Hi everyone,
I have an extended kalman filter which seems to be working OK in
floating point and I converted it to fixed point with a large number
of bits in the fractional part. What I'm observing is that the error
between the floating point and fixed point outputs seems to grow over
time. I'm using float C type in my floating-point implementation and I
have 28 bits of fractional portions in my fixed point numbers. Any one
has any experience ensuring numerical stability of kalman filters
(especially extended ones) with fixed point numbers? For example
should increasing the variance of measurement noise help make the
filter more stable?  I'm not looking for perfect match obviously but I
need to make the error between the two implementations bounded as
opposed to slowly increasing which is fatal overtime.
Muzaffer Kal

DSPIA INC.
ASIC/FPGA Design Services
http://www.dspia.com