Due to their omnidirectional view, the use of catadioptric cameras is of great interest for robot localization and visual servoing. For simplicity, most vision-based algorithms use image processing tools (e.g. image smoothing) that were designed for perspective cameras. This can be a good approximation when the camera displacement is small with respect to the distance to the observed environment. Otherwise, perspective image processing tools are unable to accurately handle the signal distortion that is induced by the specific geometry of omnidirectional cameras. In this paper, we propose an appropriate spherical image processing for increasing the accuracy of visual odometry estimation. The omnidirectional images are mapped onto a unit sphere and treated in the spherical spectral domain. The spherical image processing take into account the specific geometry of omnidirectional cameras. For example we can design, a more accurate and more repeatable Harris interest point detector. The interest points can be matched between two images with a large baseline in order to accurately estimate the camera motion. We demonstrate with a real experiment the accuracy of the visual odometry obtained using the spherical image processing and the improvement with respect to the use of a standard perspective image processing.