Real-time rotational image registration

Abstract

The majority of approaches for rotationally aligning panoramic images use feature-based methods which extract and match visual clues in both images and analytically determine their rotational offset. However, these methods can be computationally expensive and might suffer from motion blur or images taken in structureless environments (e.g. forests). An alternative approach is the visual compass, a method which simulates a wide range of possible rotational offsets and searches for the best match. While this approach has been successfully implemented for rotations around a single axis, the increasing computation time for 3D rotations limits the usability for real-time applications. It has been suggested to use real spherical harmonics (RSH) to represent panoramic images, to calculate rotations in frequency domain using sparse matrix-vector multiplications. In this paper, we present strategies which are crucial to implement a real-time visual 3D compass using RSH. We provide a software implementation of the visual 3D compass and analyze the effect of increasing rotational misalignment between pairs of panoramic images. Furthermore, we show that the visual 3D compass can be used on low-cost hardware in real-time.