- Home
- Documents
*Cuspless Sub-Riemannian Geodesics within the Euclidean ... rduits/ آ Cuspless Sub-Riemannian...*

prev

next

out of 43

View

1Download

0

Embed Size (px)

Cuspless Sub-Riemannian Geodesics within the Euclidean Motion Group SE(d)

Remco Duits, Arpan Ghosh, Tom Dela Haije and Yuri Sachkov

Abstract We consider the problem Pcurve of minimizing ∫ `

0

√ β 2 + |κ(s)|2ds for a

planar curve having fixed initial and final positions and directions. Here κ is the cur- vature of the curve with free total length `. This problem comes from a 2D model of geometry of vision due to Petitot, Citti and Sarti. Here we will provide a gen- eral theory on cuspless sub-Riemannian geodesics within a sub-Riemannian man- ifold in SE(d), with d ≥ 2, where we solve for their momentum in the general d- dimensional case. We will explicitly solve the curve optimization problem Pcurve in 2D (i.e. d = 2) with a corresponding cuspless sub-Riemannian geodesic lifted prob- lem defined on a sub-Riemannian manifold within SE(2). We also derive the so- lutions of Pcurve in 3D (i.e. d = 3) with a corresponding cuspless sub-Riemannian geodesic problem defined on a sub-Riemannian manifold within SE(3). Besides exact formulas for cuspless sub-Riemannian geodesics, we derive their geometric properties, and we provide a full analysis of the range of admissible end-conditions. Furthermore, we apply this analysis to the modeling of association fields in neuro- physiology.

1 Introduction

Curve optimization plays a major role both in imaging and visual perception. In imaging there exist many works on snakes and active contour modeling, whereas in visual perception illusionary contours arise in various optical illusions [38, 42]. Mostly, such snake and active contour models involve curve optimization in Rd ,

Remco Duits, Arpan Ghosh and Tom Dela Haije IST/e, Eindhoven University of Technology, IST/e, Den Dolech 2, NL-5600 MB, The Netherlands, e-mail: R.Duits@tue.nl, e-mail: A.Ghosh@tue.nl, e-mail: T.C.J.Dela.Haije@ tue.nl

Yuri Sachkov Program Systems Institute, Russian Academy of Sciences. e-mail: sachkov@sys.botik.ru

1

R.Duits@tue.nl A.Ghosh@tue.nl T.C.J.Dela.Haije@tue.nl T.C.J.Dela.Haije@tue.nl sachkov@sys.botik.ru

2 Remco Duits, Arpan Ghosh, Tom Dela Haije and Yuri Sachkov

d ≥ 2, that rely on Euler’s elastica curves [24] (minimizing ∫ (|κ|2+β 2)ds) in order

to obtain extensions where typically external forces to the data are included, cf. [9, 12, 50–52] .

The elastica problem suffers from the well-known fact that not every stationary curve is a global minimizer, e.g. many local minimizers exist, cf. Figure 1. Sta- tionarity of a curve can be reasonably checked by the visual system using local perturbations, whereas checking for (global) optimality [45, 46] is much more dif- ficult. Some visual illusions (e.g. the Kanisza triangle) involve corners requiring abrupt resetting of initial and ending conditions, which are difficult to explain in the elastica model. Another problem with elastica is that it is very hard to solve the boundary value problem analytically [3, 4], and typically require (2d−1)-dim. shooting schemes. On top of that elastica curves relate to modes of the direction process (for contour-completion [21, 23, 38, 50]) where the direction of an oriented random walker is deterministic and its orientation is random. Such deterministic propagation only makes sense when the initial orientation is sharply defined. In- stead Brownian motion with random behavior both in spatial propagation direc- tion and in orientation direction ( [2, 13, 19, 22]), relates to hypo-elliptic diffu- sion on the planar roto-translation group. Such a Brownian motion models con- tour enhancement [19] rather than contour completion [21], see [17] for a short overview. The corresponding Brownian bridge measures [22, 57] (relating to so- called completion fields in imaging [3, 21, 53]) tend to concentrate towards optimal sub-Riemannian geodesics [7, 13, 20, 35, 37, 45]. So both elastica curves and sub- Riemannian geodesics relate to two different fundamental left-invariant stochastic processes on sub-Riemannian manifolds on the 2D-Euclidean motion group [21] (respectively to the direction process and to hypo-elliptic Brownian motion).

In short, advantages of the sub-Riemannian geodesic model over the elastica model are:

• If d = 2, every cuspless sub-Riemannian geodesic is a global minimizer [8, 15]. • The Euler-Lagrange ODE for momentum (including normalized curvature vector

κ/ √ |κ|2 +β 2) can be reduced to a linear one,

• The boundary value problem can be tackled via effective analytic techniques, • If d = 2, the locations where global optimality is lost can be derived explicitly. • Sub-Riemannian geodesics (in contrast to lifted elastica) are parametrization in-

dependent in the roto-translation group SE(d). Here we note in case d = 2, the sub-Riemannian manifold (SE(2),∆2,Gβ ) is encoded via a pinwheel structure of cortical columns in the primary visual cortex [41].

However, the practical drawback of sub-Riemannian geodesics compared to elas- tica is that their spatial projections may exhibit cusps and it is hard to analyze when such a cusp occurs. Therefore, in this article we provide a complete analysis of such sub-Riemannian geodesics, their parametrization, solving the boundary value problem, and we show precisely when a cusp occurs. See Figure 3 and see Figure 2.

A variant of the sub-Riemannian problem that ensures avoiding cusps is the fol- lowing variational problem, which we will explain next.

Cuspless Sub-Riemannian Geodesics within the Euclidean Motion Group SE(d) 3

Fig. 1: Stationary curves of the elastica problem ( ∫ `

0 κ2(s) + β 2ds → min) do not need to be global minimizers, cf. [44]. E.g. the non-dashed elastica is a global minimum (for β = 1), whereas in dashed lines we have depicted a local minimum. This is in contrast to cuspless sub-Riemannian geodesics in (SE(2),∆2,Gβ ) where every stationary curve is globally optimal.

Fig. 2: An example of a smooth sub-Riemannian geodesic in (SE(2),∆2,Gβ ) whose spatial projection shows a cusp.

On the space of sufficiently regular curves in Rd , we define a functional E : W 2,1([0, `],Rd)→ R+, with ` ∈ R+ being the length (free) of the curves, by

E (x) := ∫ `

0

√ κ(s)2 +β 2 ds. (1)

Here, s denotes the arc-length parameter of curve x and κ : [0, `]→ R+ ∪{∞} de- notes the absolute curvature ‖ẍ(·)‖ of the curve x at each arc-length, and β > 0 is a constant.

The two dimensional case (i.e. d = 2) of this variational problem was studied as a possible model of the mechanism used by the primary visual cortex V1 of the

4 Remco Duits, Arpan Ghosh, Tom Dela Haije and Yuri Sachkov

Fig. 3: Top left: example of a spatially projected sub-Riemannian geodesic without cusp (i.e. a solution of Pcurve ). Top right: example of an elastica curve reaching points x < 0. Such a (weak) connection is not possible with cuspless sub-Riemannian geodesics. Instead we see in the bottom left figure a comparable example of a spatially projected sub-Riemannian geodesic connecting the gin = (0,0,0) with g f in = (0,y f in,0) via two cusps. Bottom right: not all points in x ≥ 0 can be reached via a globally minimizing geodesic, here we have depicted the set R of admissible end- conditions g f in = (x f in,y f in,θ f in) via black cones on half circles with radius 1 and 2.

human brain to reconstruct curves which are partially hidden or corrupted. The two dimensional model was initially due to Petitot (see [40, 41] and references therein). Subsequently, the sub-Riemannian structure was introduced in the problem by Pe- titot [42] for the contact geometry of the fiber bundle of the 1-jets of curves in the plane (the polarized Heisenberg group) and by Citti and Sarti [13, 47] for the prin- cipal bundle on SE(2) also considered in this article. The stationary curves of the problem were derived and studied by Boscain, Charlot and Rossi in [7], by Duits in [20], by Sachkov in [35] and their global optimality is shown by Boscain, Duits, Rossi and Sachkov in [8, 15]. The two dimensional problem relates to a mechanical problem completely solved by Sachkov [37, 45, 46]. It was also studied by Hladky and Pauls in [32] and by Ben-Yosef and Ben-Shahar in [6]. Within Section 3 of this article we will summarize only the main results from our previous works [8,15]. For detailed proofs of these results, we refer the reader to [8, 15].

Cuspless Sub-Riemannian Geodesics within the Euclidean Motion Group SE(d) 5

Many imaging applications such as DW-MRI require an extension to higher di- mensions, see e.g. [14, 23, 27, 50], which motivates us to study the higher dimen- sional curve optimization of the functional given by Eq. (1).

Let x0,x1 ∈ Rd and n0,n1 ∈ Sd−1 = {v ∈ Rd |‖v‖ = 1}. The goal is to find an arc-length parameterized curve s 7→ x(s) such that

x = arg inf y ∈W 2,1([0, `],Rd), `, y(0) = x0, ẏ(0) = n0, y(`) = x1, ẏ(`) = n1

E (y). (2)

We shall refer to this curve optimization problem as problem P. We assume that the boundary conditions (x0,n0) and (x1,n1) are chosen such that a minimizer exists.

Remark 1. Due to rotation and translational invariance of the problem P, it is equiv- alent to the problem with the same functional, but with boundary conditions (0,a) and (RTn0(x1−x0),R

T n1), where a∈ Sd−1 is a fixed axis, and with Rn0 is any rotation that maps a fixed reference axis a to n0 ∈ Sd−1.

Remark 2. The physical dimension of parameter β is [Length]−1. From a physical point of view it is crucial to make the energy integrand dimensionally consistent. However, the problem with