The Library
Scalable deep feature learning for person re-identification
Tools
Lin, Shan (2019) Scalable deep feature learning for person re-identification. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Lin_2019.pdf - Submitted Version - Requires a PDF viewer. Download (20Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3474416~S15
Abstract
Person Re-identification (Person Re-ID) is one of the fundamental and critical tasks of the video surveillance systems. Given a probe image of a person obtained from one Closed Circuit Television (CCTV) camera, the objective of Person Re-ID is to identify the same person from a large gallery set of images captured by other cameras within the surveillance system. By successfully associating all the pedestrians, we can quickly search, track and even plot a movement trajectory of any person of interest within a CCTV system. Currently, most search and re-identification jobs are still processed manually by police or security officers. It is desirable to automate this process in order to reduce an enormous amount of human labour and increase the pedestrian tracking and retrieval speed. However, Person Re-ID is a challenging problem because of so many uncontrolled properties of a multi-camera surveillance system: cluttered backgrounds, large illumination variations, different human poses and different camera viewing angles.
The main goal of this thesis is to develop deep learning based person reidentification models for real-world deployment in surveillance system. This thesis focuses on learning and extracting robust feature representations of pedestrians. In this thesis, we first proposed two supervised deep neural network architectures. One end-to-end Siamese network is developed for real-time person matching tasks. It focuses on extracting the correspondence feature between two images. For an offline person retrieval application, we follow the commonly used feature extraction with distance metric two-stage pipline and propose a strong feature embedding extraction network. In addition, we surveyed many valuable training techniques proposed recently in the literature to integrate them with our newly proposed NP-Triplet xiii loss to construct a strong Person Re-ID feature extraction model. However, during the deployment of the online matching and offline retrieval system, we realise the poor scalability issue in most supervised models. A model trained from labelled images obtained from one system cannot perform well on other unseen systems. Aiming to make the Person Re-ID models more scalable for different surveillance systems, the third work of this thesis presents cross-Dataset feature transfer method (MMFA). MMFA can train and transfer the model learned from one system to another simultaneously. Our goal to create a more scalable and robust person reidentification system did not stop here. The last work of this thesis, we address the limitation of MMFA structure and proposed a multi-dataset feature generalisation approach (MMFA-AAE), which aims to learn a universal feature representation from multiple labelled datasets. Aiming to facilitate the research towards Person Re-ID applications in more realistic scenarios, a new datasets ROSE-IDENTITY-Outdoor (RE-ID-Outdoor) has been collected and annotated with the largest number of cameras and 40 mid-level attributes.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > Q Science (General) T Technology > TA Engineering (General). Civil engineering (General) T Technology > TK Electrical engineering. Electronics Nuclear engineering |
||||
Library of Congress Subject Headings (LCSH): | Biometric identification, Human face recognition (Computer science), Machine learning, Video surveillance | ||||
Official Date: | September 2019 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Computer Science | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Li, Chang-Tsun | ||||
Sponsors: | Horizon 2020 (Programme) ; Nanyang Technological University | ||||
Format of File: | |||||
Extent: | xxi, 158 leaves | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year