PROBABILISTIC FEATURE-LABELING SCHEMES - MODELING COMPATIBILITY COEFFICIENT DISTRIBUTIONS

In previous work, we have developed a probabilistic method for labelli ng of 2D geometric features. One of the requirements of this method is the provision of a probability distribution for the measurement error s when generating the compatibility coefficients. This distribution ha s hitherto been specified heuristically, and a training stage was need ed to establish the distribution covariances. In this paper, we develo p an improved error model. Using as a basis the error distributions of the feature measurements, we find the distributions for the compatibi lity coefficients by propagating the covariances of the feature measur ement errors through into the calculation of the coefficients. A means of explicitly modelling small scaling errors is also included. The mo st important consequence of this development is that the training stag e is no longer required to generate the error distributions for the co mpatibility coefficients. In addition, the distributions that are gene rated by the process described in this paper are tailored to fit the i ndividual sets of feature relations, instead of being a compromise ove r all the sets of relations. As a result, we obtained better results c ompared with those obtained using the heuristic distributions.