I have written a MATLAB code for an Unscented Kalman filter for attitude estimation that I'd like to implement on Gumstix (embedded computer) in order to have better attitude estimation of an Unmanned Aerial Vehicle, UAV, in real time. I am using Ardupilot mega 2 (APM 2) from DIYDRONES.com. It contains tri-axis gyro, accelerometer, and magnetometer.

I cannot get the filter working optimally, in real time. I face a change in the pitch angle during a linear acceleration. I am using the FOAM algorithm, but I am not currently using the GPS velocity to subtract the extra acceleration done in the UAV frame. I hope people who have experience in this field help me to get a better implementation of the KF-based attitude estimation filer.

What is the preferred process model, and what is the right observer to use?