László Nemes: Pattern identification method for industrial robots by extracting the main features of objects (SZTAKI Tanulmányok 34/1975)
ABSTRACT Industrial robots controlled by minicomputer and equipped with visual inputs need not analyse and recognise a fully unfamiliar scene. It is supposed the robot should have previously been taught separately all the patterns that could possibly occur. The robot scanning the scene has to break it up into separate patterns in order to analyse them and after that to compare them wi th the pattern already learnt. I wish to suggest a method of characterizing objects and stored models by their main features by means of the graph-vector method. Since the identification process is an interactive one, at first the easily obtainable characteristic vectors of lines and surfaces of two and three dimensional objects /denoted 2D,3D/, respectively, are evaluated. In many cases, an incomplete description of an object is enough to identify it, so that this method can be used even though the difference in brightness between two adjacent planes might be small or edges obscure. If the result of the first attempt of the identification is ambigous the program requires some further information for a second attempt. Result of simulation program is presented and some limitations of the present program and proposal for the future development are also described. INTRODUCTION It is well known that the production of industrial robots is increasing and their application is fanning out into various fields. To be more precise it could be said the devices now in use are rather programmable manipulators than robots. Since machines have to be equipped with suitable sensors for them to be used as industrial robots, the man-robot communication should be facilitated and finally they should be provided with basic decision-making ability. Although extensive research has been done in the above mentioned fields it has mainly been concerned with studying comprehensive questions of artifical intelligence. The results therefore could only be a useful quideline for industrial applications. I wish to suggest a simple approach for pattern recognition for industrial robots. The aim has been to develop a very fast identification program which can be used effectively in an industrial hand-eye system. A period of ten minutes for the recognition of objects in full detail is too long for a material-handling or assembling robot. And this recognition in full detail is not essential either. In scanning the scene the objects need only be identified in terms of shape and size. This could be done by locating and