Optimizing indoor location recognition through wireless fingerprinting at the Ian Potter Museum of Art

Indoor tracking of smartphones adds context to smartphone applications, enabling a range of smarter behaviours. The predicted use cases are many and varied, and include navigation, planning, advertising and communication. Potentially, indoor tracking could become as ubiquitous as GPS - however, all of these possibilities depend on being able to produce a reasonably accurate, reliable system which does not require specialised infrastructure. While professional systems using custom devices are able to achieve very high levels of accuracy (<;1 cm), consumer no-infrastructure systems struggle to achieve reliable room-level tracking. This paper focuses on the use of WiFi received signal strength indicator (RSSI) fingerprinting, a machine learning approach which currently seems to be the most promising option for consumer smartphones. We have undertaken experimentation and optimisation in a real-world, noisy environment - the Ian Potter Museum of Art - where we developed and deployed a no-infrastructure, indoor visitor tracking application. Data was collected in a trial involving several dozen users over a few weeks, who used the system extensively. This data was analysed with a range of current WiFi RSSI fingerprinting techniques and algorithms (WASP, Redpin (kNN), SSD, SVM, Gaussian Naive Bayes and Random Forests), and their efficacy was compared and improved where possible. Known challenges such as device heterogeneity are explored, and the consistency of signal levels, including magnetic fields, are examined. Large Random Forests (200 trees) were found to have the best performance, which was further improved by calibrating for average differences in RSSI between phone models, to achieve an average of 90% correct classification of exhibits within the top five hits.