A Learning-based method for person re-identification

Oğul, Burçin Buket
Matching pedestrian images captured from different cameras is called person re-identification problem. The problem is challenging due to the low resolution of images, differences in illumination, the positional variance and possible appearance of carried objects, such as a bag, at different viewpoints. In this thesis, we investigate the discriminative ability of different features extracted from image in a binary classification framework. We finally propose a learning based method to combine different feature sets, Hue, Saturation, Value (HSV) histogram, Maximally Stable Color Regions (MSCR) and Speeded up Robust Features (SURF) matches, in a single framework. The experiments on widely used benchmark sets have shown that the best accuracy is obtained with weighted and localized histogram features. We also argue that further division of pedestrian body along the horizontal axis has the potential to increase the reidentification performance. Final integrative framework that we built outperforms the existing state-of-the-art models in terms of prediction accuracy.