Research On Person Re-Identification Key Technology In Non-Overlapping Views

Person re-identification(re-id)is an important part in the intelligent video surveillance system,which aims to match the specific person across non-overlapping camera views.Recently,person re-id is a popular research topic in computer vision.This thesis is based on domestic and foreign researches,combines the understanding of scenarios and methods,takes the improvement of the repre-sentation ability and discriminative ability of person features as the entry point,deeply studies the dictionary learning method for pedestrian re-identification in traditional methods.With the wide application of deep learning in person re-identification,The post-work of this thesis also explores the deep learning-based re-identification method.The main contributions of the thesis are as fol-lows:(1)Considering the relationship of cross-view persons,two semi-coupled discriminative dictionary learning methods are proposed.The first is coupled analysis-synthesis dictionary learning method.The analysis dictioanries are introduced to avoid the computing cost from l0 and l1 norms.A association function is constructed,which describe the characteristic of the mutual repre-sentation of person features between different camera views and improve the representation ability of synthesis dictionaries.The second is semi-coupled dic-tionary learning with relaxation label space method.This thesis uses a non-negative label relaxation matrix to construct a relaxation label space which is able to distinguish different class samples,and uses a pair of projection matrix to map the coding coefficients into the relaxation label space.This strategy not only ensure the coding coefficients of the same persons more consistent,but also enlarge the distance between the coding coefficients of different persons,so that improve the inter-class discriminative ability.In addition,Both meth-ods use local Fisher discriminant analysis to preprocess data.The experimental results demonstrate on public datasets,two methods all have better discrim-inability,effectively improve the person re-identification performance.(2)Considering the geometric characteristic of person feature space,a graph regularized and label-matched dictionary learning method is proposed.First,a regularization term is constructed in the single dictionary learning model,which makes the learned coding coefficient space preserve the local geometric structure of the original feature space.Then,a label matching term is con-structed,which further constraints the coding coefficients of the same person and make them gather more closely.The experimental results demonstrate on public datasets,compared with other methods,the proposed method achieve better person re-identification performance.(3)Considering the relathionship of person local body structures,a multi-scale local region and multi-level attention person re-id network.First,a multi-scale local feature extraction branch is constructed,use pooling and vertical segmentation to obtain the local feature maps of person image,so that the body structure features of different regions are gradually obtained from local to global.Then an attention branch is constructed to extract the attention-aware feature maps,so that the discriminative regions are captured under different se-mantic information.Last,both branches are cascade,so that obtain more robust feature representation.The experimental results demonstrate on multiple lage-scale datasets,compared with other deep learning-based methods,the proposed method achieve better person re-identification performance.