Model selection in online handwriting

The description of digitized curves has two objectives: to decrease the mass of data and to allow processes of higher level recognition. The objective of this paper is to propose a technique of parameter estimation with the help of an extended Kalman filter. This method is available to all type of conics, circles or straight lines, that allows a more accurate fitting of curves with less parameters, by choosing the best model of representation in each point of the drawing. We have exposed how curves can be modelled by sections of straight lines, circles and conics in a same formulation based on an extended Kalman filter. The results on real drawings show that the best compromise is in the use of several kinds of models and a segmentation criterion based upon this trade-off.