The complexity and sensitivity of image point correspondence 
determination for hybrid measurement is influenced by factors 
including network geometry, the number of images captured 
and to some extent the quality of imagery, including resolution 
& illumination. 
points. Only those ranked above a pre-defined threshold are 
further considered where a third image is introduced to each 
pair, forming an image triplet. Valid correspondences are 
determined if candidate point triplets adhere to the coplanarity 
condition in the triplet. The subsequent triangulated 3D point is 
then back projected into remaining images and the relevant 
image measurements that form its observations are 
automatically gathered (Otepka et al., 2002; Cronk, 2007). 
This procedure continues iteratively until the maximum number 
of object points has been established and all images and 
potential measurements exhausted. Images can be re-resected 
during each iteration, in an effort to stabilize their exterior 
orientation, improving the accuracy of the overall process. This 
is typically worthwhile until an image contains at least 20 well- 
distributed and valid measurements, after which its translation 
and orientation parameters are unlikely to change significantly. 
Photogrammetric bundle adjustments are also run periodically 
during this process, to aid in the automatic detection and 
possible rejection of outliers and to ensure image orientations 
and object point coordinates are always up-to-date. 
As can be imagined, there is a substantial number of thresholds, 
tolerances and settings involved in this procedure, and a more 
detailed explanation of these is given in Cronk (2007) and 
Fraser & Cronk (2007). One notable example is the ability to 
set the minimum number of rays before a 3D point is deemed 
valid. Fortunately, a default set of values can usually be relied 
upon to deliver a robust and fully automatic final solution to the 
image point correspondence determination. 
observations of targeted versus untargeted points, the former 
being of 0.1 pixel accuracy or better and the latter generally 
between 0.3 and 1 pixel accuracy. The majority of manually 
digitized points will be rejected, unless the error cutoff in the 
bundle adjustment accounts for this difference in precision. 
It is widely recognised that automatic, full-scene 3D surface 
extraction via image matching is rarely feasible in close-range 
photogrammetry in situations where the objects and scenes are 
well distributed over three dimensions. The hybrid 
measurement approach, however, offers the prospect of a viable 
semi-automated approach, where the required initial model 
segmentation into surfaces amenable to measurement via 
feature- and area-based matching becomes an interactive 
operation. The process of surface generation is initiated by the 
operator interactively marking the surface area to be generated, 
as indicated in Figure 2a. Feature point detection can then occur 
in either all or selected images that ‘see’ the surface. This is 
followed by image point correspondence determination to effect 
multi-image feature-based matching, there being 13 images in 
this case. The resulting surface, which comprises 4000 points, is 
displayed in Figure 2b. Subsequent refinement by area-based 
matching is also a prospect when network geometry and image 
texture are suitable. However, the computational cost of least- 
squares matching, for example, can deprive the process of the 
timeliness needed for interactive operation. 
4. MANUAL AND SEMI-AUTOMATIC PROCEDURES 
FOR HYBRID MEASUREMENT SCENARIOS 
747 
As mentioned, one of the primary motivations for the hybrid 
measurement approach is to facilitate the measurement of 
untargeted features within images forming the network. Once 
automatic network orientation and optional image point 
correspondence determination have taken place, ancillary 
manual and semi-automatic procedures can begin. 
For manual point measurement, once a single measurement has 
been indicated within an image during interactive feature 
extraction, operator assisted measurement need only extend to 
projection of the corresponding epipolar line into other images. 
After the point’s 3D position is determined from two images, its 
predicted position can be back-projected into all other images, 
to aid the operator in the manual identification of homologous 
points. The target location prediction is considerably enhanced 
in cases where either the camera has been pre-calibrated, or 
self-calibrated in the automatic network orientation phase. This 
process can be extended to the measurement of lines and 
polygons in the iWitnessPRO system. The successful 
measurement of untargeted features within an image network is 
clearly a function of the quality of the network orientation. 
During the bundle adjustment procedure, it is required that the 
observational error detection take into account the anticipated 
five-fold or so difference in precision for the image coordinate 
. * ' * ; //y ✓ > < 
> a >vy </v ■. ■•/•//>: « / ■ 
">y. \ • >V 
(b) 
Figure 2. Surface measurement: (a) Surface area to be modelled, 
and (b) extracted surface points. 
5. EXAMPLE HYBRID MEASUREMENT SCENARIOS 
5.1 Stairway Measurement 
Hybrid measurement is particularly suitable for carrying out 3D 
reverse engineering of stairways for stairlift design, fabrication