Dynamic surface reconstruction from a sequence of video images
Dynamic surface reconstruction from a sequence of video images
Computer vision algorithms are widely used for sensing in autonomous vehicles. DROID is a structure from motion vision algorithm developed by Roke Manor Research. This algorithm processes a sequence of images, extracts conspicuous point features and provides estimates of the three dimensional feature position and vehicle motion. These features are then used to reconstruct a surface representation of a vehicle's environment using planar triangular facets. Unfortunately the use of these facets results in temporal discontinuities being present in the surfaces. When viewed over several images, local areas of the surface can be seen to change markedly between frames, although no physical evidence exists for such marked differences.
This thesis is an investigation of and suggests some solutions to the problem of these temporal discontinuities. The reconstructed surface changes when the underlying triangulation changes. Examination of the triangulation between frames shows that changes are not localised or bounded across the image. Use of a simulation of vehicle motion over a horizontal plane of points shows that triangulation changes occur when points enter and leave the data set or when their image plane positions change relative to each other.
The offered solution to this problem is the use of complementary static and dynamic surface reconstruction. Firstly an improved static surface reconstruction can improve accuracy. Secondly a weighted averaging of the surfaces over time can introduce smoothness. Polynomial based static surface reconstruction were rejected as being too expensive in processing terms and poor suited to the ill constrained nature of the data set. Data Dependent triangulation appeared to offer the advantages of triangulation with improved accuracy but was disappointing. Dynamic surface reconstruction qualitatively appears to smooth the temporal discontinuities between images through use of a grid of Kalman filters sampling image plane depths at regular intervals across the image.
University of Southampton
1997
Blackwell, Ian Richard
(1997)
Dynamic surface reconstruction from a sequence of video images.
University of Southampton, Doctoral Thesis.
Record type:
Thesis
(Doctoral)
Abstract
Computer vision algorithms are widely used for sensing in autonomous vehicles. DROID is a structure from motion vision algorithm developed by Roke Manor Research. This algorithm processes a sequence of images, extracts conspicuous point features and provides estimates of the three dimensional feature position and vehicle motion. These features are then used to reconstruct a surface representation of a vehicle's environment using planar triangular facets. Unfortunately the use of these facets results in temporal discontinuities being present in the surfaces. When viewed over several images, local areas of the surface can be seen to change markedly between frames, although no physical evidence exists for such marked differences.
This thesis is an investigation of and suggests some solutions to the problem of these temporal discontinuities. The reconstructed surface changes when the underlying triangulation changes. Examination of the triangulation between frames shows that changes are not localised or bounded across the image. Use of a simulation of vehicle motion over a horizontal plane of points shows that triangulation changes occur when points enter and leave the data set or when their image plane positions change relative to each other.
The offered solution to this problem is the use of complementary static and dynamic surface reconstruction. Firstly an improved static surface reconstruction can improve accuracy. Secondly a weighted averaging of the surfaces over time can introduce smoothness. Polynomial based static surface reconstruction were rejected as being too expensive in processing terms and poor suited to the ill constrained nature of the data set. Data Dependent triangulation appeared to offer the advantages of triangulation with improved accuracy but was disappointing. Dynamic surface reconstruction qualitatively appears to smooth the temporal discontinuities between images through use of a grid of Kalman filters sampling image plane depths at regular intervals across the image.
This record has no associated files available for download.
More information
Published date: 1997
Identifiers
Local EPrints ID: 460201
URI: http://eprints.soton.ac.uk/id/eprint/460201
PURE UUID: 464a08a0-7901-4219-9366-8d8b557e7f9b
Catalogue record
Date deposited: 04 Jul 2022 18:11
Last modified: 04 Jul 2022 18:11
Export record
Contributors
Author:
Ian Richard Blackwell
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics