- Computer Graphics Bonn
Transcription
- Computer Graphics Bonn
Collision Detection in Densely Packed Fiber Assemblies with Application to Hair Modeling Gerrit Sobottka Ebadollah Varnik Andreas Weber Inst. of Computer Science II Computer Graphics Group University of Bonn 53117 Bonn, Germany [email protected] LuFG Software and Tools for Computational Engineering RWTH Aachen University 52056 Aachen, Germany [email protected] Inst. of Computer Science II Computer Graphics Group University of Bonn 53117 Bonn, Germany [email protected] Abstract— In this paper we investigate the application of bounding volume hierarchies in collision detection among densely packed fiber assemblies like hair strands or cable looms. In particular, we glance at collision detection algorithms with sub-quadratic upper bound and their practicability and performance in complex dynamic hair scenes. Unlike common collision detection techniques our approach exploits the topological structure of the underlying filament assembly and allows for fast hierarchy updates in dynamic deformable scenes. Furthermore, we compare hierarchies of the wrapped and the layered type and show the feasibility of fiber based collision detection on full human hair models. We simulate each fiber of an assembly by means of Cosserat rod models. Keywords—collision detection, fiber assemblies, hair modeling, wrapped / layered hierarchies, C OSSERAT rods Fig. 1. Left: Photo; Example of a densely packed fiber assembly, section of a blond hair strand. This strand consists of about 300 single fibers of diameter ≈ 70µm. Right: Simulated fiber assembly; colliding segments are marked red. 1 Introduction Our paper is motivated by the need for efficient collision detection methods for fiber assemblies with high packing density. Such configurations can be observed in hair strands or cable looms. A fiber or filament is an one dimensional deformable structure and can be described by a sequence of segments where each segment is connected with its successor at one end point. Although the diameter of a filament is small compared to the length it is crucially used for collision detection between two segments. In densely packed structures like hair strands the probability for inter-fiber collisions is very high. Thus, we have to deal with large collision numbers. As long as we do not consider dynamic deformable filaments space partition techniques like grids [21], octrees [5], kd-trees, and BSP-trees [2] are sufficient for interference detection. When the filaments of the assembly are dynamically deformed costly updates are inevitable. The most common technique are bounding volume hierarchies [7], [20], [11], [12], [3]. Traditionally, bounding volume hierarchies encode spatial proximity of the object features. Hence, they are particularly suitable for collision detection among rigid objects. Since spatial proximity changes over time newer hierarchy based approaches exploit topological proximity rather than spatial proximity [8], [13], [10]. Recently Guibas et. al [8] proposed a special kind of bounding sphere hierarchy for deforming ball sequences (so called necklaces) which exploits the topological proximity of the primitives. This sphere hierarchy is a balanced tree whose leaves correspond to the spheres of the necklace. At each internal node the minimum bounding sphere of the canonical sub-necklace is stored. This type of hierarchy wherein spheres enclose the associated geometry is called wrapped hierarchy. They prove that for a necklace self-collision detection has a sub-quadratic upper bound. Lotan et al. [13] pursue a similar approach for self collision detection in kinematic chains. A kinematic chain consists of a set of links which are connected by joints. They employ a chain-aligned hierarchy of oriented bounding boxes which are not perfectly tight. A chain-aligned hierarchy encodes the topological proximity of the links and can be updated in O(k log nk ) time where k are the degrees of freedom of the change. In particular, a 1-DOF change—only one single freedom joint changes—would result in update costs of O(log n). James and Pai [10] have picked up the concept of the wrapped hierarchy by introducing the bounded deformation trees (BD-Tree). These trees are constructed on polygonal models and exploit spatially coherent motion that can be described as a superposition of displacement fields. Each triangle of the underlying geometry is bounded by a leaf node sphere. They use a conservative update scheme wherein a conservative radius and a new center are computed from the displacement fields describing the deformed state of the model. Our Contribution: We examine the application of bounding volume hierarchies in collision detection among densely packed fiber assemblies like hair strands or cable looms. In particular, the practicability and performance of chain-aligned hierarchies of the wrapped and the layered type are compared. We show that suitable built hierarchies of axis aligned bounding boxes outperform the previously used sphere hierarchies with respect to memory and computation time by orders of magnitude and can be used for single fiber based interference detection in hair strands. 2 Collision Detection The application of bounding volume hierarchies for collision detection and proximity queries is a common technique [7], [20], [11], [12], [3]. However, most of these approaches are limited to the collision interference detection between rigid objects. If we consider the simulation of deformable objects in conjunction with the well-established techniques the overhead for the reorganization of the trees are high. New approaches like [8], [13], [4], [10] make use of topological proximity rather than spatial proximity which varies over time in a dynamic simulation. Given two BVH’s the standard collision detection algorithm performs a simultaneous top down traversal of the two trees. That is, it checks the root boxes for overlapping and if so proceeds with the children until no overlap occurs or it has reached the leaves. In case of self-interference detection the sub-trees of each node have to be tested against each other recursively. Usually the volumes of the tested cages are taken as a simple split heuristic to allow for a balanced tree traversal. Here we are interested in self-interference detection in one dimensional deformable structures like filaments, chains or necklaces. A filament is a sequence of segments or very thin cylinders (compared to the the length of the filament) whereas a necklace a sequence of closed balls is [8]. In a chain so called links of arbitrary shape are connected via joints [13]. On an abstract layer we can transform a filament or chain into a necklace by enclosing the segments or the links with minimum enclosing spheres. Self-collision detection of a chain consisting of n links is a Θ(n2 ) process in the worst-case. Running time can be improved if further assumptions about the chain are made, specifically, if the topological proximity of the links is exploited. In this context, one can define a so called well-behaved chain with the following properties (cf. [9]): a) the ratio of the radii of largest and the smallest sphere in the necklace is smaller than a constant, b) there exists a constant such that for each sphere the concentric ball with radius times constant does not enclose the center of any other sphere of the necklace. Then the number of balls intersecting any other balls is bounded by a constant. For example, molecules are inherently well-behaved. 2.1 Chain-Aligned and Spatially-Adapted Hierarchies We utilize chain-aligned bounding volume hierarchies for collision detection in densely packed fiber assemblies [13], [8]. The chain-aligned hierarchy encodes the chain-wise proximity of the links in a balanced binary tree. The topological order of the links is represented by the order of the leaves. Thus, neighboring nodes encode adjacent parts of the chain (cf. Fig. 5). In contrast, a spatially-adapted hierarchy bases on the spatial proximity of the links. Self-collision testing is more efficient with a spatially-adapted hierarchy at a cost of Θ(n) time in the worst case [23]. Since spatial proximity varies over time as the filaments deform the tree has to be updated at each time step at a cost of O(n log n) [7] but the hidden constants are large. Normally, this involves costly reorganizations of the tree whereas the structure of the chain-aligned hierarchy cannot change. The attempt to maintain a spatially-adapted hierarchy over several time steps can lead to situations where Θ(n2 ) overlap tests are needed to check for self-collision. It was shown by Lotan et al. [13] that in a chain-aligned hierarchy of a well-behaved chain the maximum number of 4 overlapping boxes is Θ(n 3 ). Hence, the collision detection 4 takes Θ(n 3 ) time in the worst case. Fig. 2. Assemblies. Left: Close up of blond hair fibers; Right: One of our collision detection test cases; a C OSSERAT generated hair style consisting of 50,000 hair fibers with 50 segments each. Fig. 3. A snapshot from our fiber assembly simulation environment. Left: A guide filament of a simulated fiber assembly is interactively deformed. Colliding segments are marked red. Right: Cages are displayed. 2.2 Wrapped and Layered Hierarchies We can distinguish between two types of binary bounding volume hierarchies: the wrapped and the layered hierarchy [8]. The differences can be explained by means of the spherical hierarchy: Given a sequence consisting of n spheres B = (B1 , ..., Bn ) the wrapped spherical hierarchy is a balanced binary tree where each internal node stores the minimum enclosing sphere (the so called cage) of all spheres in the canonical sub-chain. That is, the cage of a node c encloses all spheres of the necklace which are the leaves of the subtree rooted at c. The spheres themselves are stored at the leaves. So, the construction of the wrapped spherical hierarchy is done in a top-down process by recursively subdividing a given chain. In contrast, in the layered hierarchy each node stores the minimum enclosing sphere of the spheres of the two child nodes. The independence of the wrapped hierarchy of the bead order is apparent and results in tighter bounds whereas the computational effort is much higher as for the layered hierarchy. It √ can be shown, that the radius of the root sphere is at most h + 1 times the size that of the layered hierarchy where h = dlog ne is the height of the balanced tree (cf. [8]). Different kinds of bounding volumes can be used with the wrapped hierarchy. The efficiency of the construction process depends on whether one can find an efficient algorithm to compute the minimum enclosing volume of a set of volumes of the same type. We will discuss this problem later. 2.3 Bounding Volumes The first step in building a BV hierarchy is to find a suitable bounding volume type. So, we briefly discuss the assets and drawbacks of two of the most popular types and verify their appropriateness for the filament covering problem. Fig. 4. Left: A single filament packed into AABB’s. Right: Filament bounded by wrapped spheres. Note, that the child cages can stick out of the parent cages. AABB’s: Axis aligned bounding boxes (AABB) trees are very popular because of their simplicity but normally displaced downward by the more efficient OBB trees. The computation of the AABB of a single segment is straightforward: Take the two endpoints, subtract the radius from the first point’s coordinates and add it to the second point’s coordinates. Then sort the coordinates such that the first point represents the lower left corner and the second point the upper right corner of the AABB. So, only six scalars are needed. When we switch to the arbitrarily orientable OBB the additional freedom requires storage space for nine further scalars (the orientation matrix of the OBB). The side faces of an AABB are always parallel with respect to the model’s local coordinate system. Thus, in fiber assemblies there exist preferential global orientations, for instance, if all fibers are parallel to one axis. The more parallel the fibers are with respect to a coordinate axis the smaller the AABB’s will be and vice versa. This can benefit a dynamic strand simulation where deformations are normally small. As described above a wrapped hierarchy must be constructed in a top-down manner by recursively computing the bounding volumes for the canonical sub-chains at each node. In case of AABB’s the construction process simplifies when we take into account the following lemma (without proof): L EMMA : Let B = {B1 , B2 , ..., Bn } a sequence of spheres and Bi,j ⊆ B with 1 ≤ i < j ≤ n an arbitrary subsequence of B. Then if the minimum enclosing volume is an AABB the M EV (Bi,j ) is equal to M EV (Bi,j ) = M EV (M EV (Bi,k ) , M EV (Bk+1,j )) where Bi,k and Bk+1,j are arbitrary subsequences of Bi,j with i ≤ k ≤ j. Hence, for AABB’s the hierarchies of the layered and the wrapped type are identical. That is, we can compute the tree bottom-up in post-order traversal: first the AABB’s of the child nodes are computed and then we update the parent’s AABB by union the AABB’s of the two children (cf. Fig. 4). The construction can be performed at a cost of O(n) time. Note that the hidden constants are very small (cf. Fig. 8). Spheres: Sphere trees are the canonical example for BVH’s [18], [17] because of their simplicity. In particular, Guibas et al. [8] studied the properties of a wrapped sphere hierarchy. The problem of finding the minimum enclosing sphere of a set of balls can be solved by the deterministic algorithm of Megiddo [15] or the randomized algorithm of Welzl [22] both with linear expected time. Thus, the wrapped sphere hierarchy for a set of n balls can be constructed in O(n log n) time. Note, that the Welzl algorithm hides a large constant of O(δ · δ!) where δ is k + 1 and k is the dimension of the space in which the spheres are to be computed. For three-space this constant becomes apparently large. When we prefer the layered alternative construction takes O(n) time. Unfortunately, spheres do not fit the segments very well and cause much overlapping in arbitrarily oriented configurations since the size of the spheres are invariant with respect to the orientation of the filaments (cf. Fig. 4). Of course, there are other bounding volume types like kDOP’s [11] or ellipsoids which we will omit here for brevity. They are left as subject of further investigations. 3 Implementation We implemented both, a chain-aligned wrapped spherical hierarchy and a chain-aligned AABB hierarchy and compared the running times for complex static and dynamic fiber assemblies. The software was integrated in our interactive fiber assembly simulator (cf. Fig. 1, 3) which allows for interactive fiber deformation based upon a mechanical rod model which is described later in this section. For the computation of the minimum enclosing spheres we employed the algorithm of Welzl [22] (there is an efficient implementation available, cf. [6]). 3.1 Maintenance of the Hierarchy The two types of hierarchies require different maintenance approaches. For the layered hierarchy we can follow a simple bottom-up strategy as described above: we simply traverse the tree in post-order and update all bounding volumes. If the current node is a leaf the bounding volume of the associated segment is recomputed. First, we have to update the bounding volumes of the children of a node and then we proceed with computing the bounding volume of the parent node in O(1) time. For AABB’s this is simply a union of the two child AABB’s. When we switch to the wrapped hierarchy things become more complicated. Before each self-collision test the correctness of the hierarchy has to be verified in a process called cascade verification [8]. Since a sphere is determined by one to four basis spheres these entities have to be stored at the internal nodes. The idea is to check whether the basis of an internal node’s cage is still valid or not. Hence, the verification is done by traversing the tree bottom-up and checking if the leaf-node spheres of the sub-tree rooted at the current node are still in the associated cage from the previous time step. If one or more leaf spheres escaped from the cage it has to be recomputed. At least one of the escaped leaf spheres constitutes a basis sphere of the new cage. This knowledge can accelerate the recalculation of the basis cage if the underlying algorithm makes use of such information (cf. Welzl [22]). The cascade verification takes Θ(n log n) time in the worst-case while in practice the time is closer to linear. The update frequency for a single cage should be low because the assumption is that the the basis of a cage stays constant over a long time. Another question is, if the update process of a chainaligned AABB hierarchy can benefit from the application of the cascade verification strategy. The answer is no, because the cascade verification takes Θ(n log n) time in the worstcase whereas the simple bottom-up update scheme described above is always linear. Furthermore, this would necessitate the storage of the basis describing the cage associated with each internal note. Unlike spheres, which are defined by at most four basis spheres each AABB is defined by one to six basis AABB’s which have to be stored at each internal node. This would cause a noticeably increase in the memory requirements. In contrast, Lotan et al. [13] used a chainaligned hierarchy of OBB’s. In particular, they kept an account of the actual joint changes of the underlying kinematic chain in every time step. These additional information allow for search path pruning. 3.2 The Super Thread Our main aim is to efficiently detect collision in a dense bunch of filaments. The expounded approaches, though, are directed towards self-interference detection in a kinematic chain or necklace. Such chains can represent backbones of large protein molecules with a few thousand atoms. Here, we consider fiber assemblies consisting of a few hundred filaments with up to 1,000 segments and more which are fixed at one side. This necessitates a mechanism that prevents the algorithm from checking each filament against each other. In this context, we can make the following observation: Hair strands usually consist of a few hundred single hair fibers. They adhere together by friction or lipids so that the whole hair strand behaves stiff to some degree. That is, fibers do not move loosely like in a bunch of ropes but the global movement of all points on the fibers are similar. Fig. 5. Super thread: All filaments of an assembly are connected by a virtual cord. Motivated by these observations we treat the whole assembly as one big winding filament or super thread (cf. Fig. 5). To accomplish this all filaments in our model have the same length and the same number of segments. The filament number as well as the segment number are stored with the corresponding leaf node. The super thread is obtained by threading a virtual cord through the filaments. Then filament (i) is connected with filament (i+1) at the tip and filament (i+1) is connected with filament (i + 2) at the root and so forth. This link-up scheme results in gaps between the end points. The size depends on how close the connected end points are. Note, that we do not really insert segments or connecting something. This is only the direction of the traversal when we build up or update our tree. Furthermore, we do not perform a nearest neighbour search to decide which fiber end points to connect. The order solely depends on the sequence of generation. 3.3 Strand Generation To allow for interactive fiber deformation within our interactive assembly simulator and to generate non-interactive test cases we use a mechanical rod model, the C OSSERAT model [1]. The initial configuration of each filament is straight. We apply the static C OSSERAT model to dynamically deform the fibers of the hair strand. The advantage over a pure geometrical deformation approach is that we obtain a continuous deformation for every load in consideration of authentic material properties of the simulated filaments. By the application of proper force-torque pairs at the end points we can deform the filaments to the desired shape. To model a single filament it is convenient to describe its geometry in terms of a space curve r(s) with arc length s (0 ≤ s ≤ L, length L). In particular, we are interested in the equilibrium configuration of the filament under an external load. The static equilibrium of a long thin rod is governed by the well-known C OSSERAT equations [1], [14]: dn +f =0 ds and dm dr + × n + g = 0. ds ds The vectors n and m are the resultant contact force and torque transmitted at the cross sectional area, f are the external forces acting on the rod, dr/ds is the tangent vector to the curve and g are the external torques. To describe the kinematics of the curve it is convenient to equip the curve with a set of right handed orthonormal triads {di (s)} each expressed by a rotation matrix. It can (but need not) be chosen to be the Frenet frame (i.e., each triad is given by the normal, binormal, and the tangent). The change of the frame with respect to the curve parameter s is governed by d0i = u × di where u is the Darboux vector expressing the rotational strains. One can introduce an analogous vector v(s) = dr/ds describing the linear strains of the filament. Both the Darboux vector u and v can be thought of as angular and linear velocity of an evolving triad when the curve parameter s is taken as time. They should always be considered in conjunction with corresponding reference strains û and v̂ characterizing the reference configuration or undeformed state of the filament which is always straight in our case. The stresses m and n are related to the strains by constitutive laws. We use the simplest possible relations given by m = K(u − û) and n = I(v − v̂) where K and I are material dependent quantities. The filaments we consider herein are fixed at one end, i.e., the position and orientation at s = 0 are given whereas for the opposite (movable) end (s = L) a force and torque is specified. One approximate solution algorithm to this boundary value problem is described in [16]. To test stability of the hierarchies under dynamic conditions we applied a series of force-torque pairs to the end point of each filament which was prerecorded from an interactively deformed guide filament. These sequences are replayed with the strand configuration of the dynamic test cases (cf. Sec. 4). The roots of the fibers of the assemblies are distributed randomly over an elliptical area with a given density of 325 points/cm2 . This corresponds with the measured average hair density on a human scalp [19]. 4 Results We used our assembly simulator for performance testing of the two types of hierarchies and ran several tests on a Pentium Fig. 6. A C OSSERAT generated strand consisting of 500 filaments of length 30 cm with 2,000 segments each. Right: The same tress with collisions marked red. IV, 2.4 GHz machine. Specifically, we consider four different test cases: 1) Static assembly test: we created a fiber assembly and increased the number of fibers from 50 up to 1,000 at 100 segments per fiber. We have chosen the filaments of the assembly to have the same length and the same number of segments. Therefore, all bounding volumes on the filament level are of the same size. 2) Dynamic assembly test: here we consider a fiber assembly with a fixed number of filaments which is dynamically deformed. Of course, for large segment numbers this cannot be done interactively because of long collision detection and deformation computation times. To circumvent this problem we use prerecorded force trajectories of an interactively deformed guide filament which are superposed with a slight random variation and then are applied to all filaments of the assembly. We tested a small assembly of 150 filaments with 100 segments each. Starting from an initially straight configuration, i.e., all fibers are in parallel the assembly is deformed to a loop in 100 deformation steps (cf. Fig. 3). 3) Half scalp test: As a more complex example we generated a hair style consisting of 50,000 filaments (50 segments each) which covered one half of a human head model (cf. Fig. 2). To our knowledge this is the first approach on fiber based collision detection on a real hair model. 4) Strand test: Here we consider a (typical) helical strand of length 30 cm with 500 (1,000) filaments and a very high sampling rate of 0.15 mm (2,000 segments, cf. Fig. 6). Note, that this equals two times the human hair diameter (0.007 cm). 4.1 Running-Times In case of the wrapped spherical hierarchy as well as the chain-aligned AABB-hierarchy we measured the running times for construction, update, and collision detection. The measurement results for the static test configuration with varying number of segments are depicted in Fig. 7 to Fig. 9. The costs for construction of the AABB hierarchy as well as the update solely depend on the number of segments in the super thread (cf. Fig. 7). In contrast, the collision detection times heavily vary subject to the configuration. Clearly, the more the filaments are in parallel with the coordinate axes the faster the collision detection is. Normally, the fiber’s mean axis is in parallel to one global axis. To avoid preferential initial configurations we simply rotate the whole assembly by π4 about two global axes. So the AABB’s take up their maximum volume. Furthermore, the application of the prerecorded loads to each fiber results in an additional overall deformation of the originally straight assembly. As depicted in Fig. 9 collision detection with the AABB hierarchy is consistently faster than with the wrapped spherical hierarchy. The diagrams for the update and cascade verification of the spherical hierarchy show peaks for certain segment numbers. These profound oscillations must be ascribed to the application of the randomized algorithm of Welzl [22] for minimum enclosing sphere computation. Specifically, the algorithm needs more time to compute the bounding volumes of 55,000 segments than for 95,000 segments (cf. Fig. 7). Updates with the chain-aligned AABB hierarchy are very fast (cf. Fig. 8), since they linearly depend on the number of segments. For example, for the super thread from a hair tress (cf. Fig. 6) consisting of 1,000 filaments with a length of 30.0 cm and 2,000 segments each (= 2 × 106 segments) the hierarchy update takes 300 ms. Detecting the 23,000 implicit collisions in this assembly takes 1.7 secs. We doubled the number of filaments in the hair tress to 2,000. Update took 600 ms as expected, the detection of 80,000 implicit collisions took 4.3 secs with the AABB-hierarchy. For the half scalp model it takes 100 secs. to find all 550,000 collisions whereas the the wrapped spherical hierarchy needs 2, 000 secs. (cf. Tab. I). For the dynamic case we can observe a strong correlation between the number of collisions found in each simulation step and the corresponding collision detection times (results of the dynamic test case are depicted in Fig.10). This is due to the output sensitivity of the collision detection process. As can be seen the update times are independent of the underlying geometry and constitute a stable base line (the slight variations are caused by the operating system). 5 Conclusion and Future Work We have shown that collision detection for densely packed fiber assemblies is possible on standard PC’s with passable effort up to a complexity in the number of fibers and the geometric settings that resemble human hair strands or even Half scalp: SPW CABB Strand: SPW CABB Construction [s] Update [s] Detection [s] 4.609 4.109 1924.74 0.766 1984.78 100.531 2.907 2.781 91.923 0.281 15.282 1.688 TABLE I C OLLISION DETECTION STATISTICS FOR CASE SPHERICAL HIERARCHY; 3 AND 4; SPW: W RAPPED ABB: C HAIN - ALIGNED AABB HIERARCHY. the entire human scalp. Thus contrary to the common belief that fiber based hair simulation is impossible we have shown the feasibility of this approach in one important part of the entire problem consisting of simulation, collision detection, collision response and visualization. Of course the integration of appropriate collision response mechanisms—such as an adaption of the methods used for cloth-modeling [3] into the context of hair modeling—is a remaining challenge. Fig. 7. Running times for the construction of the wrapped spherical hierarchy and the chain-aligned AABB hierarchy. References [1] Stuart S. Antman. Nonlinear Problems of Elasticity, volume 107 of Appl. Math. Sci. Springer-Verlag, Berlin and New York, 1995. [2] Sigal Ar, Gil Montag, and Ayellet Tal. Deferred, self-organizing bsp trees. In Proceedings of the EUROGRAPHICS 2002, volume 21(3), 2002. [3] R. Bridson, R. Fedkiw, and J. Anderson. Robust treatment of collisions, contact and friction for cloth animation. In SIGGRAPH 2002, ACM TOG 21, pages 594–603, 2002. [4] Joel Brown, Jean-Claude Latombe, and Kevin Montgomery. Real-time knot tying simulation. The Visual Computer, 20(2):165–179, 2003. Fig. 8. Running times for the cascade verification of the wrapped spherical hierarchy and the update of the chain-aligned AABB hierarchy. Fig. 9. Times for collision detection using the wrapped spherical hierarchy and the chain-aligned AABB hierarchy. [5] F. Ganovelli, J. Dingliana, and C. O’Sullivan. BucketTree: Improving collision detection between deformable objects. In SCCG2000 Spring Conf. on Comp. Graphics, 2000. [6] Bernd Gärtner and Sven Schönherr. Smallest enclosing ellipses - fast and exact. [7] S. Gottschalk, M. C. Lin, and D. Manocha. OBBTree: A hierarchical structure for rapid interference detection. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 171–180, 1996. [8] Leonidas Guibas, An Nguyen, Daniel Russel, and Li Zhang. Collision detection for deforming necklaces. In Proceedings of the eighteenth annual symposium on Computational geometry, pages 33–42, 2002. [9] Dan Halperin and Mark H. Overmars. Spheres, molecules, and hidden surface removal. In Symposium on Computational Geometry, pages 113– 122, 1994. Fig. 10. Running times for update and collision detection with the chainaligned AABB hierarchy in a dynamic deformable scene (150 fibers with 100 segments each are deformed from a straight configuration to a loop). The third curve indicates the number of collisions found in each deformation step. [10] Doug L. James and Dinesh K. Pai. BD-Tree: Output-sensitive collision detection for reduced deformable models. ACM Transactions on Graphics (SIGGRAPH 2004), 23(3):3–18, 8 2004. [11] J.T. Klosowski, M. Held, J.S.B. Mitchell, H. Sowizral, and K. Zikan. Efficient collision detection using bounding volume hierarchies of kDOPs. In IEEE Transactions on Visualization and Computer Graphics, volume 4(1), March 1998. [12] Thomas Larsson and Tomas Akenine-Möller. Collision detection for continuously deforming bodies. In Proceedinsg of Eurographics, short presentations, pages 325–333, 2001. [13] Itay Lotan, Fabian Schwarzer, Dan Halperin, and Jean-Claude Latombe. Efficient maintenance and self-collision testing for kinematic chains. In Proceedings of the eighteenth annual symposium on Computational geometry, Barcelona, Spain, pages 43–52, 2002. [14] John Maddocks. Mathematical modelling of DNA, course notes, 2001. http://lcvmwww.epfl.ch/ [15] Nimrod Megiddo. Linear-time algorithms for linear programming in R3 and related problems. In Proc. 23rd Annu. IEEE Sympos. Found. Comput. Sci., pages 329–338, 1982. [16] Dinesh K. Pai. STRANDS: Interactive simulation of thin solids using Cosserat models. In Computer Graphics Forum, volume 21(3), pages 347–352, 2002. [17] I.J. Palmer and R.L. Grimsdale. Collision detection for animation using sphere-trees. In Computer Graphics Forum, volume 14(2), pages 105– 116, 1995. [18] Sean Quinlan. Efficient distance computation between non-convex objects. In IEEE Int. Conference on Robotics and Automation, pages 3324–3329, 1994. [19] C.R. Robbins. Chemical and Physical Behavior of Human Hair. Springer-Verlag, New York, 4. edition, 2002. [20] Gino van den Bergen. Efficient collision detection of complex deformable models using aabb trees. Journal of Graphics Tools, 2(4):1–14, 1997. [21] Kelly Ward and Ming C. Lin. Adaptive grouping and subdivision for simulating hair dynamics. In Proc. of Pacific Graphics, 2003. [22] Emo Welzl. Smallest enclosing disks (balls and ellipsoids). In H. Maurer, editor, New Results and New Trends in Computer Science, number 555 in LNCS, pages 359–370. Springer, August 1991. [23] Yunhong Zhou and Subhash Suri. Collision detection using bounding boxes: Convexity helps. In Algorithms - ESA 2000, 8th Annual European Symposium, pages 437–448, 2000.