Analysing Large-scale Surveillance Video



Zhou, Y
(2018) Analysing Large-scale Surveillance Video. PhD thesis, University of Liverpool.

[img] Text
200891265_Jul2018.pdf - Unspecified

Download (33MB)

Abstract

Analysing large-scale surveillance video has drawn signi cant attention because drone technology and high-resolution sensors are rapidly improving. The mobility of drones makes it possible to monitor a broad range of the environment, but it introduces a more di cult problem of identifying the objects of interest. This thesis aims to detect the moving objects (mostly vehicles) using the idea of background subtraction. Building a decent background is the key to success during the process. We consider two categories of surveillance videos in this thesis: when the scene is at and when pronounced parallax exists. After reviewing several global motion estimation approaches, we propose a novel cost function, the log-likelihood of the student t-distribution, to estimate the background motion between two frames. The proposed idea enables the estimation process to be e cient and robust with auto-generated parameters. Since the particle lter is useful in various subjects, it is investigated in this thesis. An improvement to particle lters, combining near-optimal proposal and Rao-Blackwellisation, is discussed to increase the e ciency when dealing with non-linear problems. Such improvement is used to solve visual simultaneous localisation and mapping (SLAM) problems and we call it RB2-PF. Its superiority is evident in both simulations of 2D SLAM and real datasets of visual odometry problems. Finally, RB2-PF based visual odometry is the key component to detect moving objects from surveillance videos with pronounced parallax. The idea is to consider multiple planes in the scene to improve the background motion estimation. Experiments have shown that false alarms signi cantly reduced. With the landmark information, a ground plane can be worked out. A near-constant velocity model can be applied after mapping the detections on the ground plane regardless of the position and orientation of the camera. All the detection results are nally processed by a multi-target tracker, the Gaussian mixture probabilistic hypothesis density (GM-PHD) lter, to generate tracks.

Item Type: Thesis (PhD)
Divisions: Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User: Symplectic Admin
Date Deposited: 21 Nov 2018 14:30
Last Modified: 19 Jan 2023 01:30
DOI: 10.17638/03024330
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/3024330