244 
245 
STEREO CORRELATION FOR LARGE SCALE PHOTOGRAMMETRY 
Gilbert Hobrough 
Theodore Hobrough 
Automatic Vision Corporation 
USA 
Commission II 
| INTRODUCTION 
For the last three years the authors have been applying electronic stereo cor- 
relation to the design of 3D vision systems for industrial robots. Methods 
developed originally for photogrammetric purposes are being redeveloped to pro- 
vide robots with stereopsis (binocular 3D vision). Our work has brought into 
sharp relief three performance limitations of existing stereo correlators: 
poor accommodation of object discontinuities such as the edges of buildings; 
poor utilization of the low contrast imagery that is characteristic of most 
surfaces; and loss of fine image data where coarse image data is absent. 
All of the above problems have now been solved and the performance of stereo 
correlators can be raised to levels previously considered unattainable. The 
solutions result partly from the advance in digital devices over the past dec- 
ade, partly from the advent of finite impulse response digital filters, and 
partly from the invention of new stereo correlation techniques. In particular, 
the techniques of 'video delay transformation', 'envelope correlation', and 
'video processing! are involved. The purpose of this paper is to explain how 
the new technologies relate to the automation of photogrammetry and in partic- 
ular to large scale urban and engineering applications. 
2 GENERATIONS. OF STEREO.CORRELATORS 
Given a single point in one image of a stereo pair it is impossible to find 
the homologous or corresponding point in the other image directly since the 
point in question is indistinguishable from the host of other points having 
the same density. Furthermore, the densities of homologous points are not nec- 
essarily equal. 
If we enlarge the given point to a sample area that includes a number of res- 
olution elements there is hope of finding the homologous area by pattern cor- 
relation. We could then regard the centres of the areas as homologous for the 
purpose of computing the position of the corresponding point in the object. 
If stereo images were identical the size of the sample area could always be 
made large enough to ensure reliable correlation, assuming that the object 
presents resolvable detail to the cameras. Of course, stereo images are not 
identical since they represent views of the object from different points in 
space. The resulting relative displacement of homologous image elements in the 
X or baseline direction imposes a limit on what we call the coherence area of 
the images beyond which correlation is not possible. 
In general coherence area is a function of spatial wavelength so that larger 
detail exhibits a larger coherence area than smaller detail. In rough terrain 
the coherence area of stereo images may fall to less than one spatial or video 
wave length. |