Unsupervised Appearance Map Abstraction for Indoor Visual Place Recognition With Mobile Robots

Visual Place Recognition (VPR), the task of identifying place where an image has been taken from, is at core important robotic problems as relocalization, loop-closure detection or topological navigation. Even for indoors, focus this work, VPR challenging a number reasons, including real-time performance when dealing with large databases ( $\sim10^{4}$ ) (probably captured by different robots), avoidance Perceptual Aliasing in environments repetitive structures and scenes. In letter, we tackle these issues proposing off-line mapping technique that abstracts dense database georeferenced images without particular order into Multivariate Gaussian Mixture Model, creating soft clusters terms their similarity both pose appearance. This abstract representation obtained through Expectation-Maximization algorithm plays role simplified map. Since querying map yields probability being cluster, exploit “belief” within Bayesian filter regards previous query between to perform more robust VPR. We evaluate our proposal two indoor datasets, demonstrating comparable precision full while incurring shorter times handling sequential