Matching scale-space features in 1D panoramas

We define a family of novel interest operators for extracting features from one-dimensional panoramic images for use in mobile robot navigation. Feature detection proceeds by applying local interest operators in the scale space of a 1D circular image formed by averaging the center scanlines of a cylindrical panorama. We demonstrate that many such features remain stable over changes in viewpoint and in the presence of noise and camera vibration, and define a feature descriptor that collects shape properties of the scale-space surface and color information from the original images. We then present a novel dynamic programming method to establish globally optimal correspondences between features in images taken from different viewpoints. Our method can handle arbitrary rotations and large numbers of missing features. It is also robust to significant changes in lighting conditions and viewing angle, and in the presence of some occlusion.