In the robotics, omnidirectional vision is favored because it increases the field of sensor's view allowing for better navigation and localization of robots. The Catadioptric sensors (combination of mirror (s) + camera (s)) are a quick and easy solution to reach a large view satisfactory. However, because of the geometry of the used mirrors, these sensors provide images with a non uniform resolution and geometric distortions. Faced up to these annoyances, researchers are divided into two categories, those who have treated the omnidirectional images as perspective ones and other who preferred to propose adapted methods to the geometry of the sensors by working on equivalent spaces (Sphere, cylinder). The main advantage of the first approach is the gain in processing time, but the quality of results is often exceeded by the adapted methods. We propose in this thesis, methods of omnidirectional images processing (matching, edge detection and corner detection), based on statistical measures, without using the projection on equivalent spaces. These methods have the advantage of browsing omnidirectional images by a fixed size window, without adaptations to intrinsic sensor characteristics and geometry of the used mirror. They also have the advantage of avoiding the derivation which accentuates the effect of noise at high frequencies of the image. The proposed methods were first validated on perspective images before being applied to the omnidirectional images. The comparative results obtained are satisfactory.
