Knowledge-Based Blunder Treatment in Bundle Block Adjustment 
Tapani Sarjakoski 
The Academy of Finland 
Helsinki, Finland 
Abstract 
Aspects of knowledge-based blunder treatment are investigated, in particular theoretical aspects 
related to the use of classification methods for blunder detection and location. It is concluded 
that an approximate, heuristic method can be implemented by using varying risk levels with a 
conventional statistical blunder detection method such as sequential data-snooping. A method to 
combine a-priori-known and estimated reference variances is introduced. 
The principles of error correction methods are studied. It is discovered that point identifiers can be 
corrected by using search and matching techniques. This can be done without auxiliary information 
or deep understanding of the sources of these errors. Possible typing errors on the coordinate values 
of object control point data can also be correted automatically but only on the most significant 
digits; errors on the least significant digits are not correctable. 
Introduction 
The treatment of blunders involves various matters like detection, location, and correction. The 
procedures always include a certain amount of knowledge or understanding of the reality in the 
context of the task. Knowledge, if correct, usually improves the capability of a system to perform a 
task. Along with the developments of digital techniques in photogrammetry, interest in increasingly 
automatic and autonomous systems is obvious. As a consequence, some of the knowledge now 
located in the human part of an man-machine system has to be incorporated in the machine part ' 
of the system - assuming that the knowledge is necessary. 
In this presentation, some possibilities to incorporate additional capabilities into a blunder treat- 
ment system are studied. The tasks under consiredation have been, until now, carried out mostly 
by humans. 
The presentation has three main sections. In first of them, Baysian classification is introduced as 
a way to combine subjective knowledge with the information from residuals to assist in decisions 
about rejecting individual observations in blunder detection and location. In the second part, a 
simple technique to combine a-priori-known and estimated reference variances is presented. In 
the third part, techniques for correcting successfully located errors are discussed. Finally, some 
implementation aspects are reviewed. 
Some aspects related to the current investigation on blunder detection have already been published 
in (Sarjakoski, 1986), where the emphasis was on studying the goal definition of a blunder search 
independently of the actual search technique. After arriving at a goal definition, different methods 
or algorithms for the task can be introduced and compared on a more objective basis. 
199