FREE ELECTRONIC LIBRARY - Dissertations, online materials

Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 16 |

«MULTI-CAMERA SIMULTANEOUS LOCALIZATION AND MAPPING Brian Sanderson Clipp A dissertation submitted to the faculty of the University of North Carolina ...»

-- [ Page 9 ] --

Our solution method is based on the observation that a feature visible in all four cameras (corresponding to a pair of cameras at two poses) constrains the relative pose of the second stereo camera to be on a sphere around this particular feature, which has a known position relative to the first stereo camera pose from triangulation, as shown in Figure 4.1. Features seen in only the left or right camera at both poses (two-view features) are labeled S1..3, and the feature seen in all four cameras is labeled Q (for quad or four-view feature). The constraint imposed by the four-view feature leaves three degrees of freedom: two degrees for the location of the second camera on the induced sphere, and one for the rotation in the tangent plane to the sphere.

It appears natural to employ two-view correspondences visible only either in the left or right view. There is a surprising degeneracy in this setting, if e.g. the left camera of the stereo pair has the same distance from the triangulated four-view feature in both poses (which can be readily verified e.g. using Macaulay 2 (Grayson and Stillman). More importantly, this degeneracy also has an impact on the numerical accuracy of the returned solution, if the camera-point distances at both time instances do not vary significantly (which is usually the case in VSLAM settings). To avoid this degeneracy and to obtain well-conditioned solutions even in the small motion case, we use a pair of two-view correspondences in the left (or right) camera and one two-view correspondence in the other camera to solve for the remaining three degrees of freedom.

The major advantage of our approach is that it uses the total field-of-view of the stereo camera system to determine the rotation, which maximizes accuracy. Given a stereo pair with a small overlap between its views, one could triangulate 3D points in the region of overlap in the first stereo camera. And, use the three-point perspective pose solution (Har

–  –  –

the second stereo camera with respect to the first camera. Due to the small overlap, the relative rotation and translation from these methods could be inaccurate because most of the 3D points would be in approximately the same viewing direction relative to the first camera. Our method fixes the rotation with features which do not have to be in the region Figure 4.1: Geometry of stereo pair relative pose problem. Features seen in only the left or right camera are labeled Si (two-view features). The feature seen in both left and right cameras at both times is labeled Q (four-view feature).

of overlap. With this method we get a more accurate rotation with small overlap, which in turn feeds into a more accurate translation measurement.

Our solver necessitates some modification to standard RANSAC (Bolles and Fischler,

1981) to account for the differing amount of information in a two-view correspondence and a twice-triangulated four-view correspondence. We develop a method to weigh the relative information in each type of inlier to determine the best solution in the RANSAC process.

We also describe how to modify the RANSAC stopping criteria to account for two classes of measurement.

4.2 Background Estimating SfM using multi-camera systems has been a topic of recent interest in the literature. In some instances, one can take advantage of a reduced degree of freedom (DOF)

–  –  –

˚o Astr¨ m, 2004) where they theoretically derive solutions for planar SfM for many different combinations of rigidly mounted cameras, features, and multi-camera system poses. Scaramuzza et al. (Scaramuzza et al., 2009) take advantage of the fact that a vehicle moves with planar non-holonomic motion to calculate the absolutely scaled structure and motion using only one point correspondence. In contrast, our approach is to be used with stereo camera systems allowed to move in all 6DOF.

A few approaches to solve for six degree of freedom motion of a multi-camera system exist which can solve for the system’s motion without any overlapping views. Kim et al.

described one such method in (Kim et al., 2007). They first calculate essential matrices for all the system’s cameras and decompose these essential matrices into rotation and translation direction. They show that the scale of the motion can then be solved as a triangulation problem.

Another approach was presented in Chapter 3 and in Clipp et al. (2008) where the same problem is addressed using a minimal number of six feature correspondences: five from one camera are used to determine the rotation and translation direction, and the sixth one from a different camera determines the translation scale. Both methods (Clipp et al., 2008;

Kim et al., 2007) are based on decomposing the translation of the individual cameras into translation of the total system and the rotation induced translation. This can only be done if these two vectors are not parallel or close to parallel. Accordingly, these methods are only applicable to certain key-frames in any general video sequence.

The generalized epipolar constraint of multi-camera systems was developed by Pless in (Pless, 2003) to allow a network of rigidly-coupled cameras to be treated as a single imaging device for SfM. Pless suggests a linear 17-point solution for relative motion using the generalized epipolar constraint but left numerical results for this approach to future work. Li et al. show in (Li et al., 2008) that this standard linear 17-point approach to solve for relative motion using the generalized epipolar constraint based on singular value decomposition does not work in many common scenarios and propose a non-minimal, linear solution that does.

–  –  –

for the relative pose of a generalized camera and develop a solution method for the relative pose of stereo cameras from six ray correspondences. They demonstrate that up to 64 solutions may exist for the proposed constraints. Their solution is degenerate if all cameras in the multi-camera system are on a line, which of course is the case with a stereo camera (our target here).

At the other end of the spectrum are methods that require all features used in pose estimation to be in the multi-camera system’s region of overlap in the first pose. These include the before mentioned three-point perspective pose problem (Haralick et al., 1994) and the

–  –  –

not vulnerable to the pure translation degeneracy, in contrast to (Clipp et al., 2008; Kim et al., 2007), they require sufficient overlap between the cameras’ fields-of-view, necessitating a large trade-off between overlap and total field-of-view.

Our approach, using one four-view feature to fix the camera translation and three twoview correspondences to find the rotation, occupies a middle ground in the space of problems considered until now.

4.3 Solution Method In this section, we describe our approach for relative pose estimation between two poses of a stereo rig. Let P0 and P1 denote the projection matrices of the left and the right camera for the first time instance, and P0 and P1 are those for the second time instance. Without loss of generality, we assume a rectified stereo system, i.e. P0 = (I|0) and P1 = (I|b), where −b is the baseline between the cameras on the stereo rig. General configurations can be reduced to this case by appropriate rotation of the 2D feature positions (corresponding to 3D camera rays emerging from the projection center). We also assume the camera intrinsics and lens distortion parameters are known, and that feature positions in the image are undistorted and normalized according to the intrinsic parameters.

Let (R|t) denote the Euclidean transformation between the time instances, i.e.

–  –  –

The 3D point visible in both cameras has coordinates x for the first time instance and y in the second instance (always with respect to the left camera in the rig). The 3D points x and y are found through triangulating their associated correspondences in each stereo pair.


–  –  –

where we used the fact that [Rx]× Ry = (Rx) × (Ry) = R(x × y) = R[x]× y for rotation matrices R.

Overall, Equations 4.1 and 4.2 allow us to express both essential matrices in terms of the unknown rotation R, i.e.

–  –  –

In general, with two correspondences in the left camera and one correspondence in the right camera, there are three equations for the three degrees of freedom of the rotation. Using e.g.

unit quaternions to represent the rotation matrix R, a polynomial system of four equations can be formulated, which contains the three epipolar constraints (two constraint equations for the two-view correspondence in the left camera and one constraint equation for the twoview correspondence in the right camera, or one left and two right)and the unit quaternion constraint. By computing the elimination ideal (e.g. (Cox et al., 1997)), a 16th degree univariate polynomial (with only even powers) is obtained. The two solutions differing only in sign correspond to the quaternions (q0, q1, q2, q3 ) and (−q0, −q1, −q2, −q3 ), which actually represent the same rotation.

–  –  –

of equations described above (using exact arithmetic in finite prime fields, as suggested in (Stew´ nius, 2005)), and generate efficient code automatically using Buchberger’s algoe rithm to solve real instances of the pose estimation problem. In order to allow real-time processing, we utilize a root finding procedure based on Sturm bracketing instead of using one of the numerically more stable, but substantially slower approaches (e.g. (Bujnak

–  –  –

grades with decreasing baselines in terms of the 3D scene depths, which reduces its utility in real-world situations. Nevertheless, the assumption of small motion (in particular small rotations) of the camera system over time (also commonly employed in differential feature tracking methods) allows us to simplify the polynomial system as follows.

First, we represent the rotation R by modified Rodrigues parameters σ (Schaub and

–  –  –

θ/4 instead of e.g. θ/2 for the classical Rodriguez parametrization. Hence, this particular representation relaxes the assumption of small rotation angles in comparison with other representations.

Under the assumption of small rotations, we approximate R(σ) by its second order Taylor approximation and insert the resulting expression into Eqs. 4.3 and 4.4. We chose to use a second order Taylor series approximation so that our solution would work with larger rotations than a first order approximation would allow. The resulting polynomial system has three equations of degree two. The corresponding Groebner basis trace leads to an 8th degree polynomial and consists of only a few steps, hence the induced solution procedure is considered to be numerically stable. Root finding and back-substitution give the modified Rodrigues parameters σ and the corresponding rotation through Eq. 4.5, which is only an approximate solution to the original problem. The reported possible rotations are nonlinearly refined to satisfy Eqs. 4.3 and 4.4 precisely.

4.4 RANSAC Considerations The minimal solution method described in the previous section uses two modes of data points—point feature matches in 3D space, and feature correspondences in 2D. Hence, we have to deviate from the uniform treatment of samples employed in traditional RANSAC settings.

The first modification addresses the refined stopping criterion to account for the potentially different inlier ratios for the two different types of correspondences. The algorithm

–  –  –

both cameras (four-view features). Without loss of generality we assume that two temporal correspondences from the left camera and one correspondence from the right camera are utilized in the minimal solver. Then the modified stopping criterion is given by

–  –  –

where n is the number of samples required to achieve confidence α in the solution. Clearly, the term is the probability of selecting only inliers as required by our solution method lrd (Section 4.3).

The inlier scoring function for a pose hypothesis also needs adjustment. In standard RANSAC the total count of inliers is used to score the current hypothesis, which assumes that all data points contain the same amount of information. In the given setup we face a mixed set of data points consisting of 3D points visible in 4 images in total, and 2D feature correspondences. In both cases the latent variables have 3 degrees of freedom per sample—the coordinates of the underlying 3D point. However, the dimension of the observed measurement is either 4-dimensional (two 2D feature positions in either the left or the right camera) or 8-dimensional (2D positions in both cameras at both points in time).

Consequently, the residual error lies in the space orthogonal to the fitted manifold (Torr, 2002). Accordingly, the error space for 3D points visible in all cameras is 5-dimensional (8-3), and the residuals for points visible either in the left or right camera is 1-dimensional (4-3). Therefore, inlier features visible in four views carry five times more information than two view correspondences. Thus, the utilized weighting to combine the respective inlier counts is given by

–  –  –

where #Inliers(3D) denotes the number of inliers among fully visible 3D points (fourview features), and #Inliers(lef t) and #Inliers(right) designate the inlier counts of the respective two-view correspondences.

4.5 Minimal Sets of Correspondences for Stereo Cameras Our method uses a novel combination of image correspondences to determine the relative pose of a stereo camera. This leads to the question of what other combinations of correspondences can be used to solve for a stereo camera’s relative pose. Four images are available as input to compute the relative pose between stereo cameras. Thus, a point feature may be detected and matched in all four (4-V), three (3-V) or only two views (2-V).

Table 4.5 shows all of the combinations of respective feature correspondences, which result in a minimal set of constraints fully determining the relative pose of a stereo camera.

Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 16 |

Similar works:

«Der Sechste Zhva dmar pa Chos kyi dbang phyug (1584–1630) und sein Reisebericht aus den Jahren 1629/1630: Studie, Edition und Übersetzung Inaugural-Dissertation zur Erlangung des Doktorgrades der Philosophie an der Ludwig-Maximilians-Universität München vorgelegt von: Navina Lamminger Datum der mündlichen Prüfung: 28. Januar 2013 Gutachter: Prof. Dr. Franz-Karl Ehrhard Prof. Dr. Jens-Uwe Hartmann Inhaltsverzeichnis Danksagung vii 1. Allgemeine Einführung und Aufbau der Arbeit 1 1.1...»


«K-12 Education as a Hermeneutic Adventurous Endeavor: Theorizing the Instrumental Approach, Education as a Sovereign Agent for Humanity, and the Relationships with Philosophy by Doron Yosef-Hassidim A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Department of Leadership, Higher and Adult Education Ontario Institute for Studies in Education University of Toronto © Copyright by Doron Yosef-Hassidim 2015 K-12 Education as a Hermeneutic Adventurous...»

«Legal Status, Education, and Latino Youths’ Transition to Adulthood A DISSERTATION SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF THE UNIVERSITY OF MINNESOTA BY Gemma Punti IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Dr. Kendall King, Adviser June 2013 © Gemma Punti 2013 Acknowledgements First, I would like to thank the six participants, as well as their families and friends who welcomed me into their communities. They opened their homes and allowed...»

«PLURALISM, CO-OPTATION AND CAPTURE: NAVIGATING THE CIVIL SOCIETY ARENA IN THE ARAB WORLD A Dissertation submitted to the Faculty of the Graduate School of Arts and Sciences of Georgetown University in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Government By Sarah Elisabeth Yerkes, M.A. Washington, DC April 17, 2012 Copyright 2012 by Sarah Elisabeth Yerkes All Rights Reserved ii PLURALISM, CO-OPTATION AND CAPTURE: NAVIGATING THE CIVIL SOCIETY ARENA IN THE...»

«ON THE LOVE OF COMMENTARY DREAMING DEATH: THE ONANISTIC AND SELFANNIHILATIVE PRINCIPLES OF LOVE IN FERNANDO PESSOA’S BOOK OF DISQUIET Gary J. Shipley Love [.] opposes itself to identification (to knowledge) of the object, which is to say that its object is necessarily charged with a heterogeneous character (analogous to the character of the blinding sun, excrements, gold, sacred things). 1 Georges Bataille INTRODUCTION: A BAPTISMAL SLEW Fernando Pessoa had many heads, seventy or more, but was...»

«Copyright by Teresa Taylor Partridge The Dissertation Committee for Teresa Taylor Partridge Certifies that this is the approved version of the following dissertation: Infant EEG Asymmetry Differentiates Between Attractive and Unattractive Faces Committee: Judith H. Langlois, Supervisor Rebecca Bigler Catharine Echols Jennifer Beer S. Natasha Beretvas Infant EEG Asymmetry Differentiates Between Attractive and Unattractive Faces by Teresa Taylor Partridge, B. A.; M. S.; M. A. Dissertation...»

«The Unique Challenges in Counseling Infertile Individuals and Couples A DISSERTATION SUBMITTED TO THE FACULTY OF UNIVERSITY OF MINNESOTA BY Ju-Ping Huang IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Dr. Kay Herting Wahl, Adviser June 2013 © Ju-Ping Huang, 2013 Acknowledgement This dissertation is a collective effort for which many individuals should be thanked. I would like to give my hearty thanks to those who have assisted and accompanied me during this...»

«Descartes' Evil Genius O. K. Bouwsma The Philosophical Review, Vol. 58, No. 2. (Mar., 1949), pp. 141-151.Stable URL: http://links.jstor.org/sici?sici=0031-8108%28194903%2958%3A2%3C141%3ADEG%3E2.0.CO%3B2-K The Philosophical Review is currently published by Cornell University. Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have...»

«Humanities 2015, 4, 885–904; doi:10.3390/h4040885 OPEN ACCESS humanities ISSN 2076-0787 www.mdpi.com/journal/humanities Article The Ubiquity of Humanity and Textuality in Human Experience Daihyun Chung Department of Philosophy, Ewha Womans University, Seoul 120-750, South Korea; E-Mail: chungdhn@ehwa.ac.kr; Tel.: +82-10-6286-9233 Academic Editor: Albrecht Classen Received: 16 June 2015 / Accepted: 19 November 2015 / Published: 27 November 2015 Abstract: The so-called “crisis of the...»

«ASTUCnf H^vUMBryALBNCE Thesis submitted for the degree of Doctor of Philosophy at the University of Leicester by Shanta Dutta MA, MPhil (Calcutta) Department of English University of Leicester June 1996 UMI Number: U084330 All rights reserved INFORMATION TO ALL USERS The quality of this reproduction is dependent upon the quality of the copy submitted. In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if material had...»

«PHARMACOKINETICS AND PHARMACODYNAMICS OF GLYCOPYRROLATE IN THE HORSE By MARC J. RUMPLER A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2012 1 © 2012 Marc J. Rumpler 2 To Tyler and Evelyn: Whatever it is you wish to accomplish in life, may you pursue it with passion, attack it with perseverance, triumph with success and reminisce with pride. 3 ACKNOWLEDGMENTS I...»

<<  HOME   |    CONTACTS
2016 www.dissertation.xlibx.info - Dissertations, online materials

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.