Camera Registration in a 3D City ModelMin DingCS294-6 Final PresentationDec 13, 2006Goal: Reconstruct 3D city model usable for virtual walk- and fly-throughs• Fast• Scalable• Virtual reality• Urban planning• Simulation• Special effects• Car navigationObjectives:• Automated• PhotorealisticAerial Image Registration for airborne modelingShortcomings of the existing approach3D city model reconstruction from aerial LIDAR and oblique aerial photos aloneHigh scalability, fast acquisitionManual correspondence or extensive computation for aerial photo texture mappingAutomated texture mapping system is necessaryCamera registration algorithm overviewNeed to recover the intrinsic (focal length) and extrinsic (rotation, translation) parameters of a cameraAssume zero skew, unit aspect ratio and principal point at the image centerStamp GPS and electronic compass readings to each aerial image•obtain estimate of translation parameters and yaw angle(φ)Find focal length, pitch(θ) androll(ψ) angles from vanishing pointsRefine estimates by projecting 3D points to an image and solving point correspondences on this image•use 3D corners as feature pointsVanishing points detection – literature reviewExisting techniques look for intersections among groups of lines Expectation Maximization Algorithm [Kosecka et al. 2002]RANSAC [Aguilera 2005]Gaussian sphere / Hough transform [Barnard 1983, Shufelt1999]GPCA [Vidal et al. 2004]Perform well on indoor image or outdoor image with only a few buildings of simple geometryDifficult to apply to aerial image of complex urban scenes where multiple vanishing points existVanishing points detection – literature reviewVanishing points detection – detection algorithmIteratively find vanishing pointsDoes not require a priori knowledge of number of vanishing pointsRemove line segments in each iterationGuaranteed convergenceInitialize vanishing point to be intersection among nearly parallel linesNot pick up real intersection in 3DRefine vanishing point position with Levenberg-Marquardt algorithm at the endVanishing points detection – selection algorithm1.Fix the vanishing point with most number of segments2.Choose two other points which make the orthocenter of the formed triangle closest to the image centerAssume principal point at the image centervviivvkkvvjjvviivvkkvvjjVanishing point detection – entire processCamera calibration – intrinsic parameterStandard uncalibrated camera modelThree orthogonal vanishing points correspond to in homogenous coordinateCamera calibration – extrinsic parametersObtained R does not belong to SO(3)R’ is the closest unitary matrix to R in FrobeniusnormDecompose R’ into yaw, pitch and roll anglesR’ = Rroll*Rpitch*RyawUpdate yaw angle from GPS readingR” = Rroll*Rpitch*R’yaw3D corners detection – depth map1.Apply Harris corner detection on digital surface model (DSM)2.Label a Harris corner as a 3D corner when two sufficiently long lines intersect at a right angleFrom 299 Harris corners to 189 3D corners3D corners detection – aerial image1.Start from the end points of all the segments corresponding to the identified three orthogonal vanishing points2.Label an end point as a 3D corner if there are two sufficiently long lines converging to the other two vanishing points in a region near this end point From 1964 end points to 283 3D corners (99 are real 3D corners)Point correspondences on an image (?)Originally intended to run RANSAC to identify correspondence pairs based on the same fundamental matrixVanishing point based automatic algorithm:f = 2566.2Pitch = 50.1502°Roll = -5.0192 °Manual correspondence Lowe’s algorithm:f = 2555.7Pitch = 60.5141°Roll = -0.9834°Precision analysis in a controlled experiment… …… …Fix camera pose and rotate a calibration rigApply vanishing points based automatic calibration algorithm to find pitch and roll which should be constantpitch: 66.3° (1.3 °)roll: -14.5 °(0.5 °)pitch: 82.5° (2.2 °)roll: -2.43 °(0.17 °)Conclusions and future directionsDeveloped a fast and robust vanishing point detection for complex urban scenesExamined precision of vanishing point based camera calibrationDifficult to obtain accurate parameters just from vanishing points in a complex urban settingPossible improvementsinclude additional hardware (eg. 3-axis compass)apply stereo-vision (eg. video
View Full Document