A Special Hardware for Detection of Attaching or Overlapping Objects

A high-speed processor for the recognition and location of object features is presented. The svstem is equipped with a powerful coprocessor to compute the normalized cross-correlation. An application in industry demonstrates the systems benefits. Flexible systems for industrial automation dema~d sophisticated techniques for visual observation of the scene. Completeness checks, alignment and gauqing as well as forms of dimensional checks require the ability to precisely locate object features in an image. In industrial environments one has to cope with the presence of noise in the images, image degradation, variations of lighting as well as variations of the visual appearance of the objects of attention. Finding the position of a feature of interest in an imaqe can optimally be done by template matching, measuring the similarity hetween a reference pattern (template) and portions of the image. The research community has been using the normalized correlation to compute similarity measures for feature matching for years /Pratt/. For practical applications the benefits of normalized correlation were obstructed by its enormous computational cost, making it improper for real.-time applications on availabe processing hardware. Instead approaches were made based on edge detection methods trying to extract higher level information to perform faster search in a feature space with suhstantial1.y reduced amounts of data /HBttich/. These approaches were restricted to few practical applications due to the efforts needed to maintain a sufficient performance under industrial conditions, the methods and algorithms used still being far from perfect. Me consider template matching by normalized correlation search a very effective method to locate features in an imaqe if the performance of the method is improved by proper use of todays technology. A system to perform template matching in real time was reported /Silver/. Unfortunately for many industrial applications , this system does not perform rotation of the templates while computing the correlation coefficients .