You are on page 1of 4

Asymptotic Analysis of Discrete Normals and Curvatures of Polylines

Torsten Langer∗ Max-Planck-Institut f¨ ur Informatik Computer Graphics Group Alexander G. Belyaev† Max-Planck-Institut f¨ ur Informatik Computer Graphics Group Hans-Peter Seidel‡ Max-Planck-Institut f¨ ur Informatik Computer Graphics Group

Accurate estimations of geometric properties of a smooth curve from its discrete approximation are important for many computer graphics and computer vision applications. To assess and improve the quality of such an approximation, we assume that the curve is known in general form. Then we can represent the curve by a Taylor series expansion and compare its geometric properties with the corresponding discrete approximations. In turn we can either prove convergence of these approximations towards the true properties as the edge lengths tend to zero, or we can get hints on how to eliminate the error. In this paper, we propose and study discrete schemes for estimating tangent and normal vectors as well as for estimating curvature and torsion of a smooth 3D curve approximated by a polyline. Thereby we make some interesting findings about connections between (smooth) classical curves and certain estimation schemes for polylines. CR Categories: I.3.5 [Computer Graphics]: Computational Geometry and Object Modeling; Keywords: error analysis, discrete approximation, tangent vector, normal vector, Frenet frame, curvature, torsion, polyline

P−1 c P−2 d P0 e ϕ f P1 P2

Figure 1: A space curve. curve normal can be computed. For the estimation of curvature and torsion, various methods have been suggested by Boutin [2000]. With our approach, we yield simpler formulae which, nevertheless, exhibit at least the same accuracy. There are basically two ways to evaluate the quality of any of these methods. On the one hand, they can be applied to a specific polyline interpolating an analytical curve, and the result can be compared to the exact tangent vector (or any other approximated geometric property) at the corresponding point. On the other hand, an asymptotic analysis can be applied. In this case, the analytical curve is given in general form, usually represented by a Taylor series expansion. Then the outcome of the discrete approximation can again be compared to the real tangent vector. Both methods have advantages and drawbacks. The first one cannot state general results, but only for certain test curves. The second method holds for all (analytical) curves and can give clues for design and improvement of the approximations. But it is only helpful for dense polylines where dense is not well-defined. It has successfully been applied for planar curves [Anoshkina et al. 2002]; for space curves, pioneering work has been done in [Boutin 2000]. In real world applications, all these computations have often to be done in the presence of noise. In this paper, we assume that all points lie exactly on a smooth curve since the definitions for differential properties are valid only in that case. Though we make this assumption for the development of our discrete approximation formulae, this does not mean that our work is useless for real data. The estimation error of every approximation scheme is composed of a systematic error inherent in the utilized approximation scheme and of an error introduced by noise. The goal of this paper is to minimize the former. The main focus of this paper is developing a mathematical apparatus for the asymptotic analysis of arbitrary curves, and applying it to derive new, asymptotically correct estimations for tangents, normals, curvatures, and torsions of space curves. A uniform evaluation for existing approaches and our newly proposed approximations is given. In particular, we prove the convergence of our approximations and can show their optimality in many cases. To estimate torsion, which is a third derivative, we need at least four points for the approximation, but we consider also estimations using five points to obtain better results. The case of planar curves [Anoshkina et al. 2002] is consistently included.



Reliable approximations of differential properties of a curve form the basis of many algorithms in computer graphics and computer vision. Curvature, for example, can be used to define the smoothness of a curve. Furthermore, the understanding of discrete normals and curvatures of curves is a precondition for the even more important – and more difficult – task of understanding discrete normals and curvatures of a surface. In this sense, our work also lies the foundation for reliable estimates of normals and curvatures on meshes. The problem of estimating differential properties of discrete approximations has already been treated in the classical literature of differential geometry [Sauer 1970]. But in that context, the speed of convergence was not an issue, and often very simple approximations were used, yielding, for example, only linear approximations for the tangent vector. Today, one common way to obtain tangent and normal vectors at a vertex of a polyline is to compute it as a weighted average of the incident edges (or as a weighted average of the edge normals, respectively). Various weights have been proposed for that purpose. The perhaps most popular schemes are uniform weighting, weighting by edge lengths, and weighting by inverse edge lengths. It was shown in [Anoshkina et al. 2002] that the last of these methods yields the best results for planar curves. But this result holds not necessarily for space curves, in particular there exists no unique edge normal from which the (uniquely defined) ‡ e-mail:
† e-mail: ∗ e-mail:


Approximation of space curves

Let a smooth curve r be interpolated by the five points P−2 , P−1 , P0 , − − − → P1 , and P2 , with the corresponding edges Pi Pi+1 denoted by c, d, e, and f, and their lengths denoted by c, d , e, and f , see Figure 1. Then these edges can be expressed by their Taylor expansions in the coordinate system given by the Frenet frame of r with tangent t, normal n, and binormal b = t × n. Let further κ denote the curvature at P0 and τ denote the torsion at the same point. For the exact

If there were curves with other quadratic terms we could gain a tangent estimation and in turn an estimation of the normal for planar curves of the same accuracy. Proof. 2 2 ¯ = κ + e − d κ + d − de + e κ ¯ κ k 3 12 d 2 + de + e2 2 d 2 − 2de + e2 3 − κτ + κ + O(d . 2004.5 Theorem (torsion. we use inverse edge lengths as weights for the edges: 2. P−1 . The equation can directly be derived from the Taylor expansions in Appendix A. estimated using the discrete curvature vector (2) or angle approximation. Then d − e + 3g κ e−d +g τ+ τ1 = τ + τ + O(d . this is the only linear combination of d and e. Proof. f )3 (beapproximation from equation (1). however. Note that in the planar case knowledge of tangent and normal is equivalent. and even worse. f )2 . d +e+ f 6 κ 4 ˆe with the results It is interesting to compare the above estimation τ from Boutin [2000]. P0 and P1 . Horn 1983]. Now we apply the method of angle approximation ˆe (located at the edge to these binormals to compute the torsion τ e) from the angle ηe between b0 and b1 . e. d . every tangent formula can be used to compute normals of plane curves. e and f . b1 × b0 should approximately be orthogonal to n and is thus approximately aligned ˆe with t. e)3 . Also. In fact. First.expansions have a look at Appendix A. A more direct approach is to compute the curvature vector. Pi−1 . If all edges have equal length. 3 12 36 Since sin ϕ equals ϕ up to quadratic error. that yields a (at least) linear approximation of the real normal vector. and sin ηe approximates ˆ d from b−1 and b0 ηe up to second order). ˆ con2. P1 . Let g be the − − → distance P0 P2 .4 Theorem (Euler’s elastica). From the curvature vector. 2002]. respectively. and Pi+1 . and thus of the true normal vector. Another possibility is to estimate the curvature by angle approximation. 2. Therefore. e)3 . They were first introduced by Euler [1744] and have applications in computer graphics as well as in computer vision today [Cerda et al. the computation of normals is more difficult. and P2 . the convergence is even quadratic. but this is not possible. we gain the curvature as the norm. − 36 These estimations are optimal among all three-point approximations in the sense that the linear terms cannot be different from the ones that show up here. 2 e d ¯ k − d +e e d d − e 2 d 2 − de + e2 κ − κκ + O(d . If all edges have equal length. Pi . e) κ + +n κ − 3 12 2 2 d − de + e d −e κτ − (2κ τ + κτ ) + O(d . P0 . e)3 .2 Theorem (curvature vector).1 Theorem (tangent vector). who uses the same four points. for example d×e b0 = d ×e . we use the fact from b the Frenet equations that d ds = τ n. That approach is based on the definition of curvature as the rate of angular change of the tangent vector along the curve. 2. see [Anoshkina et al. e. 2002]. e)4 κ − 6 24 2 2 de d e − de + b − κτ + (2κ τ + κτ ) + O(d . and we define η b1 × b 0 . the lower order error terms vanish for an even broader class of curves. The curvature estimation κ verges of fourth order for elastica if all edges have equal length. see Figure 1. Yet another way to estimate the curvature is as the inverse of the radius of the circle passing through P−1 . 6 24 This estimation is optimal among all three-point approximations of the tangent in the sense that the quadratic term in the normal component cannot be different from the one that shows up here. and τ2 = τ + 6 κ 4 . κ 12 2 see Appendix A. 36 32 2ϕ e−d d 2 − de + e2 κ3 κ + =κ+ (κ + ) d +e 3 12 2 d 2 + de + e2 2 κτ + O(d . e)4 8 12 d 2 e − de2 de +n (1) (κ − κτ 2 ) + O(d . we can compute an approximation for κ as 2 d×e ˆ ≈κ de(d + e) without significant loss of accuracy. In fact η cause ηe depends linearly on d . b1 × b0 = sin ηe . But instead of taking the norm of the cross product. P0 and P1 is a second order approximation of the real tangent of the curve r: d de e ˜ t + d + e d 2 e2 de d 2 e − de2 = t 1 − κ2 + κκ + O(d . Again the equations can be derived from Appendix A and optimality can be reduced to the planar case [Anoshkina et al. e. e)3 . see Appendix B for a derivation. converges linearly towards the true curvature. Curvature.3 Theorem (curvature). e)4 . Using four of the five points P−2 . Let ϕ be the angle between d and e. Mumford 1994. Also note that for d = e the expansion of the angle approximation becomes e2 κ3 ˆ = κ + (κ + − κτ 2 ) + O(e4 ). This has been done in [Boutin 2000] and yields d 2 − de + e2 d 2 + de + e2 2 e−d κ =κ+ κ + κ − κτ + O(d . The tangent of the circle passing through P−1 . which is computationally rather expensive. f )2 . because the oscillating plane is unknown. We define analogously η ˆ e) and get (see Appendix A for the Taylor expansion of η 2. and here the quadratic term vanishes for a special class of curves called elastica. whereas torsion is a signed property. Again all claims can directly be proven from the Taylor expansions given in Appendix A. four points). e. e)3 . +b 3 12 Furthermore. =t 4 6 (2) d −e d 2 − de + e2 2 3 (κ − κτ ) + O(d . In 3D. ˜ t where ˜ t denotes the tangent ˆ e = ηe + O(d . always yields positive values. The last statement of the proposition can easily be derived using the Taylor expansions of d and e from Appendix A. the convergence is even quadratic. Proof. torsion can be approximated linearly as follows: ˆd 3η c−e κ c + 2d − e ˆd τ τ + O(c. f )2 6 κ 4 d +e+g κ e−d +g τ+ τ + O(d . this is the only linear combination of d and e that yields a second order approximation. Therefore. e)2 . It can be done after estimating the binormals which determine that plane. =τ− τ− c+d +e 6 κ 4 ˆe 3η f −d κ d − 2e − f ˆe τ =τ+ τ− τ + O(d . The finite difference approach yields a linear approximation of the true curvature vector. characterized by minimizing the bending energy κ 2 ds −→ min ˆ κ while fixing end points. for example using finite differences. Binormals bi at Pi can be estimated by the normal of the plane defined by three consecutive points.

and ϕ1 between e and f: c + 2d 2ϕ−1 κ−1 κ + O(c. Five points are sufficient to obtain a second order approximation for torsion: ce − d f + d 2 − e2 κ5 ˆ+ ˆ = τ + O(c. 2002. in particular. A NOSHKINA . and the Frenet frame of a space curve. S AUER .Our approximation is more symmetric in the sense that the first linear error term vanishes if all edge lengths are equal. we approximate curvatures at P−1 and P1 from the angles ϕ−1 between c and d. International Journal of Shape Modeling 5. 1806–1810. we want to examine the influence of noise on normal and curvature estimations. We assume without loss of generality that the curve is parameterized by arc length. This facilitates the problem to express discrete properties in geometrical meaningful terms like curvature and torsion by using Taylor series along with the well known Frenet equations [do Carmo 1976. f )2 . Modeling. 491–506. we provided a useful toolbox for the analysis of polylines in three-dimensional space. = τ n. KOENDERINK . five points). Proc. we can get a five-point approximation of the torsion at P0 that converges quadratically for arbitrary edge lengths. ACM Trans. and Visualization 2002. E ULER .. and curvature and torsion of curves on surfaces. International Journal of Computer Vision 40. and κ1 =κ+ e+ f 3 From this.6 Theorem (torsion. Solid Shape. R. E. M. A Taylor series expansion of space curves In this appendix. H. Acknowledgements. and so on. e. K REYSZIG . we get five-point estimates for curvature 1 κ5 ((2e + f )κ−1 + (c + 2d )κ1 ) c + 2(d + e) + f = κ + O(c. Sauer 1970]: s3 s5 s4 s2 e = s1 r + 1 r + 1 r + 1 r(4) + 1 r(5) + O(s6 1) 2 6 24 120 s3 s4 = t s1 − 1 κ 2 − 1 κκ 6 8 s5 + 1 κ 4 + κ 2 τ 2 − 4κκ − 3(κ )2 + O(s6 1) 120 s3 s4 s2 + n 1 κ + 1 κ + 1 (κ − κ 3 − κτ 2 ) 2 6 24 s5 − 1 (6κ 2 κ + 3κττ + 3κ τ 2 − κ ) + O(s6 1) 120 s3 s4 + b − 1 κτ − 1 (2κ τ + κτ ) 6 24 s5 + 1 (κ 3 τ + κτ 3 − κτ − 3κ τ − 3κ τ ) + O(s6 1) . Additamentum ‘De Curvis Elasticis’. G. (κ 4 + κ 2 τ 2 − 4κκ − 3(κ )2 )t − (6κ 2 κ + 3κττ + 3κ τ 2 − κ )n + (κ 3 τ + κτ 3 − κτ − 3κ τ − 3κ τ )b. 1744. E.. the notion of torsion has no meaning for planar curves. r = t = κ n. anyway. Sci. In Methodus Inveniendi Lineas Curvas Maximi Minimive Probprietate Gaudentes. A. K. f ). 211–216. J. L. Acad. Koenderink 1990] dt db dn = κ n. d . Numerically invariant signature curves. c + 2(d + e) + f and finally 2. Differential Geometry. L. We would like to thank the anonymous reviewers for their valuable comments. P. B. Springer. respectively. AND S EIDEL . J. f )2 .. 2. s0 = 0. 1983. C. 2004. on Math. Furthermore. E. we can even get an expression completely without linear terms if d = e and c = f . A. Our treatment is based on the work of Anoshkina et al. D. B ELYAEV. By estimating torsion using the angle between b−1 and b1 . DO C ARMO . University of Toronto Press. Polygonal curve evolutions for planar shape modeling and analysis. A better way to obtain such a symmetric expression is to take the ˆe such that the term involving ˆd and τ (unique) weighted average of τ τ vanishes completely and the term involving κ κ τ vanishes for d = e and c = f : 1 ˆ τ ( f + 2e − d )τd + (c + 2d − e)τe c+d +e+ f =τ− ce − e2 + d 2 − d f κ τ + O(c. Ed. Also. M.. YOSHIZAWA . 3(c + d + e + f ) κ It can be further improved by estimating κ κ τ and eliminating the corresponding error term. r(5) = r = (κ n) = κ n − κ 2 t − κτ b. and κ5 are defined as above. L. e. f )2 . In that way. r(4) = −3κκ t + (κ − κ 3 − κτ 2 )n − (2κ τ + κτ )b. for its derivative (also suggested in [Boutin 2000]) 3 κ5 (κ1 − κ−1 ) = κ + O(c. interpolated by a polyline as in Figure 1. f )2 . Berlin. 1959. = −κ t − τ b. MIT Press. such that they converge towards their smooth counterparts as edge lengths tend to zero. In Algebraic Geometry and its Applications. τ5 τ τ 3(c + d + e + f ) κ5 ˆ .. 503 pages. B ELYAEV. d )2 =κ− c+d 3 f + 2e 2ϕ1 κ + O(e. 1999. The curve of least energy. (3) ds ds ds where t. Its application yielded several formulae to estimate curvature. Asymptotic analysis of three-point approximations of vertex normals and curvatures. 2000. AND YANO . e. with r(si ) = Pi . Natl. n and b are the unit tangent. M UMFORD . C ERDA . The elements of draping. κ5 . J. torsion. 1994. e.. Here. . but we have to take into account the higher complexity of three-dimensional space. we proved the optimality of our estimates in many cases.) Differentiating the curve r(s) then yields r = t. we plan to extend our research to the asymptotic properties of estimations of normals and curvatures of meshes. In Vision. and κ and τ are curvature and torsion. n. b) is an 3 Conclusion We have presented a mathematical framework to evaluate and develop approximation schemes to estimate differential properties of discrete curves. τ B OUTIN . Bajaj. M. (We omit the position s since the equations hold for all (fixed) s and we are interested only in the case s = 0. 120 In the next step. AND PASINI . 3. E. V. 1970. In the future. Differential Geometry of Curves and Surfaces. 1976.. [2002]. Thus. G. M. 1990. respectively. d . V. d . 235–248. M AHADEVAN . Now we can use Taylor expansion to express the edge − − → e = P0 P1 = r(s1 ) − r(0) in the local canonical form [Kreyszig 1959. Elastica and computer vision. S. Differenzengeometrie. 7 (February). Software 9. 441–460. The research of the authors has been supported in part by the EC-IST FP6 Network of Excellence “AIM@SHAPE”. References A NOSHKINA . d .-P. P. Prentice-Hall. 195–217. H ORN .. we conduct an asymptotic analysis of an arbitrary curve r(s). the unit normal and the unit binormal vector. Since (t. For this purpose. we express e only in terms of its length e without using the possibly unknown geodesic length s1 . USA 101.

we obtain e3 2 e4 κ + κκ 24 24 e5 (27κ 4 − 8κ 2 τ 2 + 72κκ + 64(κ )2 ) + O(e6 ). 24 16 d +e d 2 − e2 d 3 + e3 3 2 2 (κ − κτ + 2κ ) ϕ= κ− κ + 2 6 48 3 d 2 e + de2 2 κτ + O(d . we obtain d +e d 2 − e2 (d − e)(d 2 − e2 ) 2 sin ϕ = κ− κ + κτ 2 6 36 2 2 3 3 d e + de 3 d +e + (κ − κτ 2 ) − κ + O(d . e Using these series. Using the Frenet equations we get ˆ dr = t + ε −hκ t + (h + kτ )n + (k − hτ )b . ds Therefore. 18 The terms for the binormals b−1 at P−1 and b1 at P1 are similar. 5760 After inverting the Taylor series for e. e. 24 16 Note that the quadratic terms vanish for d = e. 12 8 Since the norm of the above vector equals sin ϕ . In this section. we follow the treatment given in [Belyaev et al. we have ˆ ˆ ds dr dr ˆ t= = = t + ε (h + kτ )n + (k − hτ )b + O(ε 2 ). We consider a small perturbation of r(s) ˆ (s) r(s) + ε h(s)n + k(s)b r where r(s) is an elastica parameterized by arc length s. and ε is a real number.orthonormal basis. ˜ t = η τ− τ 3 18 κ 2 2 2 d − de − 2e − 3e f − f − τ + O(d . this shows: κ3 κτ κ + − κτ 2 = 0 and κ τ + = 0. ds ˆ (s ˆ by arc length. e)4 . κ Now we can compute. + 5760 Substituting the expansion of s1 into the formula for e and dividing by e yields e2 e3 e = t 1 − κ 2 − κκ e 8 12 e4 − 9κ 4 − 8κ 2 τ 2 + 24κκ + 16(κ )2 + O(e5 ) 1152 e2 e3 e + n κ + κ + (κ − κτ 2 ) 2 6 24 e4 (3κ 2 κ − 6κττ − 6κ τ 2 + 2κ ) + O(e5 ) + 240 e3 e2 + b − κτ − (2κ τ + κτ ) 6 24 e4 (3κ 3 τ − 2κτ 3 + 2κτ + 6κ τ + 6κ τ ) + O(e5 ) . f )3 12 where ˜ t is the tangent approximation from equation (1). we can estimate the angle between two consecutive binormals as d +e+ f d 2 + de − e f − f 2 κ ˆ e = b 1 × b0 . e)4 κτ + +n 6 24 e2 − d 2 d +e +b κ+ κ 2 6 d 3 + e3 d 2 e + de2 3 + (κ − κτ 2 ) − κ + O(d . With those in turn. the situation for elastica in the three-dimensional space is more complex. e)4 . that means while fixing position and tangent of the two end points. we can compute the cross product of d d and e : 2 2 d e + de 2 d e κ τ + O(d . and in the same way we c get the expressions for d d and c . f = r(s1 + (s2 − s1 )) − r(s1 ). Hereby. − 240 In a similar fashion. we obtain f f from the difference of two Taylor expansions. e)3 . 1999] and [Mumford 1994] for elastica in the plane. 2 2 . and ˆn ˆ2 = κ ˆ 2 = κ 2 + 2εκ h(κ 2 − τ 2 ) + h + kτ + 2k τ + O(ε 2 ). ds ˆ ds d s ˆ dˆ t ds dˆ t ˆn ˆ= κ = ds ˆ ds d s ˆ = κ n + ε (−h κ − kκτ )t + ε (h(κ 2 − τ 2 ) + h + kτ + 2k τ )n + ε (−hτ − 2h τ − kτ 2 + k )b + O(ε 2 ). − 36 s1 = e + and for d = e e3 (2κ + κ 3 − 2κτ 2 ) + O(e5 ). Then Let r ˆ) be a parameterization of r ˆ dr ds ˆ= ds = (1 − ε hκ + O(ε 2 ))ds. we will derive necessary conditions for a space curve r(s) to be an elastica. 24 We can also compute the normalized binormal at P0 by d ×e d×e b0 = d e d×e sin ϕ de =t κτ + O(d . using integration by parts: ˆ 2ds κ ˆ= +ε = κ 2ds ˆ h(2κ 3 + 2κ − 2κτ 2 ) − 2k(κτ + 2κ τ )d s ˆ + O(ε 2 ) h(κ 3 + 2κ − 2κτ 2 ) − 2k(κτ + 2κ τ )ds + O(ε 2 ). Nevertheless. 360 s3 s4 e e = s1 − 1 κ 2 − 1 κκ 24 24 s5 + 1 (3κ 4 + 8κ 2 τ 2 − 72κκ − 64(κ )2 ) + O(s6 1 ). we can compute e in terms of s1 by s5 s4 1 2 e 2 = s2 κ − 1 κκ 1− 12 12 s6 4 1 + (κ + κ 2 τ 2 − 9κκ − 8(κ )2 ) + O(s7 1 ). e)3 τ+ τ + +n 3 12 18 κ (d − e)2 2 +b 1− τ + O(d . e)4 . κ 2 ds + ε Because h(s) and k(s) are arbitrary functions with compact support and the integral κ 2 ds is minimal for r(s). ϕ = eκ + d =e B Euler’s elastica for space curves κ 2 ds −→ min. e)4 × =t d e 12 d 3 + e3 e2 − d 2 (2κ τ + κτ ) + O(d . e)3 6 d 2 + de + e2 κ d 2 − de + e2 e−d τ + O(d . The same is true for fourth order terms: d e d =e e3 2 × =t κ τ + O(e5 ) d e 6 e3 +n (2κ τ + κτ ) + O(e5 ) 12 e3 e3 +b eκ + (κ − κτ 2 ) − κ 3 + O(e5 ) . h(s) and k(s) are real functions with compact support.