Motion Model

The general process of image acquisition (e.g., taking an image by a camera) can be modeled by

$$g(x,y)=\int_0^T \int \int_{-\infty}^{\infty} h(x,y,x',y',t) f(x',y',t)dx'dy'dt +n(x,y)$$

where T is the exposure time, n(x,y) is some additive noise, and h(x,y,x',y',t) is a function characterizing the distortion introduced by the imaging system, caused by, for example, limited aperture, out of focus, random atmospheric turbulence, and/or relative motion. If the imaging system is ideal, spatial and time invariant, and noise-free, i.e.,

$$h(x,y,x',y',t)=\delta(x-x',y-y')$$

then the imaging process becomes

$$g(x,y)=\int_0^T f(x,y,t) dt$$

If the signal is also time invariant (a stationary scene), i.e., f(x,y,t)=f(x,y), the image obtained is simply

$$g(x,y)=T\;f(x,y)$$

Now assume there exists some relative planar motion (only in the x-y plane) between the object and the camera system, i.e., the signal f(x,y,t) is no longer time invariant. This planar motion can be described by its two components in x and y directions $\{x_d(t), y_d(t)\}$, and the image of this moving object becomes

$$g(x,y)=\int_0^T f(x,y,t) dt=\int_0^T f(x-x_d(t),y-y_d(t)) dt$$

For simplicity, we assume 1D linear motion in x direction only:

$$x_d(t)=vt,\;\;\;\;\;\;\;\;\;y_d(t)=0$$

where v is the speed of the motion.

If we introduce a new variable x'=vt, we have dt=dx'/v and the integral from 0 to T with respect to t becomes integral from 0 to $L\stackrel{\triangle}{=}vT$ with respect to x', the imaging process can be described as

$$\int_0^T f(x,y,t) dt=\int_0^T f(x-vt,y) dt$$ g(x,y) =

$$\frac{1}{v}\int_0^Lf(x-x',y) dx' = \int_{-\infty}^{\infty} f(x-x',y) h(x') dx'$$ =

= f(x,y) * h(x)

where the function

$$h(x)\stackrel{\triangle}{=} \left\{ \begin{array}{ll} 1/v & ... ...{if $0 \leq x \leq L$ } \\ 0 & \mbox{else} \end{array} \right.$$

can be considered as the impulse response function, or the point spread function (PSF) of the imaging system. Note that variable y can be treated as a parameter and in this case motion restoration is essentially a 1D problem.