Research Proposal 的模板、写作建议以及样本 计算机

Research ProposalN.MavaddatDepartment of Computer Science and Software EngineeringThe University of Western AustraliaCrawley,W.A.6009navid@.auApril2002A.Proposed Study1.TitleAutomatic Determination of Vanishing Points and Geometric Cues for Single View Rectification and Metrology.2.ContributionFrom a single two-dimensional image of a three-dimensional object or scene it is possible to extract substantial information about the original object or even partially reconstruct the image of the object from a different viewing angle.[4]This has many applications in thefields of architecture, archaeology,forensic science and many others.This analysis can provide,in the case of metrology, measurement of the original scene,and in the case of rectification,the necessary information to construct a virtual model of the scene.A critical part of the analysis process is to accurately identify attributes of interest.One such attribute of interest are the‘vanishing points’of a scene.These are defined as points at which the extensions of parallel lines appear to converge in the perspective view of the image. In practice the detection of the vanishing points is made by visual inspection of the image and is both cumbersome and not very accurate.An automatic method of the vanishing point detection would eliminate the subjective inspection of the image and expedite the process.In my Honours project I was able to advance a method for automatic detection of the vanish-ing points through the combination of image analysis techniques and projective geometry.The method was successfully applied to the detection of the vanishing points for several images and reconstruction of images from different perspectives.However,the above method can only be readily implemented when there are sufficient straight lines in the image that can be detected by computational methods and their convergence point determined.In general the two-dimensional images may contain none or a very few straight lines.Hence the proposed method has to be much improved and augmented to be a useful tool for practical applications.The attributes of interest are not confined to vanishing points.Other attributes such a horizon lines are also useful for single view image analysis.These attributes have yet to be explored and their detection processes automated.Further work includes the development of prototypes for proposed applications of single view metrology and rectification.In addition to vanishing points,other geometrical cues can greatly assist the accuracy and feasi-bility of single view metrology.These geometrical cues include determination of right angles andcircles in a scene.B.Research Plan1.Time Estimates for CompletionDate Activity05/2002Literature review and research proposal.08/2002Comparison of current vanishing point detectors with respect to efficiency and accuracy 10/2002Extension of existing techniques beyond straight lines to include horizon detection11/2002Investigation of detection of known angles.02/2003Investigation of detection of known curve shapes.05/2003Analysis of textures.01/2004Error analysis of techniques investigated04/2004Thesis Outline.05/2004Thesis Composition.10/2004Thesis Review.11/2004Final Submission.2.Project Aims•Replicate process of single view rectification and metrology without the need of camera calibration.•Improve view reconstruction by the analysis of geometric cues.•Compare accuracy and efficiency of vanishing point detection algorithms.•Investigate the accuracy of geometric feature detection with relation to human interaction in the decision making scheme.•Investigate the significance of less commonly used geometric feature in the process of image rectification.3.Existing Approaches3.1.Manual Vanishing Point DetectionManual vanishing point detection involves visually isolating pairs of lines image that are knownto parallel in the scene,and interpolating their crossing points.These crossing points define the vanishing points.An image of a three dimensional scene has a vertical vanishing point and each plane with a horizontal component defines a vanishing point,which will lie on the horizon line. Lines that are known to be parallel in the scene and that are also parallel in2d image of the scene have vanishing points that lie infinitely far away.Only the direction of these vanishing points can be computed.This approach requires manual input and therefore its accuracy is dependant on the accuracyof the measurement of parallel lines.Vanishing points are often far offthe image and hence a slight inaccuracy in the measurement of a line will result in a larger error in the position of the calculated vanishing point.3.2.Vanishing Point Detection using Hough TransformWhen lines that are parallel in world coordinates are mapped through a projective transform to an image plane,they will no longer be parallel unless the vanishing point is at infinity.These line will in fact cross and this crossing point defines the vanishing point.The Hough Transform is a widely used algorithm for determining lines and other geometrical primitives in computer vision.For the purposes of vanishing point detection,the Hough Transform is optimised in order to detect lines that would be parallel in the world coordinates of a scene. Using these parallel lines,the vanishing point of the image can be calculated[22]3.3.Gaussian SphereMany vanishing point detection algorithms required the searching of many possible vanishing point candidates.Vanishing points and other possible points of interest often lie far outside the image area.To search through these the traditional way is a time consuming and cumbersome task as it is important to test each candidate relative to the others.This means that the space to be searched is not bounded.The time taken to evaluate each vanishing point candidate increases exponentially for each addition candidate,as it has to be tested against every other candidate. For this reason it is desirable to try to transform the candidates into a bounded search area.One method for making the search space bounded is using the Gaussian sphere.The Gaussian sphere is basically a unit sphere.The lines of interest are projected on it forming circles on its surface. The intersection of the different circles are analogous to the intersection of lines.In this manner vanishing points can be determined in a bounded space and then transformed back to the image coordinates.This method introduces complexity in the calculation of intersection point,as now the intersection of circles needs to calculated instead of simply lines.[13,25,16]3.4.Project GeometryEuclidean geometry is a special case of perspective geometry.Projective geometry deals with the geometry of the perspective projection.Projective geometry is a mathematical tool with numerous properties.There are four main prop-erties that facilitatefinding vanishing points.[1]1.The cross product of two points in homogeneous coordinates results in a representation ofthe line they lie on.2.The cross product of two lines results in the point of intersection.3.Perspective,aspect ratio and sheer transformations do not cause the loss of dimensional dataas is the case in Euclidean geometry.4.Points and lines at infinity can be described and manipulated.These properties simplify the calculation of intersection points as they only require a simple cross product,and the result is kept intact even if at infinity.4.Research Method4.1.Types of Geometric CuesThe aim of geometric cue detection is to isolate features of the image that can be used to decipher the type of projective distortion it has gone through.I have identifiedfive main classifications of geometrical cues that are relevant to view reconstruction and metrology.Vanishing Points Vanishing points are primary geometrical cue used to determine projective distortion.Horizon Line Horizon lines are defined as the line formed by the vanishing points of all the planes that have a horizontal components.As well as being useful geometric cues,the horizon line can also assist in the error analysis of the vanishing points.Known-Angles If an image has a feature which forms a known angle then it is possible to determine the projection transformation it has undergone.This analysis is particularly well suited to images of architectural environments,with right angles being the most common angles to be used as cues.Simple Shapes These include triangles,quadrilaterals and simple curves and spheres. Textures Textures and repeating patterns that form surfaces can encode information about the perspective transformation they have undergone by how they have been distorted.4.2.Devising detection algorithmsThe aim of my investigation is to compare the efficiency of geometric cue detection schemes and more importantly the accuracy of these schemes.There has been a vast increase in computational power since this research in to automating the detection of vanishing points began.There are two main outcomes as a result of this.Firstly,new applications of view reconstruction and metrology that are now viable can be improved with accurate and robust vanishing point detection.Secondly, new methods of detection that require the processing of textures,for example can now been carried out within a useful period of time.4.3.Testing and analysis using generated control images.The accuracy of the geometric cue detection algorithms will be testing using artificially generated images.These images will be generated such that the position and nature of the cues are known and‘noise’in the image is controlled.This phase of testing will determine the usefulness of the detection algorithm proposed.4.4.Analysis of results using‘real world’images.After the algorithm has been validated using generated images,the real world applicability needs to be determined.Thefirst step would require testing of the algorithm with images of simple scenes where attributes of the cues can be manually determined.4.5.Applicability of detection techniques to view reconstruction and metrology The main use of geometric cues is in view reconstruction and metrology.It is important to determine the relative usefulness of the different cue detection techniques.I plan to investigate the significance of the different geometrical cues,such that more effort can be put into the accurate detection of cues so that errors in detection are not amplified when applied to practical situations.5.Duplicated WorkBoth my supervisor and I have conducted thorough searches of online indices and produced a comprehensive literature review.No material was found that duplicated this project.C.ScholarsDavid Liebowitz(dl@)Department of Engineering Science,University of OxfordOxford OX13PJ,UKAntonio Criminisi(antcrim@)Microsoft ResearchRedmond,WA98052-6399,USAAndrew Zisserman(az@)Department of Engineering Science,University of OxfordOxford OX13PJ,UKRichard Hartley(hartley@.au)Department of Systems Engineering,The Australian National UniversityCanberra ACT0200Frank A.van den Heuvel(F.A.vandenHeuval@geo.tudelft.nl)Faculty of Civil Engineering and Geosciences,Delft University of Technology2629J Delft,The NetherlandsD.BibliographyReferences[1]S.Birchfield.An introduction to projective geometry.Unpublished.Avail-able:/˜birch/projective/,1998.[2]B.Brillault-O’Mahoney.New method for vanishing point puter Vision,Graph-ics,and Image Processing.Image Understanding,54(2):289–300,September1991.[3]A Criminisi,I.Reid,and A Zisserman.A plane measuring device.Image and Vision Com-puting,17(8):625–634,1999.[4]A.Criminisi,I.Reid,and A.Zisserman.Single view metrology.In Proc.7th InternationalConference on Computer Vision,Kerkyra,Greece,pages434–442,September1999.[5]A.Criminisi,A.Zisserman,L.Van Gool,Bramble S.,and pton.A new approachto obtain height measurements from video.In Proc.of SPIE,Boston,Massachussets,USA, volume3576,1-6November1998.[6]A.Criminisi.Accurate Visual Metrology from Single and Multiple Uncalibrated Images.PhDthesis,University of Oxford,Dept.Engineering Science,December1999.D.Phil.thesis. [7]A.Dick,P.Torr,and R.Cipolla.Automatic3d modelling of architecture.Proc.11th BritishMachine Vision Conference(BMVC’00)Bristol,pages372–381,September2000.[8]O.Faugeras.Three-Dimensional Computer Vision.The MIT Press,1993.[9]I.Faux and putational Geometry for Design and Manufacture.Ellis HorwardLimited,1979.[10]R.Gonzalez and P.Wintz.Digital Image Processing.Addison Wesley,1987.[11]Visual Geometry Group.Visual geometry group web page./˜v gg.[12]P.D.Kovesi.Invariant Measures of Image Features From Phase Information.PhD thesis,The University of Western Australia,May1996.[13]J.C.H.Leung and G.F.McLean.Vanishing point matching.International Conference onImage Processing,1996.Proceedings.,pages305–308,1996.[14]D.Liebowitz,A.Criminisi,and A.Zisserman.Creating architectural models from images.In Proc.EuroGraphics,volume18,pages39–50,September1999.[15]D.Liebowitz and A.Zisserman.Metric rectification for perspective images of planes.In Proc.IEEE Conference on Computer Vision and Pattern Recognition,pages482–488,June1998.[16]E.Lutton,H.Maitre,and J.Lopez-Krahe.Contribution to the determination of vanish-ing points using hough transform.In IEEE transactions on pattern analysis and machine intelligence,volume16,pages430–438,1994.[17]M.J.Magee and J.K.Aggarwal.Determining vanishing points from perspective images.Computer Vision,Graphics,and Image Processing,26(2):256–267,May1984.[18]puter Vision IT412Lecture Notes,chapter1.University of Western Australia,1999.[19]C.Rother.A new approach for vanishing point detection in architectural environments.Pro-ceedings of the11th British Machine Vision Conference(BMVC’00),1:382–391,September 2000.[20]J.A.Shufelt.Performance evaluation and analysisof vanishing point detection techniques.In ARPA Image Understanding Workshop,pages1113–1132.Morgan Kaufmann Publish-ers,Palm Springs,1996.[21]M.Straforini,C.Coelho,and M Campani.Extraction of vanishing points from images ofindoor and outdoor scenes.In Image and Vision Computing,volume11(2),pages91–99, 1993.[22]T.Tuytelaars,M.Proesmans,and L.Van Gool.The cascaded hough transform.InternationalConference on Image Processing,1997.Proceedings,.[23]C.Tyler.The rules of perspective./CWTyler lab/CWTyler/PrePublications/ECVP1999/PerspectiveRules.html,1999.[24]Art Deparment University of Idaho.Introduction to perspective.Available at:/drawing/111/lectures/lecture04.html.[25]F.A.van den Heuvel.Vanishing point detection for architectural photogrammetry.In Inter-national Archives of Photogrammetry and Remote Sensing Vol.XXXII part5,pages652–659, 1998.[26]G.Q.Wei and Z.Y.He.Determining vanishing point and camera parameter:New ap-proaches.In Ninth International Conference on Pattern Recognition(Rome,Italy,November 14–17,1988),pages450–452,Washington,DC,puter Society Press.E.Facilities1.SupervisionProfessor Robyn Owens and Dr Peter Kovesi are available to supervise this research project. 2.Special EquipmentThis project will require a computer terminal with a consumer-level3D graphics hardware.I currently have access to a computer terminal and have gained approval to purchase the necessary graphics hardware.3.Special TechniquesNo special techniques are required for this project.4.Special LiteratureNo special literature is required for this project.5.Statistical AdviceNo statistical advice is required for this project.F.Estimated CostsNo costs other than those normally borne by the Department are anticipated.G.Confidentiality and Intellectual PropertyI intend to make the products of this research(source code etc.)freely available to the academic community.H.ApprovalsNo ethical or medical approvals are required for this project.。