understand the kalman filter for video stabilization

asked 2019-11-19 20:53:17 -0500

Downey gravatar image

I am trying to understand the procedure of video stabilization, which is below :

Compute goodFeature()
Compute Optical Flow using calcOpticalFlowPyrLK()
Keep only good points
Estimate a rigid transformation
Smoothing using Kalman filter
Warping of the picture.

My question is I don't understand how to employ the Kalman filter to compensate for the affine transformation model.
For example, we get an affine model for two adjective frames, I_n, I_n+1.
x = ax' + by' + c
y = dx' + ey' + f
x,y are the locations in the current frame. x',y' are the locaitons for the pair points in the previous frame.

I read some papers, which mentioned give the a,b,c,d,e,f, as input to the kalman filter and generate the compensated a',b',c',d',e',f' for the affine transformation.

From my understanding, to use the kalman filter the estimation value, observation value, error information are needed to compensate for the affine transformation model.

I can understand that the above a,b,c,d,e,f is the estimation values. But what is the observation value, error information?

Please help. Thanks

edit retag flag offensive close merge delete