VISUAL LEARNING OF OBJECTS AND SCENES

Models of objects or scenes represent data obtained from sets of training images. A database that contains such models serves us for recognition tasks (e.g. recognition of new input images). Principal Component Analysis (PCA) is one of the widely used methods for appearance-based modeling. However, its drawback is that it is not reliable when training sets of images contaminated with non-Gaussain noise are used. This particular noise is present in most of the realistic images (e.g. Unwanted object occlusions, specular reflections, people on the scene). Here we present a more robustPCA based on the traditional PCA. We introduce the least-squares estimation that is used by traditional PCA with the statically more robust M-estimation. We describe the algorithm for minimazing the M-estimation, which is based on the non-linear iterative Newton method. Most of the outliers are detected after the minimization and the influence of all pixels isreduced with respect to their deviation. Experimental results prove that a robust PCA is more reilable on the noisy training data than traditional PCA. They show method's efficiency on the set of scene images with illumination variations. We also succesfully apply the robust method on training sets of panoramic images with varying illumination. This specific model enables surveillance and view mobile robot localisation