Using topological constraints as context for the joint classification of image regions in a traffic environment

This paper presents an approach using context knowledge in form of topological constraints to improve the classification result for an automated scene analysis. The approach consists of an initial segmentation, in which the image is divided into a set of disjoint regions based on their respective color values and a subsequent joint classification, where the generated image regions are assigned to object classes. The classification is performed according to extracted feature measurements and context knowledge about the spatial relationships between the different object classes. The classification task is formulated as an optimization problem using a maximum a posteriori estimation rule. The classification criteria are combined using the Bayesian theorem, where feature measurements and context are coded as conditional probability density and a priori probability, respectively. A Markov random field model is used to get an analytical expression for the symbolic context knowledge. Through its associated Gibbs distribution a systematic way for designing the appropriate functional form of context is found. The optimization task is solved by applying an evolutionary algorithm. Due to a problem-related formulation of the necessary operators this has been proven to be a very efficient search strategy. The performance of the presented approach is demonstrated on synthetic and real world images, showing typical scenarios of road traffic on motorways.