Site Prediction |
The recording of Bet Giorgis, a 12th century Rock-Hewn church in Ethiopia
Conduct of the Project
Three fundamental data sources were involved in the Bet Giorgis documentation. The first of these was existing 1:10,000 scale aerial photography from which ortho-imagery and a DTM of the Lalibela area and church surroundings could be generated. The second was ground survey data to facilitate control for the aerial imagery and to provide a tie between the local XYZ reference coordinate system used in the close-range photogrammetry and the geodetic network. The third was imagery, to facilitate photogrammetric triangulation, ortho-image generation and texture mapping for the visually realistic Bet Giorgis model. In the present account, the discussion will be confined to the close-range photogrammeric survey.
In recognition of the flexibility and favourable metric performance of modern low-cost, off-the-shelf digital cameras, a decision was made to accomplish the 3D measurement of significant feature points on the church and pit walls by photogrammetric means, and to fully carry out this task on site. As a consequence it was planned that upon leaving Lalibela after the 4-day fieldwork phase, a sufficiently dense array of feature points would be surveyed. These would provide a preliminary point cloud for the later to be determined CAD model of the church, and facilitate straightforward exterior orientation determination for any images that were to be employed in subsequent ortho-image generation and model rendering. Thus, facilities were required in the field to support camera calibration and photogrammetric bundle adjustment, as well as supplementary object point triangulation (with given exterior orientation) and exterior orientation determination. These needs were all met by the Australis software system for off-line digital close-range photogrammetry (Fraser and Edmundson, 2000).
A further consequence of the use of 'amateur' CCD cameras as metric imaging devices is that on a project such as the Bet Giorgis modelling there can be expected to be no shortage of cameras. In this instance, comprehensive sets of imagery were recorded with three digital cameras, a Kodak DCS330, a DC210 and a SONY Cybershot. Also, a Leica R5 35mm semi-metric analogue camera was employed for comparative studies. The focus of the SONY Cybershot was manually set to infinity and the zoom to the shortest focal setting, and calibration parameters from a prior laboratory testfield calibration were employed. The DCS330 was calibrated for two lenses (focal lengths of 14mm and 28mm) in the field via self-calibration, as was the DC210. This highly automated process consumed only a few 10s of minutes for each camera/lens combination and required the establishment of an array of 40 retrotargets on the wall of a nearby building. Confirmation of this stand-alone calibration was also afforded in the subsequent multi-image triangulation used to position interest points on the church.
These 'interest points' comprised both artificial targets attached to the rock face and natural feature points on the church. The artificial targets, 2.5cm white dots on a black background, would constitute 'framework points' for an optimal accuracy, multi-image bundle adjustment. The natural points, on the other hand, tended to be located on features of interest for the building of a CAD model, e.g. on edges, corners, etc. Moreover, within Australis the artificial targets could be measured automatically to, nominally, 0.1 pixel accuracy, whereas the natural points required manual image measurement to about 0.5 - 1 pixel accuracy. Shown in Figure 3 is a representative DC210 image highlighting targeted (green) and natural points (yellow). Beyond a height of about 6m it was necessary to use natural points due to lack of access for affixing targets.
Figure 3: Image measurement within Australis.
Figure 4 illustrates the network of 57 DC210 images and 210 object points which was computed within Australis to provide an initial array of accurately positioned object points for subsequent exterior orientation determination (for images not in the network) and point densification through spatial intersection. Redundant, surveyed control points were employed to transform the reference coordinate system of the photogrammetric survey into the geodetic system. An analysis of checkpoint residuals indicated a photogrammetric triangulation accuracy of close to 1.5cm, which was consistent with expectations for an average imaging scale of 1:3000 (4.8mm lens and 15m object distance) and the 1-pixel measurement precision anticipated for the natural targets.
As is indicated in Figure 4, images were recorded at intervals of a few metres around the top of the pit and also around the bottom, where stations were primarily at the front and rear of the church. The figure shows only the images employed to establish the framework points and not those taken as stereopairs for purposes of fine-detail feature modelling and texture mapping. The photogrammetric triangulation phase was not restricted to use of the DC210. Similar networks were also processed for the DCS330 and Sony Cybershot imagery using Australis and Photomodeler (www.photomodeler.com), respectively, with consistent results being obtained. From the full block of SONY Cybershot images, only those of the church itself have been used so far for triangulation and CAD model generation (40 images recorded from the top of the trench and 31 from the bottom). The reason for carrying out a separate triangulation with these images, based on the points generated by the DC210 block, stems from a particularity of Photomodeler. To build up a CAD model with Photomodeler it is more convenient to include all images in a single project, since this affords faster access. Also, the resulting CAD model can be expected to be more geometrically consistent and free of problems in the overlap regions of photogrammetric models. Moreover, experience has demonstrated that if only single models are measured sequentially, points and features in the overlap areas are very often either missed or measured twice.
Figure 4: 57-station, 210 point photogrammetric network to provide 'framework' points.
- CAD Model Generation
The result of the photogrammetric triangulation phase is a point-cloud of feature points, some of which are useful to the CAD model/wireframe generation process, and some which while not being of prime use in this regard are nevertheless crucial to the later processes of image registration, ortho-image generation and texture mapping. The next requirement, however, is generally to densify the object point array in order to include feature points needed for the geometric modelling. Given that there was already a comprehensive network of images of known exterior orientation, this task involved straightforward spatial intersection from typically just 2-4 images, a process which can be carried out interactively and rapidly in either Australis or Photomodeler. The majority of such points were measured in individual stereomodels via monoscopic image measurement. It is noteworthy that every new feature point was also included in the bundle adjustment such that the strength of the network improved in a stepwise fashion with the collection of CAD model points.
Visual interpretation and manual measurements were found to be absolutely necessary for producing a decent CAD model, and at this point there was little room for automation. The church possesses a much larger amount of surface detail than was expected at first sight, and the primary surfaces contain many irregularities in geometry, such that straight edges of substantial length and planar surfaces of significant size are not in abundance (e.g. features include horizontal mouldings, projecting corner beams in the doors, projecting gutter spouts and crowned ogee-arches in the upper row of the windows). Therefore, a hierarchical measurement strategy was used with feature refinements at each iteration step. The resulting point cloud was then turned into a line model and finally into a surface model. Since these two operations were carried out manually, they constituted the most time consuming of the reconstruction processes.