Abstract
This manuscript presents new approaches to process omnidirectional images. These methods differ from classical approaches related to image processing because they are adapted to the omnidirectional image geometry and applied directly in the image space to avoid interpolated data. One of the main originality of this work is related to the use of fuzzy sets theory to represent the imprecisions of the luminance measure due to the image sampling. This modeling associates to each pixel of the image, a fuzzy domain. To process omnidirectional catadioptric images, a projective space, constituted by the unified projection model and the surrounding cylinder, is defined. The cylinder space is locally equivalent to the space of all perspective images acquired by a rotating camera. The operators are geometrically defined on the cylinder and projected on the omnidirectional image. First, we have adapted the convolution masks by taking into account the imprecision on the grey level localization. Next, we have studied edge detection and developed a new non-additive gradient based approach. The gradient value in each pixel is estimated by an interval which provide a natural threshold. The validity of our approach is illustrated with experimentations on synthetic and real omnidirectional images. The synthetic images are first used to test the sensibility of our detectors to noise and then to study the good detection and good localization of the obtained edges. The real images allow to test the method's robustness towards classical disruptions in image processing such as brightness variation and acquisition noise.