Naskou Athanasios, Konstantinos
(2018)
Photogrammetric computer vision methods for the automatic 3D reconstruction of virtual sports environments.
PhD thesis, University of Nottingham.
Abstract
The scan-to-BIM conversion of 3D point clouds of infrastructure works is essential for their as-built documentation and maintenance; automating this tedious and labour-intensive task is a challenging civil engineering problem. Related work on the subject has provided remarkable results with respect to the automatic detection and 3D modelling of planar, spherical, and cylindrical shapes in laser scans of buildings or industrial sites, e.g. walls and pipes. However, long and narrow spaces comprising free-form surfaces with reflective properties unfavourable to optical measurements pose additional challenges, especially when the precise definition of their boundaries is also required. Image-based 3D scanning techniques, in particular, are more liable to yield noisy and incomplete point clouds that include outliers and clutter, which complicate the 3D modelling process.
This thesis is concerned with the dense 3D reconstruction and parametric modelling of objects with a characteristic elongated geometry that can be represented as generalised cylinders, such as race tracks or roadways. The study is focused on a piece of infrastructure which is essential to a sports engineering application. The aim was to explore the potential of designing a robust workflow for the extraction of the 3D model of an approximately 2.5×1500 m long artificial combined bobsleigh, skeleton, and luge sliding track, which can be carried out by non-specialists in geomatics processes. The geometry of the track’s concrete structure and ice layer is of interest, so that it can be used in computer simulations for optimal trajectory identification and track construction verification or within a virtual reality system as an aid to both amateur and elite athletes for site familiarisation, training, performance enhancement and, more importantly, accident prevention.
The emerging field of infrastructure computer vision provides several tool for this purpose, based on the application of computer vision methods for the collection and analysis of real world data in order to convert it to useful information. Some of the problems that it aims to solve overlap with the objectives of this thesis: the provision of guidance for proper data collection, combining existing information with human input and sensor output, separating signal from noise in photogrammetric measurements, compensating for missing data due to the difficulties inherent in dealing with certain materials and surface types, resolving the scale ambiguity in generating 3D information from monocular image-based reconstruction methods, as well as, point cloud segmentation and geometric shape detection and fitting.
The first part of the research that is presented in this thesis is concerned with photogrammetric network design simulations in search of an imaging configuration simple enough to be implemented on site with a hand-held camera and compact enough to allow for image acquisition within the time allocated for a typical track walk. In addition, the potential of facilitating the dense reconstruction process by automating an interactive graph-based image segmentation algorithm using 3D information from the sparse reconstruction phase is investigated. In the second part, an automatic clothoid spline fitting method for horizontal alignment extraction is proposed, in order to support a sweeping-and-morphing algorithm that detects the across-track boundaries of the track’s surface and fills in missing information along the track’s length.
The main hypothesis was that the image-based 3D scanning of the entire length of an ice track can be successfully completed within the limited time frame of a typical track walk by a single person with a hand-held, high-resolution, non-metric, pre-calibrated DSLR camera with a moderately wide prime lens and that the image orientation could be retrieved automatically, without the requirement for scene signalisation or manual image observations, by low-cost software. The process should allow for the parametric 3D model of the track’s surface to be extracted from the generated 3D mesh of the site without requiring user interaction. The proposed techniques were validated on two track sites using data from image-based and laser scanning surveys. It was concluded that this scenario is feasible but the integrity of the result and the repeatability of the process are not guaranteed. Subdecimetre accuracy can be achieved for the track’s sparse reconstruction if ground control points are measured, with the typical GNSS RTK technique accuracy, at 100–150 m intervals along the track and used in a constrained self-calibrating bundle adjustment. If only 3 ground control points are available, submetre accuracy is achievable; a result at this accuracy level can be expected from the Zeb1 hand-held laser scanner within the same time frame but with a potential to greatly simplify the 3D modelling process, yet complicate the generation of an immersive virtual environment in terms of visual fidelity, unless it is combined with an image-based scanning process in order to retrieve colour information. The proposed sweeping-and-morphing algorithm can provide an initial solution for the parametric 3D model of the track’s surface without requiring user interaction but the result needs refining, particularly around the edges along the transitions in and out of the bends. Specifically for a track that was scanned in the iced state, and especially if the 3D model is to be used in a virtual reality system, sliding simulations for the verification of the result are required.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)