People re-identification over non-overlapping camera views

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Person re-identification is a computer vision task of recognizing an individual from similar background across non-overlapping camera views. In this research, we present an appearance-based method for people re-identification. We studied two different descriptors for characterization of people: shape based descriptors such as HOG-LBP and color based descriptors. The training and testing process are based on the pairs of images. If two images come from the same people, we consider it as positive example; however, two images represents two different individuals, they are negative examples. In order to find the best model, which can distinguish a given people, SVM and metric learning algorithms are applied for the training and testing procedure. SVM is a common classifier widely used in computer vision. However, metric learning is focus on calculate the distance and try to find a better metric A to get the better results. Cumulative match characteristic (CMC) are used to evaluate the performance. As a result, we applied two different features the HOG-LBP shape feature and the color histogram feature. And for each feature we compare the results of metric learning with support vector machine (SVM). Then we compare our results with other metric learning results.