# when can the Kalman filter return strange values?

Sorry if I'm asking a dumb question, but I'm not very good on statistics. I have tried to do an implementation of Kalman filter. I have double checked and the implementation seems correct with test values. But when I try it with real data (computer vision tracking system) I'm getting some nan or inf values.. So I thought maybe it could be wrong input data that are wrong.

A colleague suggests that maybe Kalman filter fails if it is getting bad input like "if it get for several times a 0 value it will get infinite"

So.. here is the question: is it possible that the Kalman filter gives as result an infinite number? and when?

• Why not take a set of inputs where it goes infinite/nan and stop a few steps before that happens and watch the calculations at each step? You'll note that you have divisions and matrix inverses (which you won't do by actual inversion, hopefully, but by solving systems of equations by more stable methods). These are where your problems are most likely to arise. – Glen_b Jun 4 '13 at 23:51