CODEWITHSUNDEEP
androidmathstatisticskalman-filter
user1396033
user1396033

Reputation: 215

Location Tracking - Sigma Point Kalman Filter with a Pedometer and WIFI

I'm buliding an indoor location tracking app using android.

I've successfully implemented a pedometer that counts steps and convert them to changes from the last WIFI/GPS fix to track the user location more accuratlly inside a building.

the pedometer works like charm, and i am getting very high accurate results, BUT, when a new WIFI fix is obtained the path become unaccurate and jumpy.

I've searched in the web and read many thesis and case studies and i found that i need some kind of filter to smooth the results.

I've narrowed it down that i need to use either Particle filter or Sigma-Point kalman filter (or extended kalman filter)

I've read many articles about these filters, i've watched the course in UDACITY.com and still i can't get it.

in a brief my pedometer output is the new latitude and longitude based on the steps made an the phone heading, and the WIFI location manager ouput is also latitude and longitude.

can some one provide me a short explenation about what i should do to smooth the new WIFI fix to my current pedometer path?

Upvotes: 3

Views: 1149

Answers (1)

Dave
Dave

Reputation: 8109

To use a Kalman filter (either linear or extended) you need to:

  1. Define your state space: presumably this is the position and velocity of the person (rule of thumb: only include acceleration if you are actually measuring it). You could probably do this in Lat/Lon coordinates and not worry about the non-linearities, since people don't walk that far.
  2. Define your measurement space: i.e. what your sensors tell you and how that is related to the state space.
  3. Define a plant-noise model: determine the real-world motion characteristics of your target; usually this is just a covariance matrix that describes, statistically, how the target's motion can change between observations.
  4. Define a measurement noise model: estimate the errors in the data reported by your sensors; Again, this is a statistical model, usually expressed in terms of the covariance of the measurement errors.

My guess is that your problem comes from (4): you are not appropriately modeling the measurement errors from your pedometer based sub-system, and the GPS fixed based sub-system.

It also might be worth trying out a basic linear Kalman filter on your problem; the math involved is a bit less complicated, and can more or less be used as a black box (once you define the model above); the extended approaches (sigma-point KF or particle filtering) usually require more problem-domain specific tuning.

Upvotes: 3

Related Questions