Comp 259: Project Final Report

Comp 259: Course Project (Spring 2002)

Hair Simulation

Final: May 10, 2002
Kelly Ward

[ Overview | Hairstyling | Collision Detection | Other Work | Future Work ]

Updated Hair Research Information, through 2005

Simulating hair is a costly computation due to the complex nature of hair. The cost of performing this simulation is improved by using simulation levels of detail. The hair is represented as one of three representations. The coarsest level is a patch representation, which is modeled as a Dynamics NURBS (D-NURBS) surface. Next is the cluster representation, which is a generalized cylinder that is covered by a D-NURBS surface. And the finest representation is to model individual strands as D-NURBS curves.

With the following framework in place, there are still a number of ways to improve the performance of the simulation as well as make the hair modeled look more realistic. In order to add to the realism of the hair simulation, an interactive hairstyling tool was created in order to make a wider variety of hairstyles in an intuitive and easier manner than had previously been done. Also, friction was added to the simulation during collision detection with the head so that the hair no longer appeared to slide along the head. Instead, the hair moves in a much more natural way around the head.

One of the most expensive calculations to perform is collision detection for the hair since there are so many hairs on a human head. The collision detection between the hair and the head was improved which is explained in the section on hair-head collision detection below. Also, hair-hair collision detection is added to the simulation.

Hairstyling

Moreover, the user can save a hairstyle so that it can be loaded again later. The ability to load old hairstyles onto a model makes creating new hairstyles even easier. The user can load an old hairstyle and then add more hair sections to the model or change the parameters of the existing hair. Therefore, a style that was short and straight can quickly become long and curly.

In addition, the user can control certain level of detail parameters. The user can specify if certain sections of hair should not fall below a certain LOD (such as around the part of the hair where more detail is desired) or that a certain section of hair should not go above a certain LOD (if the hair is at the bottom of the head and will not be seen much).

Figure 1: Different hairstyles created with hairstyling tool.

Collision Detection

Head-Hair

The collision detection between the hair and the head model was performed using hierarchies of Swept Sphere Volumes (SSVs). This is implemented using the PQP package. These hierarchies were created around the head model as well as around each section of the hair. A section of hair is defined by the rigid sections of the underlying skeleton model (the sections between each control point). Figure 2 shows an SSV around a section of hair. The SSV depicted is a line swept sphere (LSS).

The collision response is calculated by determining the normal direction on the head where the hair is colliding. The normal direction is calculated based on the normal directions of the triangles of the head model that are in collision with the hair. Using the hierarchy in order to perform collision detection enables the exact triangles that are in collision with the hair to be returned. Since the normal direction of the hair sections are not needed for collision response the triangles that define the hair section are not needed. Furthermore, the highest level SSV of the hierarchy provides a tight fit around each hair section. Thus, when the highest level SSV of a hair section collides, this provides enough information for the collision and the rest of the hierarchy is not necessary. By not using the hierarchy of SSVs for the hair model during collision detection, the simulation saves a lot of computation time. Figure 3 shows a chart of the performance gain by not using a hierarchy for the hair model in comparison to when a hierarchy was used. There was a noticeable increase in performance that was noticed for each hair representation (patches, clusters, and strands).

Figure 2: An SSV around a section of hair.

Figure 3: Performance increase in hair-head collision detection

Hair-Hair

Since there are so many hair sections used to simulate hair, it does not seem appropriate to use the same method of collision detection for hair-hair collision detection as head-hair collision detection. Also, each section of hair is only going to come in contact with a small percentage of the other hair sections. Rather than performing collision detection tests with every other hair section, we can save calculation time by only performing intersection tests with other hairs that are near it. We break up the space around the head into cells or grids and only perform intersection tests on hair sections that fall within the same cell.

In order to break the area up into cells, the HCollide collision detection package was used. HCollide is setup to insert triangles into these grids and perform bounding box intersection tests against each other. A section of hair is inserted into the cells based on the placement of its underlying skeleton model. The skeleton model also serves as the line of the corresponding LSS. Hair-hair collision between patches and other sections of hair is ignored since the patch is only used when the hair is not very visible to the viewer. The ability to insert lines into the cells instead of triangles was added to the collision package. Each line segment of the hair sections are then added to the cells. The cells that contain more than one section of hair are then returned so that SSV intersection tests can be performed on them. Figure 4 shows the grids layout of the hair.

Figure 4: Hair falling into grids around the head for collision detection

Other Work

When the hair was adjusted to follow the motion of the head it became apparent that the hair appeared to slide around the head. As the head moved, the hair appeared to remain in one place. Frictional forces were added to the simulation in order to make the hair move in a more realistic manner. When the hair collides with the hair, a frictional force is calculated and added to the hair. The frictional force is calculated by using the normal direction of the head at the point where the hair is contacting it and finding the plane tangent to the head at that point. Then, all of the forces that are acting on the hair are summed and projected on the plane. The frictional force is then applied in the opposing direction as that force since friction opposes the motion of an object. The result is a much more realistic looking motion to the hair. When the head moves, the hair seems to be pushed by the head as opposed to sliding frictionlessly around it.

Future Work

Hair-hair collision response - Different than response for head-hair collision
Fly-Aways around silhouettes
Speed-up hair-hair collision detection
Hairstyling
- More flexibility to allow ease of creating hairstyles
- Automatic creation of LODs
Model different hair types (fine vs. coarse)
Substances in hair: hairspray, gel, water, etc.

http://www.cs.unc.edu/~wardk/
wardk@cs.unc.edu