网格数据上的有效卡尔曼滤波器实现

Question

I have written a very simple Kalman filter that operates on time series (with data gaps included). It works nicely, but I happen to have a data cube of data (an array of shape Nt, Ny, Nx , say), and I want to apply my temporal Kalman filter for each pixel in the data cube. I have done the obvious (loop over the last two dimensions), but this takes quite a long time.

Ultimately, I always end up having to extract data from individual "pixels", and constructing the relevant matrices/vectors, so the process is quite slow (note that the gaps in each individual time series are different, and typically, so is the H matrix that links the state to the observations). I'm not familiar with cython, and it might help (only that I'm not familiar with it).

I was just wondering whether a clever rephrasing of the problem, or a clever data structure, might allow to do the temporal filtering far more efficiently. I'd rather this only used numpy/scipy, and not OpenCV, as otherwise it's a hassle depending on an extra package.

Answer 1

I made a simple vectorised Kalman filter like this for processing movie frames. It's pretty quick but currently limited to 1D inputs and outputs, and it doesn't do EM optimisation of any of the filter parameters.

import numpy as np

def runkalman(y, RQratio=10., meanwindow=10):
    """
    A simple vectorised 1D Kalman smoother

    y       Input array. Smoothing is always applied over the first 
            dimension

    RQratio     An estimate of the ratio of the variance of the output 
            to the variance of the state. A higher RQ ratio will 
            result in more smoothing.

    meanwindow  The initial mean and variance of the output are 
            estimated over this number of timepoints from the start 
            of the array

    References:
        Ghahramani, Z., & Hinton, G. E. (1996). Parameter Estimation for
        Linear Dynamical Systems.
        http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.55.5997

        Yu, B., Shenoy, K., & Sahani, M. (2004). Derivation of Extented
        Kalman Filtering and Smoothing Equations.
        http://www-npl.stanford.edu/~byronyu/papers/derive_eks.pdf
    """

    x,vpre,vpost = forwardfilter(y, RQratio=RQratio, meanwindow=meanwindow)
    x,v = backpass(x, vpre, vpost)
    x[np.isnan(x)] = 0.
    return x

def forwardfilter(y, RQratio=10., meanwindow=10):
    """
    the Kalman forward (filter) pass:

        xpost,Vpre,Vpost = forwardfilter(y)
    """

    y = np.array(y,copy=False,subok=True,dtype=np.float32)

    xpre = np.empty_like(y)
    xpost = np.empty_like(y)
    Vpre = np.empty_like(y)
    Vpost = np.empty_like(y)
    K = np.empty_like(y)

    # initial conditions
    pi0 = y[:meanwindow].mean(0)
    ystd = np.std(y,0)
    R = ystd * ystd
    Q = R / RQratio
    V0 = Q

    xpre[0] = xpost[0] = pi0
    Vpre[0] = Vpost[0] = V0
    K[0] = 0

    # loop forwards through time
    for tt in xrange(1, y.shape[0]):

        xpre[tt] = xpost[tt-1]
        Vpre[tt] = Vpost[tt-1] + Q

        K[tt] = Vpre[tt] / (Vpre[tt] + R)
        xpost[tt] = xpre[tt] + K[tt] * (y[tt] - xpre[tt])
        Vpost[tt] = Vpre[tt] - K[tt] * (Vpre[tt])

    return xpost,Vpre,Vpost

def backpass(x, Vpre, V):
    """ 
    the Kalman backward (smoothing) pass:

        xpost,Vpost = backpass(x,Vpre,V)
    """

    xpost = np.empty_like(x)
    Vpost = np.empty_like(x)
    J = np.empty_like(x)

    xpost[-1] = x[-1]
    Vpost[-1] = V[-1]

    # loop backwards through time
    for tt in xrange(x.shape[0]-1, 0, -1):
        J[tt-1] = V[tt-1] / Vpre[tt]
        xpost[tt-1] = x[tt-1] + J[tt-1] * (xpost[tt] - x[tt-1])
        Vpost[tt-1] = V[tt-1] + J[tt-1] * (Vpost[tt] - Vpre[tt]) * J[tt-1]

    return xpost,Vpost 

If anyone knows of a vectorised Kalman smoother implementation in Python that supports multidimensional inputs/outputs and EM parameter optimisation, I'd love to hear about it! I made a feature request to the PyKalman maintainers, but they said that clarity was a higher priority than speed.