Person Re-identification Techniques For Smart Video Surveillance Using Deep Learning

Person re-identification is the task of recognizing an individual across different cam-eras.it is a crucial capability needed by many camera surveillance systems.However,many problems still prevent the task of achieving high accuracy.In this work,we address the problem of person re-identification in many perspectives.We first propose a solution to address the problems of over-smoothness and of the lack of large scale datasets fount in person re-id.The proposed framework performs intelligent data augmentation and assigns a partial smoothing label to the generated data.It first exploits the clustering property of existing person re-id datasets to create groups of similar objects.Then,it uses each group to generate synthetic images through adversarial training.A non-uniform label distribution is finally assigned to the generated samples,and a new regularized loss function is defined for training purposes.The second solution is designed towards improving the representation learning of person re-id by introducing a self-attention mechanism coupled with cross-resolution.The proposed self-attention module reinforces the most informative parts from a high-resolution image using its internal representation at the low-resolution.The high-resolution image is used to learn a high dimensional feature representation while the low-resolution image is used to learn a filtering attention heat-map.Our self-attention module can easily be plugged in any existing person re-id model.The third solution proposes a multi-view model coupled with a new n-pair loss to eliminate the complex view discrepancy for robust person re-id.We exploited the com-plementary representation shared between views and proposed an adaptive similarity loss function to better learn a similarity metric.The three supervised method are incorporated into an ensemble learning to improve the overall re-identification process.Finally,considering the lack of annotated datasets,we propose a model that addresses the problem of unsupervised domain adaptation(UDA)in which the model learns from an unlabeled target domain using a fully annotated source domain.we introduce a tech-nique that enforces three properties:(1)a target invariance used by the target domain for supervision signal,(2)camera invariance,formed by the unlabeled target images and their camera style transferred,and(3)a hierarchical clustering optimization technique that ex-ploits the similarities between the target images to constraint the supervision information.In this thesis,the proposed models are extensively evaluated to demonstrate their performance.