July 01, 2004

GPPS: A Gaussian Process Positioning System for Cellular Networks (More info)

Lighthouse beaconing into the fog

This paper presented a method of performing location estimation based on observed wireless signal strength.

There are basically two ways of approaching this problem: Fingerprinting and Access Point Modeling.

Fingerprinting means taking a bunch of samples in the area over which you want to localize. Then when you want to infer location you match the signal strength that you see with the calibration points and conclude that you are collocated with the most similar calibration point. There are many ways to extend this technique through intelligent weighting of the k-nearest calibration points, or adding a user model that provides a prior and prevents inferring a location which is too far from your previous position estimate.

Access Point Modeling means that you try and develop an accurate model of the wireless beacons whose signal strength you are observing. Then when you make an observation of signal strength you can determine, based on the model, which location is likely to generate such a reading.

This paper presented a method of fingerprinting which is a generalization of the k-nearest neighbors to include all the calibration points. It is bad because it does in fact require calibration points, but it is better than most fingerprinting techniques because it gets reasonable accuracy with far fewer calibration samples. Based on the calibration points it used a kernel method they called Gaussian Process Positioning.

The idea is to treat each calibration point as a Gaussian process and learn a kernel function which can provide the entries of a covariance matrix. Then when you want to predict the signal strength at a new location, you pick a location, plug it into the kernel function to create a covariance matrix for all the calibration points plus the new location and use the covariance matrix to estimate the signal strength.

They tested the system in a complex indoor environment. They performed worse than RADAR with dense calibration, but gradually did better as the number of calibration points becomes sparser. Notably the choice of kernel is critical to the success of the system when you have a density of calibration points of between 1 every 25 m^2 and 1 every 12 m^2. Less dense and GPPS does better than nearest neighbor techniques, more dense and nearest neighbor does better.

This system is basically learning which calibration points vary in a similar way and is therefore learning an infrastructure based model, it assumes that all dynamic variance is modeled by the gaussian white noise error model. There is no dynamic calibration point selection.

Each access point has its own kernel function.

Nearest neighbor is a specific case of the GPPS technique. GPPS uses all the calibration points to estimate position and nearest neighbor uses some subset. The kernel method allows for a principled way of weighting all the calibration points.

Posted by djp3 at July 1, 2004 09:46 AM | TrackBack (0)
Comments
Post a comment









Remember personal info?