Harlan Hile (harlan@cs), Alan Liu (aliu@cs)

Project description
Our project investigated increasing Placelab's accuracy in determining indoor location, using wi-fi access point measurements. This would allow building-scale applications to be built that could use this increased accuracy.

As a first cut, we focused on determining floor location, and then tried to improve location accuracy in the allen center.

Platform
We did our testing on Tablet PCs

Placelab modifications
We extended Placelab's particle filter to understand the difference between floors (before position of all APs and estimates were in two dimensions). We also experimented with other methods for floor determination.

The major steps we took were

• Changing the Beacon class to support extra meta-data, in our case floor. We found GPS coordinates and other data for the allen center AP's and populated the database.
• Adding the notion of floor to each Particle estimate
• Changing the motion model to deal with walking indoors, instead of driving outdoors
• Changing the sensor model to account for wireless signal attenuation due to floors

Experiments
We logged measurements and ground truth on a tour around the CSE building, and then compared the error of the Placelab particle filter tracker to the following (separate) modifications:

• We first focused on floor estimation, using two techniques: Placelab's original sensor model, with a floor attention factor of 0.8, and taking the mode of the stronger half of access points
• We then looked at location accuracy with a few techniques: Placelab's standard particle filter, with the same floor attenuation model; Centroid of the access points of the floor (floor determined by above mode method); particle filter using only access points on the floor (floor determined by above mode method)
• We then looked at increasing accuracy using improved sensor models. Using our logs, we learned sensor models fit to the building. We followed a binning (histogram) approach like placelab's original model, but populated the bins with averages from our logs. In one model, we ignored floors in calculating the signals, and then used the same 0.8 attenuation as before. In another, we first binned on floor distance and then on 2d distance. We also tried a model where likelihood of particles was influenced by the above mode method for determining floor. Below are graphs, with comparison to original.