Applying kalman filtering for mobility tracking of users using GeoCordinates

282 views Asked by At

I am trying to implement a simple kalman filter that will be used for tracking user movement using a set of geo-cordinates or place IDs from twitters tweet geo data.

Essentially, I am working on calculating mobility patterns using Twitter data? Given a large set of tweets some with geocoordinates some with location place id's, what methods can we be used (or develop) to check where do the users go, how much do they move, and when do they come back 'home'.

Can anyone shed some light into this?

1

There are 1 answers

0
VECTOR On

I will say that the question is very ambiguous in its current form and very much incomplete. I am not familiar with twitter's geo data, would the geo-coordinates be the latitude and longitude? or will they be represented in some other geo-centric coordinate system. And can I assume that you can convert the place ID's into the same representation? If so, your filter implementation will get alot simpler.

All in all, yes it is possible to track the coordinates of users, given the frequency of measurements is good enough and you are applying the right motion and observation model for each user. If you already have an ID associated with each geo-coordinate (which would be the case I think), then your filter implementation further simplifies to multiple Kalman filter instances running (one for each user).

Finally I assume that the geo-data is noisy maybe that is why you want to track the coordiantes using a Kalman filter.

Do you already know where their home is? Or would you like to detect that too using your filter by analyzing their movements?

Lastly I will discuss some problems that you may have to solve and if you are new to filtering I would recommend that you start by tracking the motion of one user and start reading www.kalmanfilter.net guide. If the geo-data is providing lat/lon then the utm (x,y) coordinates would be obtained by a non-linear function, pushing you to use a filter with non-linear observation model. As far as motion model is concerned, I think you can get away with a simple linear holonomic model (that should be sufficient for a pedestrian), but what if the user is driving a car? And a key thing for good tracking would be the frequency of measurements. How often would you be getting those position measurments? would they be sufficient for your application?

Maybe you should further clarify what is it exactly that you want to track? What does "How much they move?" mean. Would only tracking the amount of visits and time spent in a finite amount of places (like nodes in a graph) using a discrete filter would satisfy your requirements?