Automatic Registration of Multiple Texel Images to Form a 3-Dimensional Texel Image

Three-dimensional (3D) imagery has gained a lot of importance in today\u27s world, be it in the field of entertainment, documentation, or defense. Multiple methods for creating 3D images have been proposed in the past. A few famous methods used for 3D image matching are those that include usage of 2D images as stereo pairs or computing 3D rigid body transformations based on range information of points. The Iterative Closest Point algorithm (ICP) and its variants are well known for registration of point clouds, which can be used to create 3D surfaces. This thesis provides an algorithm, which is a continuation of the work done previously at Utah State University, to create accurate 3D images based on texel images obtained from the handheld texel camera built at USU. The first part of the thesis briefly reviews the structure and working of the handheld texel camera and the technique of creating texel images using the device and calibrating the images to mitigate the effect of lens distortions. A method is then suggested to reduce the errors in the range information in the image caused by walk error and wiggling error and also to compensate for the timing error induced in the individual pixels of the lidar sensor. A way to add a correcting factor to the range information to compensate for any oset in the origin assumed by the sensor and the actual center of perspective (COP) of the sensor is suggested in the later part of the thesis, thus correcting the images for the inaccuracies caused by the oset. The second half of the thesis brie y goes over the work previously done on 3D image matching and registration to produce 3D images. A few changes are suggested in some parts of the existing method, which use concepts of epipolar geometry in the RANSAC algorithm and use planar interpolation to accurately obtain the 3D co-ordinates of points from 2D coordinates. An iterative solution is proposed to correct erroneously chosen correspondences or reject bad correspondences to improve the rigid body transformation. The transformation thus obtained is used to compute more point matches, which are in turn used to estimate a more accurate least squares solution for the rigid body transformation. Results show that the calibration techniques and the changes implemented in the point cloud matching algorithm, suggested in this thesis, improve the accuracy of the images and produce 3D images with correct matching