Two-dimensional object modelling for robotic vision

The authors define modeling requirements for robotic vision, describe previous work in developing models, and discuss how models are used in vision processing. They describe how Honeywell research results improve on previous manual techniques by using heuristic model feature selection and reduce the need for train-by-showing images by substituting synthetic imagery. Particular, attention is given to efforts to construct an automated modeling system which develops 2-D part recognition models without the use of the online vision system and which requires only minimal operator expertise. The focus has been on two key components of this automated model building system; automated feature selection and ranking, and synthetic image generation. The value of these advances is discussed, along with an assessment of future research needs.<>