which separates the unknown parameter corrections associated with the model 
A, from those associated with the camera orientations Ao . Assuming, 
as a first step, Ag a null vector, one obtains a Ay vector by a series 
of inversions of 3x3 matrices, each of which contains the coordinate cor- 
rections of a single control point. Multiplying this Ay vector with the 
submatrix; By(AP A) 'B, ; and adding this result to the absolute column, 
BL(AP™'A M , à Ag vector can be computed by inversions of & series of 
maximum (9x9) matrices, each of which contains the corrections to the orien- 
| | tation elements of a single camera. This new Aq vector is multiplied with 
| the submatrix, BI (AP"'A'Y'B,, and added to the absolute column, BI(AP MW)" 
and a new Ay vector is computed using the aforementioned computational steps, 
which in turn lead to the computation of a new ^g vector. This method will 
converge although very reluctantly. A geometrical analogue of such a solution, 
although not entirely descriptive, leads to the following approach. Starting 
with certain approximations for the orientation parameters, sets of coordinates 
  
i of all points of the model are computed with formulas (56) and (57), which may 
be subjected to an after-treatment according to formula (39) with Ag as null 
vector. In any case, the maximum size of the matrix which must be inverted is 
(3x3). With the thus obtained spatial coordinates of all points of the model, 
a series of resections in space 1s computed, thus obtaining the orientations 
of all cameras. In these computations, matrices of (9x9) maximum size must be 
inverted. With the thus computed orientation parameters, & new set of coordi- 
nates of the model are computed, on which a new set of camera orientations can 
  
be hased. By repeating these two phases, alternately, the final orientations 
and correspondingly, the final coordinates of all measured points of the model 
can be computed. Agaln,' the convergence is extremely slow. An increase in 
the slope of convergence would remedy this situation. The associated numerical 
effort may be considerable. Such methods are described in [17]. On the other 
hand the development of electronic computers progresses at an impressive rate, 
with respect to both storage facilities and computing speeds. In the near 
future, it should be possible to handle, economically, numerical solutions 
requiring & very large number of iteration cycles. It is believed that this 
situation will make possible a solution based on relaxation techniques for the 
80 
  
  
: En T ;