DETECTION OF ROUNDABOUTS IN SATELLITE IMAGES

In this article, we present a new algorithm for the detection of objects in satellite images. Our method requires a learning of the specific structures we wish to detect but no “a priori knowledge” about them. Indeed, the user just has to provide example images of the different objects he wants to find in large images. This algorithm combines an angular local descriptor and a radial one. These descriptors are rotation invariant and provide coefficients using the Fourier analysis on some well-chosen unidimensional signals in the image. These features are computed in the learning images and an euclidan distance is applied for comparison with the same features computed in the images used for detection. The efficiency of these features is compared to the first Zernicke moments on a roundabout database extracted from highresolution SPOT images in various urban areas.