Year: 2011
Type: dissertation
Abstract: This thesis presents a 3D reconstruction software pipeline that is capable of generating \npoint cloud data from uncalibrated underwater video. This research project was undertaken \nas a partnership with 2G Robotics, and the pipeline described in this thesis will become \nthe 3D reconstruction engine for a software product that can generate photo-realistic 3D \nmodels from underwater video. The pipeline proceeds in three stages: video tracking, \nprojective reconstruction, and autocalibration. \n \nVideo tracking serves two functions: tracking recognizable feature points, as well as selecting well-spaced \nkeyframes with a wide enough baseline to be used in the reconstruction. Video tracking is accomplished \nusing Lucas-Kanade optical flow as implemented in the OpenCV toolkit. This simple and \nwidely used method is well-suited to underwater video, which is taken by carefully piloted \nand slow-moving underwater vehicles. \n \nProjective reconstruction is the process of simultaneously calculating the motion of the \ncameras and the 3D location of observed points in the scene. This is accomplished using \na geometric three-view technique. Results are presented \nshowing that the projective reconstruction algorithm detailed here compares favourably to \nstate-of-the-art methods. \n \nAutocalibration is the process of transforming a projective reconstruction, which is not \nsuitable for visualization or measurement, into a metric space where it can be used. This \nis the most challenging part of the 3D reconstruction pipeline, and this thesis presents a \nnovel autocalibration algorithm. Results are shown for two existing cost function-based \nmethods in the literature which failed when applied to underwater video, as well as the \nproposed hybrid method. The hybrid method combines the best parts of its two parent \nmethods, and produces good results on underwater video. \n \nFinal results are shown for the 3D reconstruction pipeline operating on short under- \nwater video sequences to produce visually accurate 3D point clouds of the scene, suitable \nfor photorealistic rendering. Although further work remains to extend and improve the \npipeline for operation on longer sequences, this thesis presents a proof-of-concept method \nfor 3D reconstruction from uncalibrated underwater video.
Source: UWSpace (University of Waterloo)
Author: Neil Cavan
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2591235595