Accuracy Characeterization for Metropolitan-scale Wi-Fi Localization (Mobisys 2005)

Similar goals to PlaceLab original paper, but study WiFi only.
Study the impact of limited calibration on the accuracy. Can achieve accuracy of 13-40 meters.

Active Campus work locate devices based on the known positions of APs.

Introduce the "response rate" metric - the fraction of times the AP is heard in a set of scans

Compare the following localization algorithms:

Centroid - First, a location of each AP is estimated during training by averaging the positions of the reported readings. Then, calculate the location of a testing point as a center of all APs heard during the testing scan (may apply weighting)

Fingerprinting - match fingerprints that have at most P different APs between the fingerprints in the training set and the testing point. P = 2 provided good accuracy.

Ranking - Uses Spearman rank-order to match fingerprints (the Spearmant coefficient (between -1 and 1) is being substiuted for the distance in signal space + negated (to make similar ranks have a smaller number).

Particle Filter

Data collection of 30 minutes, scanning 4 times per second

For a relatively well populated area (Ravenna) there is a little variation between algorithms.

In lower density areas, the particle-filter performs much better

For the centroid with sufficient number of observations, the GPS noise cancels out (the same for particle filter). The fingerprinting uses GPS data, so the accuracy goes down.