Visual navigation in a conspicuously planar world

Sammanfattning: This thesis addresaee an important topic in computer vision,namely finding the ground plane and potential obstacles in theimages taken by an indoor mobile platform. This is ofimportante, since the platform then automatically can detect ifit can continue on its path. Orienting the cameras to view theground plane, generally implies that the ground plane is thedominant surface in view. The particular problem addressedhere, is to segment the images into regions correrponding tothe dominant plane, and others regions that do not. Threedifferent methods are presented in this thesis. Thefirst mehtoduses a monocular camera, in which the imagemotion has a certain structure for points corresponding toplane, varying with its orientation and position. Theorientation of the ground plane is required a priori. Theglobal 3D motion is estimated directly from the spatio-temporalderivativea of the image sequence captured by the camera,assuming an infinitesimal displacement between consecutiveframes. Using robust estimation, points corresponding toobstscles are disregarded. Obstacle points result in highresiduals, and an automatic thresholding produces a binaryimage marking potential obstacles.When the displacemcnt between consecutive frames is Iarge,the binocular camera model is more appropriate. Thesecond methoduses detected points in uncalibratedstereo images to find the dominant projectivity. The stereocorrespondence is not required a priori. A robust samplingprocedure aided by image structure, conjectures candidateprojectivities, and after an evaluation, the best candidatecorresponds to the dominant plane.Thethird methodrequires an initial estimate of thedominant projectivlty, given from a previous step, and thecorrect dominant projectivity is the global minimum of anenergy, minimized with respect to the stereo correspondence andthe projective transformation. When a robust function isincluded in the energy, to allow for obstacles, thestraight-forward two-step method is not guaranteed to converge.Instead, the energy is augmented into an auxiliary one, whosesolution is the same as for the original energy. A locallyconvergent two-step method is proposed, finding the dominantprojectivity.Keywords:computer vision, purposive vision, binocularvision, uncalibrated stereo, obstacle detection, dominant planedetection, motion estimation, stereo correspondence, projectivetransformation, auxlliary variables, convergent minimization,robust estimation