Speeded-up Bag-of-Words algorithm for robot localisation through scene recognition

Abstract

This paper describes a new scalable scheme for the real-time detection of identical scenes for mobile robot localisation, allowing fast retraining to learn new environments. It uses the image bag-of-words algorithm, where images are described by a set of local feature descriptors mapped to a discrete set of dasiaimage wordspsila. This scheme uses descriptors consisting of a combination of a descriptor of shape (SURF) and a hue histogram, and this combination is shown to perform better than either descriptor alone. K-medoids clustering is shown to be suitable for quantising these composite descriptors (or any arbitrary descriptor) into visual words. The scheme can identify in real-time (0.036 seconds per query) multiple images of the same object from a standard dataset of 10200 images, showing robustness to differences in perspective and changes in the scene, and can detect loops in a video stream from a mobile robot.