[Autopilot] rough process noise determination
Status: Alpha
Brought to you by:
tramm
From: Jimmy T. <jts...@ya...> - 2003-03-28 07:50:05
|
Hi Aaron, My crude plan for getting the error covariance matrix: First, integrate 3dm-g IMU values to get position, (body) velocity, and orientation. Then, determine the difference between predicted values and the actual values. f=[x,y,z, u,v,w, q0, q1,q2,q3]' dx/dt = f(x(t),t) x[k+1]=x[k]+f(x(t),t)*period+wk wk = x[k+1]-x[k]-f(x(t),t)*period process noise covariance Qk = E[wk*wk'] For example, move the 3dm sensor at x axis for 1 meter and check the difference between the output of state propagation and the true value. That error divided by the number of state propagation cycles will be errors of my model. Assuming the state error covariance matrix is a diagonal matrix, its diagonal terms should be close to the squares of above errors. of course, since my models are not totally correct the error tends to blows up as time goes on. I was hoping it not to blow up that quickly so that I can a least make a reasonable estimate for 5 seconds. I think because of the accelerometer offset, I am having trouble getting reasonable state propagation output. Would you measure the kalman process error covariance this way? Or can you suggest me another way? Thanks! --------------------------------- Do you Yahoo!? Yahoo! Platinum - Watch CBS' NCAA March Madness, live on your desktop! |