Semi-automated registration with direct linear transformation and quality evaluation of digital photogrammetry and terrestrial laser scanning

Abstract

Photogrammetry and Laser Scanning can be used to complement one another, during instances where digital images are taken of the object of interest with the intention to merge the 3D data and image in order to reconstruct photorealistic virtual models with photo quality and metric realism. Laser scanning acquires 3D data points and intensity information of objects but is unable to directly obtain photorealistic colour in most cases. To get photorealistic colour, some laser scanners come with an onboard camera, or alternatively a separate camera is used, and registration is required for both cases. One example uses a specially designed camera mounting for the laser scanner and another is to transfer colour information from 2D images captured from near the scanner to the 3D points using close-range photogrammetry. Currently limited methods exist for the registration of the data from multiple-sensors. This research outlines the evaluation and semi-automated registration of a single colour image to laser scanning point cloud data, using the canonical transformation and Direct Linear Transformation (DLT) methods for registration.The method presented in this thesis is to directly reconstruct three dimensional data from a single image with the assistance of estimated depth information. Laser scanning point cloud information is used to supplement the recovery of the estimated depth information, which is then assigned to the image data. Two primary aspects for this research are (1) the Synthetic Camera Image, following on from previous work reported in the literature on utilising synthetic imagery created from point-clouds, and (2) the Direct Linear Transformation model, which is used to provide the transformation parameters between the 2D and 3D datasets.Synthetic datasets were used to provide an indication of expected results in terms of range, incidence angle and image resolution. The image resolution is an important factor to consider. It should be as high as possible as it affects the outcome of precision. Testing with real data, the proposed method resulted in a precision of 2 mm for the data of a model T-Rex dinosaur dataset, and 19mm for a typical indoor scene. The variations in precision levels are due to different values for range, incidence angle and image resolution. Overall the results achieved the expectations producing a colour point cloud with metric assessment.