Yep, that’s right, selected just for you! Oh forgot to tell you about the little present we got from IBM, a travel alarm clock and body fat analyzer in one! Never seen that combination in a single device yet (and this little jerk machine reminds me sadly that I should really do some more sports). Sadly no iphone/nike iPod connectivity yet 😉
Improving Location Fingerprinting through Motion Detection and Asynchronous Interval Labeling – Philipp Bolliger (ETH Zurich, CH); Kurt Partridge (PARC, US); Maurice Chu (PARC, US); Marc Langheinrich (University of Lugano (USI), CH)
My ETH colleague Philipp Bolliger started with stating that Location is a prominent contribution to context. Calibration and maintenance of fingerprints is time consuming and must be done continuously, so current solutions do not scale well. Real signals vary a lot over time due to many factors such as other devices, physical setup, etc. With the system he present, users can create labels for their location. Then the fingerprinting is improved by using the accelerometer to detect when a device is stationary. They also label intervals, instead of just instant measurements (as long as they know the user is not moving). They use only symbolic locations, not really coordinates at this point.
In their setup they have 6 laptop that measure signal once per minute (and this for 2 weeks). Signals were very noisy, and the more access points you have the better. The RSSI over long period of time follows a normal distribution. The correlation between access points is quite low. Short term variance is always present, so it’s important not just take a reading at a given time, but average out multiple readings. Even long term variance is substantial (night/day/weekends), so you also need to update the radio map frequently. Why should people contribute? Participation and use of this system must be fun and rewarding. There are already successful systems (geocaching, OpenStreetMap, MapServer, …).
Interval labeling: learn locations while being stationery and confirmed. Intervals can be also labeled asynchronously. People will also tend to remember the labels of the locations, and he shows a little widget to label your location. He then shows the motion detector (that most modern laptops/phones have), and the signals must be smoothed out (all three axis). Excellent performance for the motion detection. They use a probabilistic model of the likely reading of RSS, and location is estimated by choosing the model with maximal likelihood. By choosing intervals between motion as unit, labeling process is less obtrusive and labels are much easier to collect.
