Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.626767
Title: Photogrammetric multi-view stereo and imaging network design
Author: Hosseininaveh Ahmadabadian, A.
ISNI:       0000 0004 5363 5069
Awarding Body: University College London (University of London)
Current Institution: University College London (University of London)
Date of Award: 2014
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Please try the link below.
Access through Institution:
Abstract:
This thesis proposes a new approach, photogrammetric multi-view stereo, for accurate and dense 3D reconstruction, including scale recovery from images. The novelty of the method can be seen in exploiting the length of the stereo camera base separation to define scale within a robust adjustment algorithm. The method is tested by imaging a series of known objects with stereo camera systems of varying quality. In each case, the baseline scaled network output is used as input into four different state-of-the-art dense matching packages in order to generate a series of dense (detailed) point clouds. Results demonstrate that networks, comprising some 50 images captured with consumer grade digital SLR cameras, can deliver 3D point data with an uncertainty around 100μm. Results are shown to be of comparable quality to a metrology grade triangulation laser scanner. Experiences gained with these stereo systems exposed a key data handling limitation in that both image capture and processing time are highly dependent on the number of camera views used. In particular, the volumes of data make the dense matching process impractical on current consumer level computing hardware. This problem demands a logical clustering and selection of the most suitable viewpoints (vantage viewpoints) from the large image dataset to provide a reduced network with similar overall capability. The second part of the research described in this thesis addresses this challenge through the development and testing of a novel methodology capable of structuring the viewpoints into clusters and then selecting vantage images resulting in more effective processing whilst ensuring a specified level of coordinate precision and point cloud completeness. The developed method is compared against the established CMVS clustering method using an in-house Imaging Network Designer (IND) suite of software developed during the course of the research. Results demonstrate that IND can provide a better image selection for subsequent dense reconstruction than CMVS in terms of completeness. The ability to select vantage images raises another research question in the feasibility of designing a complete imaging network from scratch. This issue is investigated and validated through a novel stereo imaging network design strategy. Again, this is implemented within the IND software framework and evaluated with both simulation and practical tests. In simulation tests, IND performance was tested by: 1) a comparison between spherical and ellipsoidal imaging network configurations 2) capability to select an appropriate stereo camera system to achieve a given level of point cloud precision. In practical tests, IND was evaluated using a purpose built imaging robot, INDRo, to capture images from each designed camera posture. The images were then used for accurate and dense 3D reconstruction using the photogrammetric multi view stereo method in two modes: 1) resolving scale with stereo camera base separation; 2) resolving scale with control points. The system provided this opportunity to test the effect of incidence angle, one of the key internal IND parameters, which affects the density of the network and should be adjusted with respect to the performance of feature matching algorithms. The results of this test showed that the matching algorithms work effectively for incidence angles of 10o. The results demonstrated that the final point cloud generated with the system in resolving scale with either control points or stereo camera base separation has an agreement of 0.3mm with a dataset generated with an Arius3D laser scanner. Furthermore, work on inverse robotic kinematics demonstrated the feasibility of using Particle Swarm Optimization to achieve the required camera poses with this non-standard robot.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.626767  DOI: Not available
Share: