KF with missing measurements

[kvnsng]

Hi, thanks for making this great package available. I have a quick question.

I tried looking through the issues but had a hard time finding how to deal with missing measurements. Just wondering if anyone could point me towards a way to use filterpy KF on a dataset with missing measurements. Thanks!

[nourani]

Hi Kevin, What is exactly the issue? What do you mean when you say you have missing measurements? What kind of measurements do you have? How are they sampled / how frequently do you read them? Are all missing or just some missing (at times)? Some more information can help me. Cheers

[kvnsng]

Hi, thanks for responding. I have a series of x,y coordinates of a moving object and at certain times, I have missing measurements. I found in the documentation that we can insert None for missing measurements, but not sure how I need to structure the data now.

here's my function for running Kalman Filter and smoothing:

def kalman_filter_smooth(data, r_noise=5.):
    cols = data.shape[1]
    f1 = KalmanFilter(dim_x=cols*2, dim_z=cols)
    f2 = KalmanFilter(dim_x=cols*2, dim_z=cols)

    initData = []
    for i in range(data.shape[1]):
        initData.append(data[0,i])
        initData.append(data[1,i]-data[0,i])
    f1.x=np.array(initData)

    # for matrix H
    Hmat = np.zeros((cols, cols*2))
    for r in range(Hmat.shape[0]):
        c = r*2
        Hmat[r,c] = 1.

    # for matrix F
    Fmat = np.eye(cols*2)
    for r in range(Fmat.shape[1]):
        if r % 2 == 0:
            Fmat[r,r+1] = 1
    f1.F = Fmat
    f1.H = Hmat
    f1.P *= 50.
    f1.R *= r_noise

    xs0,ps0,_,_ = f1.batch_filter(data)
    xs1, ps1, ks, _  = f1.rts_smoother(xs0, ps0)

I feed in a data with (n,2) shape. If I have missing measurements at certain times, do I need to insert a row with [None, None]?

[kvnsng]

I figured it out. Just realized you have to put in None instead of [None, None].

[RoyLarson]

Is there a way to have just some of the measurements missing per update?

Say I am tracking something where I get a position from one set of instruments and a velocity from another
set of instruments, with different update frequencies. So my measurement timeseries might look like

t	z_pos	z_vel
0	1	1
1	2	None
2	4	None
3	6	1.5

All of the KF matrices would be standard for a 1-d kinematic problem

Is there something that automatically takes care of the Nones or do I have to make something that makes an H matrix for each situation?

Thanks

Thanks