Ren, Jie
Robust moving object detection by information fusion from multiple cameras.
Doctor of Philosophy thesis, University of Liverpool.
PDF
RenJie_Jan2014_19013.pdf - Author Accepted Manuscript Available under License Creative Commons Attribution No Derivatives. Download (3MB) |
Abstract
Moving object detection is an essential process before tracking and event recognition in video surveillance can take place. To monitor a wider field of view and avoid occlusions in pedestrian tracking, multiple cameras are usually used and homography can be employed to associate multiple camera views. Foreground regions detected from each of the multiple camera views are projected into a virtual top view according to the homography for a plane. The intersection regions of the foreground projections indicate the locations of moving objects on that plane. The homography mapping for a set of parallel planes at different heights can increase the robustness of the detection. However, homography mapping is very time consuming and the intersections of non-corresponding foreground regions can cause false-positive detections. In this thesis, a real-time moving object detection algorithm using multiple cameras is proposed. Unlike the pixelwise homography mapping which projects binary foreground images, the approach used in the research described in this thesis was to approximate the contour of each foreground region with a polygon and only transmit and project the polygon vertices. The foreground projections are rebuilt from the projected polygons in the reference view. The experimental results have shown that this method can be run in real time and generate results similar to those using foreground images. To identify the false-positive detections, both geometrical information and colour cues are utilized. The former is a height matching algorithm based on the geometry between the camera views. The latter is a colour matching algorithm based on the Mahalanobis distance of the colour distributions of two foreground regions. Since the height matching is uncertain in the scenarios with the adjacent pedestrian and colour matching cannot handle occluded pedestrians, the two algorithms are combined to improve the robustness of the foreground intersection classification. The robustness of the proposed algorithm is demonstrated in real-world image sequences.
Item Type: | Thesis (Doctor of Philosophy) |
---|---|
Additional Information: | Date: 2014-01 (completed) |
Subjects: | ?? TK ?? |
Divisions: | Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science |
Depositing User: | Symplectic Admin |
Date Deposited: | 01 Aug 2014 11:00 |
Last Modified: | 16 Dec 2022 04:42 |
DOI: | 10.17638/00019013 |
Supervisors: |
|
URI: | https://livrepository.liverpool.ac.uk/id/eprint/19013 |