Visual Place Recognition to Support Indoor Localisation


Place recognition is an important capability of robots or mobile devices, that enables the localisation within large indoor environments such as public buildings, even if no connection to a Global Navigation Satellite System (GNSS) such as Global Positioning System (GPS) is possible. After successful place recognition, the semantic information about the place can be subject to further functionality such as navigation to another place.

A broad overview of place recognition and related research fields is given and various challenges and different approaches of solving this task are compiled and analysed regarding requirements such as scalability, viewpoint- and illumination-invariance, impact on restricted platform resources and robustness regarding the confusion of places.

Current visual place recognition systems mostly use images as place representation and image classification or image retrieval approaches are therefore often deployed.

The developed concept of a place recognition system, which fulfils the above-mentioned requirements is based on image retrieval and uses the flexible VLAD-PQ framework for efficient search in a database of vectorized images depicting the environment under different viewpoints and dynamic configurations. With this approach, a hierarchical search is made possible, since initially the place can be recognised by the system and afterwards a more precise position within the local coordinate system of the recognised place can be conducted. The recognised place may also directly provide a user with the requested semantic localisation information.

For this work, an environment was recorded under different dynamic configurations. It consists of eleven places, represented by a total of 4000 training images. The chosen dynamic indoor environment is a part of the first floor of the Technische Universität Dresden, which contains places of similar visual appearance and with high architectural symmetry.

Within the challenging environment, the system can correctly recognise the places 90.9% on average for a large test set. This high precision value is achieved by extensively ‘looking around’ and including the votes of many query images. While using many query images improves place recognition performance, capturing many query images of the robots’ surrounding requires more time.

Download Thesis

You can download the full submission as PDF file (2 MB) here.