1: \begin{abstract}
2: We present a fast convolution-based technique for computing an approximate,
3: signed Euclidean distance function on a set of 2D and 3D grid
4: locations. Instead of solving the non-linear, static Hamilton-Jacobi equation
5: ($\|\nabla S\|=1$), our solution stems from first solving for a scalar field $\phi$
6: in a \emph{linear} differential equation and then deriving the solution for $S$
7: by taking the negative logarithm.
8: In other words, when $S$ and $\phi$ are related by $\phi = \exp \left(-\frac{S}{\tau} \right)$
9: and $\phi$ satisfies a specific linear differential equation corresponding to the extremum of a variational problem, we obtain the
10: approximate Euclidean distance function $S = -\tau \log(\phi)$, which converges to the true solution in the limit as $\tau \rightarrow 0$.
11: This is in sharp contrast to techniques like the fast marching and fast sweeping methods which
12: directly solve the Hamilton-Jacobi equation by the Godunov upwind discretization
13: scheme. Our linear formulation results in a closed-form
14: solution to the approximate Euclidean distance function expressible
15: as a discrete convolution, and hence efficiently
16: computable using the fast Fourier transform (FFT). Our solution also
17: circumvents the need for spatial discretization of the derivative
18: operator. As $\tau\rightarrow0$,
19: we show the convergence of our results to the true solution and also
20: bound the error for a given value of $\tau$. The differentiability
21: of our solution allows us to compute---using a set of convolutions---the
22: first and second derivatives of the approximate distance function.
23: In order to determine the sign of the distance function (defined to
24: be positive inside a closed region and negative outside), we compute
25: the winding number in 2D and the topological degree in 3D, whose computations can also be performed via
26: fast convolutions. We demonstrate the efficacy of our method through a set of experimental results.
27: \end{abstract}
28: