Motion model analysis

- 2 mins

Motion model analysis


Motion Model Analysis via DSP Methods

1. The Motion Model

We consider a time-varying affine transformation:

$$ f(x,y,t) = \begin{bmatrix} \theta_{k1} + \alpha_{k1}t + (\theta_{k2} + \alpha_{k2}t)x + (\theta_{k3} + \alpha_{k3}t)y \\ \theta_{k4} + \alpha_{k4}t + (\theta_{k5} + \alpha_{k5}t)x + (\theta_{k6} + \alpha_{k6}t)y \end{bmatrix}. $$

In compact matrix form:

$$ f(x,y,t) = (W_0 + t W_1)\phi(x,y), \quad \phi(x,y) = \begin{bmatrix}1 \\ x \\ y \end{bmatrix}, $$ where $W_0, W_1 \in \mathbb{R}^{2 \times 3}$.

2. DSP Interpretation

From a DSP perspective:


3. Frequency-Domain Effects

For an image $I$ with Fourier transform $\hat I(\omega)$, affine warp gives:

$$ \mathcal{F}\{I(A(t)\mathbf{u}+b(t))\}(\omega) = \frac{1}{\|\det A(t)\|} e^{-j \omega^\top A(t)^{-1} b(t)} \hat I(A(t)^{-T}\omega). $$

4. Motion Planes in 3D Spectrum

For pure translation $b(t)=vt$, the 3D Fourier transform of the video lies on motion planes:

$$ \omega_t + v_x \omega_x + v_y \omega_y = 0. $$

5. Implementation as DSP Operations


6. Parameter Estimation as Demodulation

This is analogous to carrier phase synchronization in communications: maximizing alignment of warped frames is a coherent integration problem.

comments powered by Disqus