Rotation invariant features from omnidirectional camera images using a polar higher-order local autocorrelation feature extractor

Proposed in this paper is a component for extracting low-dimensional rotation invariant feature vectors directly from omnidirectional camera images. The component is based on higher-order local autocorrelation (HLAC) functions, but with a modification that makes the extraction result in rotation invariant representations. As the component provides a static mapping to feature vectors, it requires no setup or learning phase and is well-suited for lifelong learning scenarios where input distributions can be nonstationary. Experiments with an actual robot system are presented and results show that the extracted feature vectors manage to capture structures in the environment. When used as the perceptual component of a sequential Monte Carlo localizer, the location of the robot can be tracked without access to long-range distance sensors. Important limitations and suitable uses for the extracted representations are also discussed.