The paper describes a method for automatic scene segmentation and nonlinear shape-preserving filtering for precise detection of road signs in real traffic scenes. Segmentation is done in the RGB color space with a version of the fuzzy k-means method. The obtained posterior probabilities are then nonlinearly filtered with the novel version of the shape-preserving anisotropic diffusion. In effect more precise detection of object boundaries is possible. Thanks to this, the overall quality of the detection stage was increased, as it was confirmed by many experiments.
展开▼