Abstract
In vision-based mobile robotics, retrieving the image of a place that looks most similar to the current camera view can be used to localize a robot in a familiar environment. This technique is referred to as visual place recognition (VPR). A key challenge of VPR are appearance changes, e.g. due to differences in illumination. In this work, we focus on Fourier signatures as an especially efficient VPR method. Fourier signatures describe panoramic images by a subset of the amplitudes of their frequency spectrum. By representing the panoramic image in the frequency domain and discarding phase information, the resulting signature of a place becomes invariant to cyclic shifts along the horizontal axis, corresponding to a rotation of the mobile robot. We show that whitening these signatures with Principal Component Analysis (PCA) substantially improves the robustness against illumination changes and attribute the improvements to the combination of scaling and decorrelation effects provided by PCA whitening. Comparing Fourier signatures with AnyLoc, a modern deep-learning-based image description method, we observe competitive VPR quality at substantially lower computational cost.