Research On Pedestrian Re-identification Algorithm Based On Sparse Representation And Metric Learning

Due to the interference of illumination,tone and other factors,images of the same person collected from different cameras usually have great visual differences,while the images of different person may be very similar.So,it is difficult to linearly distinguish heterogeneous samples from different views.Therefore,we propose a Cross-View Kernel Collaborative Representation Classification(CV-KCRC)framework and apply it to person re-recognition by non-linear extension of the traditional Collaborative Representation Classifier(CRC),which can only be used in single-view scenes.CV-KCRC not only enhances CRC's ability to deal with the linear inseparability problem of cross-view heterogeneous samples,but also improves the model's discrimination and robustness.Secondly,considering the problem of local misalignment between images,and making up for the shortcoming that CV-KCRC can only make global matching of images,we proposed the Cross-View Local Block Metric Learning(XLBML)algorithm.XLBML uses the Graph Matching algorithm to learn the cross-view block discriminant subspace and block distance metric function between the most relevant local blocks.It reduces the distances between most related blocks of positive samples while enlarging the block distances between related blocks of negative samples,and finally obtains the block-wise distance metric to accurately match cross-view samples.The main research works of this paper are as follows:(1)The traditional CRC is extended to cross-view scenarios,and nonlinear mapping is introduced to enable it to deal with the Same Direction Distribution Problem of cross-view samples.Then,a low dimensional discriminant subspace with cross-view robustness is learned in a high dimensional nonlinear feature space,and the samples in the subspace are collaboratively represented,so as to obtain the robust cross-view-consistent collaborative representation codes of samples.(2)Aligning the local correlated blocks between the positive sample pairs by graph matching,we obtain the block correlation matrix.Then use the block correlation information to guide the local metric learning process.In this way to conduct the locally metric learning between most related blocks,which can be more accurately shorten the distances between the related blocks of positive sample pairs and enlarge the distance between the related blocks of negative sample pairs.Finally,the global distance computed by CV-KCRC and local block distance calculated by XLBML are integrated and used for the final matching.In this way to overcome the limitations of the single global model or the single local model.(3)Experiments are carried out on some commonly used person re-recognition datasets,and good performances are achieved.Experimental results demonstrate the effectiveness and superiority of the proposed algorithms.