Off the beaten track: Predicting localisation performance in visual teach and repeat

Abstract

This paper proposes an appearance-based approach to estimating localisation performance in the context of visual teach and repeat. Specifically, it aims to estimate the likely corridor around a taught trajectory within which a visionbased localisation system is still able to localise itself. In contrast to prior art, our system is able to predict this localisation envelope for trajectories in similar, yet geographically distant locations where no repeat runs have yet been performed. Thus, by characterising the localisation performance in one region, we are able to predict performance in another. To achieve this, we leverage a Gaussian Process regressor to estimate the likely number of feature matches for any keyframe in the teach run, based on a combination of trajectory properties such as curvature and an appearance model of the keyframe. Using data from real traversals, we demonstrate that our approach performs as well as prior art when it comes to interpolating localisation performance based on a number of repeat runs, while also performing well at generalising performance estimation to freshly taught trajectories.

“Dave...I can assure you ...that it’s going to be all right ...” A Definition, Case for, and Survey of Algorithmic Assurances in Human-Autonomy Trust Relationships

“…In the context of image classi cation, Paul and Newman [108] introduced 'perplexity' as a metric that represents uncertainty in predicting a single class and is used to select the 'most perplexing' images for further learning. ere have also been several a empts to use Gaussian processes (GPs) to actively learn and assign probabilistic classi cations [13,31,48,49,91,[131][132][133]. As with perplexity-based classi ers, the key insight is that if a classi er possesses a measure of uncertainty, then that uncertainty can be used for e cient instance searching, comparison, and learning, as well as reporting a measure of con dence to users.…”

kRadar++: Coarse-to-Fine FMCW Scanning Radar Localisation

“…We expect that the significance of this work will be persistent and anticipate follow-on investigation of efficient methods for selecting which experience to trust for localisation in a multi-experience TR scenario as in Reference [52], as well as predictive systems for characterising localisation performance in TR envelopes around the taught path as in Reference [53].…”

Fit for Purpose? Predicting Perception Performance Based on Past Experience

“…Similar problems are reported for localisation performance. [10] and [11] propose embedding spatial models of expected localiser performance in localisation maps in order to aid trajectory planners.…”