Optimization of Person Re-Identifi cation Through Visual Descriptors

Abstract

Person re-identi cation is a complex computer vision task and plays a key role in providing authorities a great tool to maintain high level security. Human appearance is a critical part of information with high discriminating power. Many methods employ the combination of such types of visual features and solve one particular problem of re-identi cation. The combination of Gaussian of Gaussian and weighted histograms of overlapping stripes latest version is incorporated in the proposed approach. The two descriptors have complementary property and combination is robust to changes in light and pose. Evaluation is performed over several benchmark datasets including VIPeR, CUHK01, CAVIAR4REID, GRID, 3DPeS, iLIDS, ETHZ1 and PRID450s. Experiment and evaluation on eight datasets shows the e ectiveness of the proposed approach.