Pedestrian Re-identification

Pedestrian re-identification aims to identify the given target pedestrian in the photostaken from multi-camera surveillance systems. This technique can be applied in differentimportant applications, e.g. security and finding someone in the public place. However, thetime complexity of the monitoring environment causes a great challenge. Existing researchstudies mainly focus on proposing discriminative features and similarity measures. However,only little investigation discusses on complexity of labeling, and the imbalanced dataset,which are also critical issues in pedestrian re-identification.In this study, we proposed a SVM-based active learning pedestrian re-identificationmethod. Rather than learning from all the training samples, the proposed method selects themost valuable samples according to the current knowledge of the classifier. Experimentalresults show that our proposed method not only reduces the number of labeling but alsoachieves a higher accuracy with using less training samples.In addition, the number of images related to the target pedestrian is much less thannon-target images, i.e. the data is imbalance. To solve this problem, this study proposes theresampling method during the process of active learning, including oversampling andundersampling. Oversampling method used is based on SMOTE algorithm, and ideologicalunder-sampling method is removing one of the nearest two data in the dataset. Experimentalresults show that over-sampling approach than under-sampling method improve the accuracyof the identification.The results of this paper were compared with some existing pedestrian re-identificationmethod, including: RankSVM, SDALF, ELF and PLS, by comparing the CMC curve foundthat the method of this paper is much better than other methods.