly,we are concerning the relative comparision between the distance of a positive pair and the minimum distance of all related negative pairs,rather than comparing the positive pair with each of the related negative pair,thus formulating a new objective function.Secondly,the trace-norm regularization is introduced into the proposed method to obtain a learned feature projection matrixes with low rank and to relieve the over-fitting problem,thereby,it simplifies the complexity of the model.Finally,experiments on the PRID 2011 and iLIDS-VID datasets verify that generalization ability of the proposed model.2.Perturbation Linear Discriminant Analysis is introduced into the Cross-view Quadratic Discriminant Analysis(XQDA)algorithm to solve the XQDA algorithm caused by the singularity of the matrix.In the algorithm,the perturbation random vectors are introduced to learn the difference between the class empirical mean and its expection,meanwhile,the optimal metric matrix is obtained by solving the generalized Rayleigh entropy form.Finally,experiments verify the effectiveness of the parameter optimization.