This work details a new method for loop-closure detection based on using multiple orthogonal projections to generate a global signature for each image of a video sequence. The new multi-projection function permits the detection of images corresponding to the same scene, but taken from different points of view. The signature generation process preserves enough information for robust loop-closure detection, although it transforms each image to a simple and compact representation. Thanks to these characteristics, a real-time operation is possible, even for long sequences with thousands of images. In addition, it has proved to work on very different scenarios without the need to change the parameters or to perform an onffline training stage, which makes it very independent on the environment and camera configuration.
Results of an extensive set of experiments of the algorithm on several datasets, both indoors and outdoors and including underwater scenarios, are presented. Furthermore, an implementation, named HALOC, is available at a public repository (see https://github.com/srv/libhaloc) as a C++ library for its use under the BSD license.