automatic methods to extract and to match tie points; this limits 
the amount of measurements to be performed by the user. In this 
context, this paper describes the solution developed at the 
University of Parma to deal with the orientation of close-range 
block and the perspectives towards a complete automation of 
this task. 
What this paper is about 
In this paper we present a system of programs developed to 
perform image orientation in close range blocks that, according 
to testing and checks we executed, is considerably robust with 
respect to block configuration. The current implementation (in 
the following referred to as TRIACR) has only batch 
functionality: it receives as input image coordinates and the 
coordinates of GCPs and works out the bundle block 
adjustment. The final goal of the system is nevertheless to 
interface the program to the measurement process, in order to 
improve possible network configuration deficiencies as well as 
the robustness with respect to blunders in the data set. After this 
testing stage, the efforts will focus in that direction. The goal is 
not setting up an on-line orientation process, but rather allowing 
fast editing and correction after completing the measurement on 
the images. 
The system implement algorithms for space resection, space 
intersection and bundle block adjustment integrated in a 
management module which oversees the sequence of 
computations and the whole orientation procedure. While the 
adjustment program is our own development, neither the 
algorithm of space resection nor that of space intersection have 
been originally developed by the authors: as it will be specified 
in the following, changes have only been made to the resection 
algorithm, which is now significantly more robust than in its 
original release. The main contribution of this work is therefore 
to be found in the procedure handling the block orientation. 
In the next section, the orientation procedure will be introduced, 
outlining how the workflow fits naturally within any block 
geometry, including, perhaps with an additional measurement 
effort, aerial blocks with standard overlap. Section 3 introduces 
the two algorithms employed. The space resection algorithm, as 
in other approaches, relies on the knowledge of 4 object points 
to discriminate among different solutions. Space intersection is 
based on a vector formulation leading to a closed formula and to 
the evaluation of the parallax in object space. 
Section 4 presents the checks already implemented to verify the 
quality and reliability of the computed solution, the results of 
the procedure on two of the several real and simulated blocks up 
today oriented with TRIACR. 
THE PROCEDURE FOR BLOCK ORIENTATION 
Outlook of the Strategy 
The core of the program TRIACR is based on repeating, until 
the whole block is oriented, a sequence of three operations: a 
space resection of single images followed by a space 
intersection and finally a bundle adjustment including all 
information (GCP coordinates, image coordinates of tie and 
GCPs and EO parameters) available at that stage. Accordingly, 
there are three main modules: 
1. an algorithm of space resection (RESECT), which 
requires at least four points to compute a solution but no 
approximations of the EO parameters; 
2. an algorithm to compute the approximate object 
coordinates of a point measured in a pair of images; 
3 a least squares bundle adjustment program (CALGE), 
As mentioned above, we assume that block orientation starts 
when the measurement of the image coordinates of all tie and 
GC points has been completed, so that the block geometry is 
consistent. Apart from a remark discussed later in this section, 
criteria for number and distribution of tie points and control 
points are just those one would to apply to that aim in any 
block. 
Our goal is to run a bundle adjustment of the block without 
providing initial values (either for EO parameters and for 
ground coordinates of tie points). We will use RESECT to 
provide approximate values for the EO of every image and a 
closed formula (Cooper and Robson, 1996) to compute the 
(approximate) ground coordinates of tie points from an image 
pair. To this aim, we have to provide RESECT with at least 4 
known object points for every image: we do so by an 
incremental (iterative) process where, starting from a kernel of 
oriented images and by analysing all measured image 
coordinates, we orient new images and compute preliminary 
ground coordinates of tie points as soon as this is feasible. It can 
be thought as of an extension of the strip concatenation of single 
models used in the past. To start the process, we need in the 
block at least one kernel of images (two or more), which can be 
registered by RESECT (i.e. containing at least 4 GCPs). The EO 
of kernel images, determined by RESECT, is used to determine 
by space intersection the approximate ground coordinates of 
every tie points shared by the images of the kernel. To complete 
the first sequence of operations, EO and ground coordinates of 
tie points are refined by computing a bundle adjustment. 
The list of the known object points (true GCPs and tie points 
just determined) is therefore updated and the image point list of 
the whole block is searched for images containing, based on the 
updated object point list, at least four known object points: this 
should be the case if proper overlap between the images of the 
block exist and if tie point selection was adequate so that the 
block geometry is stable and consistent. Additional GCPs are 
also included, as soon as the images they are measured come 
into play. 
The procedure is then started again from the new images which 
can be oriented by RESECT. This is repeated until the whole 
block is completed, adding new images and new tie points at 
every iteration. The success of the method depends just on the 
correct design of the block, particularly on the distribution of tie 
points. As already stressed, this does not put any true additional 
constraint on the standard triangulation procedure. Take for 
instance an aerial block with its typical regular shape: for the 
method to proceed smoothly across the whole block, we need: 
- atleast one model with 4 GCPs; 
- along every strip, at least 4 tie points rather than the 
standard 3 in the von Gruber positions; 
- if there are no or just a few GCP at the block ends, at least 
one image pair with 4 tie points across each strip. 
None of these features looks as serious constraint (the method 
was successfully tested on aerial blocks with 60% forward and 
20% side overlap), taking into account that the method is 
devised for use in terrestrial blocks, where a larger number of 
tie points is used and overlaps are on average larger than in 
aerial blocks. There is no need for the procedure to start froma 
“corner” of the block: the initial kernel grows in every direction 
where tie points allows it to do. There may even be more than 
one kernel. 
It may happen that the configuration of known object points is 
not suitable to enable RESECT to compute the solution. If this 
is the case, the image will not be oriented at that iteration. If 
other tie points in the image are determined from nearby images 
at a later stage, this gives another chance for the orientation of 
—454— 
the im 
or mo 
compu 
Since 
combi 
a tie p 
disper: 
measu 
is too 
block 
image: 
thresh« 
T 
The pr 
system 
GCPs, 
metho 
approa 
values 
not li 
conver 
compl 
param 
propos 
requiri 
camere 
interio: 
require 
Killian 
(1989) 
on the 
paper; 
We sta 
of Cre 
change 
propos 
image 
exterio 
The R] 
genera 
r;= € 
where 
point v 
R, the 
each re 
scalar 
enougl 
€, 3 ro 
is not 
equatic 
elimin: 
by a se 
uzL 
where 
space.