[Back]

G. Retscher, J. Joksch:
"Comparison of Different Vector Distance Measure Calculation Variants for Indoor Location Fingerprinting";
Talk: 13th International Symposium on Location-Based Services LBS 2016, Wien (invited); 2016-11-14 - 2016-11-16; in: "13th International Symposium on Location-Based Services LBS 2016", (2016), Paper ID 10, 24 pages.

The study-at-hand discusses Wi-Fi location fingerprinting in an indoor environment. Wi-Fi is a predestinated signal-of-opportunity which can be used for positioning of a mobile user as most devices nowadays incorporate a Wi-Fi card and it is available in many buildings and public spaces. For the determination of the user location in the fingerprinting method signal strength observations are carried out in two phases. In the first training phase signal strength measurements from the visible Wi-Fi Access Points are collected to build-up a fingerprint database. In the following positioning phase, a user can be located and tracked if he carries out similar measurements and compares them with the values in the fingerprinting database. For the matching a distance criterion is applied to obtain the best estimation of the users´ location. In analytical form the use of nine different vector distances for such an approach is investigated. The selected distances included the Manhattan, Euclidean, Chebyshev, Canberra, Cosine, Sorensen, Hellinger, Chi-square and Jeffrey vector distance. In the test bed in an office environment four multiple-SSID (Service Set Identification) Wi-Fi networks existed at a physical single Access Point location. From the results in this investigation it could be seen that not the use of all signal strength measurements yields to a better positioning solution but the measurements to one network out of the four provides a better performance. The achievable positioning accuracies depend mainly on the selection of the vector distance and matching algorithm. Furthermore, the Access Point architecture and configuration are determinant factors. In most tests in the selected office environment the Cosine vector distance provided the overall best performance followed by the Euclidean and Hellinger distance. Only with the Chebyshev distance significantly larger positioning errors occurred. In average a minimum mean distance error of around 1.4 m could be achieved when using a single network in a multiple-SSID configuration.

Location fingerprinting, matching algorithms, vector distance measures