Optimization problems in statistical object recognition

This paper treats the application of statistical principles for 3D computer vision purposes. Both the automatic generation of probabilistic object models, and the localization as well as the classification of objects in compound scenes result in complex optimization problems within the introduced statistical framework, Different methods are discussed for solving the associated optimization problems: the Expectation-Maximization algorithm forms the basis for the learning stage of stochastic object models; global optimization techniques like adaptive random search, deterministic grid search or simulated annealing are used for localization. The experimental part utilizes the abstract formalism for normally distributed object features, proves the correctness of 3D object recognition algorithms, and demonstrates their computational complexity in combination with real gray-level images.