Visual loop-closing with image profiles

Abstract

This paper investigates the ability of image profiles, pixel-intensity sums across subsets of a video stream, to support the crucial robotic skill of place recognition through visual information alone. Building from work in which image profiles are the fundamental image representation for a model of biological neural processing [3, 4, 5], this paper offers a conceptually simpler approach to simultaneous localization and mapping via a single camera (monocular SLAM). In contrast to feature-based approaches in which extraction and statistical post-processing dominate the computation, this work uses a representation suitable even for very simple autonomous platforms. Experiments demonstrate the ability of our profile-based path segments to compensate for the inevitable inaccuracies in odometry when creating consistent world maps.