SELF-LEARNING IMAGE GEOMETRICAL DISTORTION CORRECTION

A method (200) of distortion correction in an image captured by a non-rectilinear camera (210) is provided, including obtaining multiple images (212) of a scene captured by the camera over time, determining (220) where bottom portions (222) of objects having moved over a horizontal surface in the scene are located in the images, determining (230) a boundary (232) of the horizontal surface in the scene based on the determined locations of the bottom portions, generating (240) a three-dimensional model (242) of the scene by defining one or more vertical surfaces around the determined boundary of the horizontal surface of the scene, and correcting (250) a distortion of at least one of the images by projecting the image onto the three-dimensional model of the scene. A corresponding device, computer program and computer program product are also provided.