Using perceptron-like algorithms for the analysis and the parametrization of object motion
Marco Mattavelli, E. Amaldi · 2002
A new approach based on the extraction of maximum consistent subsystems of linear systems is proposed for addressing the general problem of determining the linear motion parameters of unknown moving objects from a sequence of images. This type of task can be tackled using simple but effective variants of the well-known perceptron algorithm that aim at maximizing the number of patterns that are correctly classified. Unlike in the usual perceptron applications, the weight vectors determined during the training phase are not used to classify new patterns but to extract the structure and to provide the parameters of the considered piecewise linear model. The potentialities of the new approach are demonstrated for the segmentation of the optical flow. Experimental results obtained for fields from synthetic and natural images indicate various advantages of our approach with respect to some classical alternatives.