Re-identifying people in the crowd

Riachy, Chirine (2019) Re-identifying people in the crowd. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral Thesis) riachy.chirine_phd.pdf - Submitted Version Download (18MB) | Preview

Abstract

Developing an automated surveillance system is of great interest for various reasons including forensic and security applications. In the case of a network of surveillance cameras with non-overlapping fields of view, person detection and tracking alone are insufficient to track a subject of interest across the network. In this case, instances of a person captured in one camera view need to be retrieved among a gallery of different people, in other camera views. This vision problem is commonly known as person re-identification (re-id).

Cross-view instances of pedestrians exhibit varied levels of illumination, viewpoint, and pose variations which makes the problem very challenging. Despite recent progress towards improving accuracy, existing systems suffer from low applicability to real-world scenarios. This is mainly caused by the need for large amounts of annotated data from pairwise camera views to be available for training. Given the difficulty of obtaining such data and annotating it, this thesis aims to bring the person re-id problem a step closer to real-world deployment.

In the first contribution, the single-shot protocol, where each individual is represented by a pair of images that need to be matched, is considered. Following the extensive annotation of four datasets for six attributes, an evaluation of the most widely used feature extraction schemes is conducted. The results reveal two high-performing descriptors among those evaluated, and show illumination variation to have the most impact on re-id accuracy.

Motivated by the wide availability of videos from surveillance cameras and the additional visual and temporal information they provide, video-based person re-id is then investigated, and a su-pervised system is developed. This is achieved by improving and extending the best performing image-based person descriptor into three dimensions and combining it with distance metric learn-ing. The system obtained achieves state-of-the-art results on two widely used datasets.

Given the cost and difficulty of obtaining labelled data from pairwise cameras in a network to train the model, an unsupervised video-based person re-id method is also developed. It is based on a set-based distance measure that leverages rank vectors to estimate the similarity scores between person tracklets. The proposed system outperforms other unsupervised methods by a large margin on two datasets while competing with deep learning methods on another large-scale dataset.

Actions (login required)

View Item

Downloads