Collision detection & Response



by Telemachos




Read this first! As of 21/07 2003 this document is outdated and a new and much improved version is available here. The text below contains algorithmic errors that has been fixed in the new version.


Introduction

Hi again guys'n'gals.

Today's topic is collision detection in a general 3d environment made from a collection of triangles. We'll take a look at how to detect collisions in a fairly fast and precise way but also on how to handle collisions when they occur in a way a player would find normal in a 3d computer game. In the approach I'll discuss here we'll approximate the body we wish to move through 3d space with a sphere or ellipsoid because that's a fairly close approximation to most player controlled shapes (like a humanoid or animal). Also spheres and ellipsoids has a shape which is very "friendly" to the environment when a collision actually occurs. They have no sharp edges which can get us stuck and their smooth bodies glides much more easily through the geometry.

The algorithm was first presented by Paul Nettle a.k.a. Midnight on www.flipcode.com under the name "Generic Collision Detection For Games Using Ellipsoids". My original motivation for writing a new tutorial on the subject was to correct a few flaws in the original algorithm. Since I have had a very interesting email conversation with Paul and the result is that his document is now available in an updated version which includes new ideas from both of us. I now consider the two documents (his and mine) to be quite closely related and both has strengths and weaknesses :

Paul's tutorial describes the technique in a fairly short way and provides pseudo-code for the outlines of an implementation. Paul's tutorial might be the best for people already skilled in linear algebra (vector math in particular) and people who just wants the idea and then work out the details of an implementation themselves. Get it here : www.GraphicsPapers.com (in the "docs" section)

This text contains a more detailed explanation of the areas I think people might find hard to understand. I will offer actual C++ code for the functions used and explain the math behind most of the theory. I'll also talk about things which are not really algorithmic question but more questions found in implementation. For example : how do we keep track of collisions if working with big levels split up into multiple meshes? How do we detect and handle errors which are bound to occur now and then because of the limited precision of floating point math. I will also talk about collision detection in a non-static environment (objects which might move like doors swinging open) and describe how to implement lifts which movement will actual "push" the player with them.

Update 3/11 2002: Since I originally wrote this document it has come to my attention that the algorithm can fail when colliding from a sharp angle with a low surface - for example a step on a staircase. The problem is when calculating the polygon intersection point (described somewhere near the middle of this page) and the plane intersection is found to not be within the polygon. Then I describe how you should chose the nearest point on the edge of the triangle - this is not always correct! But unfortunately I cannot tell you how to calculate the correct polygon intersection point as I have had no time to figure it out myself. Do send a mail if you know how to calculate this point correctly. It happens fairly rarely and when you detect such an "edge collision" you can use a brute force subdivision along the velocity vector to figure out where the sphere and polygon collides.

The demo & source code:

I tried to think of a way I could present code for a full implementation of this algorithm in a clear and non-confusing way. After some thought I decided that it would take too much "wrapping" code to give you full source for a demo on collision detection. The important points would be buried in code on geometry loading, texture loading, d3d setup etc. etc.

Therefore what I'll do is to give you C++ source files with all the coded needed to implement collision detection but not in a program which compiles and runs. You'll have to add it to your own engine yourself!! On the other hand, if you cannot do that then reading this tutorial might be a bit premature!

You can also browse the source files online :
  • vectormath.h
  • vectormath.cpp
  • collision.cpp
  • response.cpp


  • In case you don't trust me blindly and wants proof of the algorithm before you invest hours in an implementation I'll also include binaries of my own 3d engine along with a scene which demonstrates all the features we discuss here - including the advanced stuff we'll discuss last like collision against dynamic geometry !! The code used in that demo is almost the same as the one I'll give you with the exception of certain engine specific features to handle the scripted animations and stuff like that. Sorry, I wont give you the source for that engine - so don't bother to ask ;)

    Download the entire package with the demo and this html document (1.3mb):
  • pxdtut10.zip


    How to read this document :

    First of all let me say : This stuff is NOT easy!! It'll take a long time to understand correctly and probably several readings of this text. So give yourself time to fully understand what's written in a section before moving on to the next. To fully understand the theory presented you will need to have a fairly good idea about linear algebra - i.e. vectors, matrices, vector spaces and the like.
    The tutorial is split up in two major parts - one on collision detection and one on collision response. Now what's the difference I hear you ask ?

  • Collision detection deals with finding out whether or not we'll bump into something when we move our ellipsoid though 3d space and if so, what do we hit and where ?
  • Collision response deals with what to do when a collision DO occur. Do we stop? Do we continue movement in another direction ? What do we do ? Collision response is important for the FEEL of the game you produce.
  • We will assume that the mesh is placed in the same space as the player - in world space. In other words we assume that each mesh's world matrix is the identity matrix. I'll discuss moving or displaced geometry in the section called "Dynamic geometry".


    Vector spaces & translations between them :

    This algorithm makes heavy use of non-standard vector spaces to simplify certain calculations. Those of you already familiar with the definition of vector spaces and change-of-basis matrices to move between spaces can skip this section entirely. The rest of you read on but be aware that this is only a very quick crash-course on this matter. A full explanation would include too much math and is beyond the scope of this tutorial anyway.

    OK, most of you will never have encountered any other vector spaces than the standard two and three dimensional spaces we usually think of from school math. These are called R2 and R3 and are those we use in a standard coordinate system (like the one d3d or openGL operates in). As it turns out we can easily define vector spaces or 4,5 or even infinitely (!!) many dimensions by listing a few rules (axioms) which every vector space must follow. R2 and R3 follows these rules too!

    Of course humans cannot visualize spaces of dimension greater than 3 and this might be the greatest challenge when trying to work with such spaces. Still such spaces can be useful to a game programmer. To give an example I can mention the quaternion which can be thought of as vectors in 4D. They are sometimes used to represent orientations for animations because we can interpolate smoothly between them by using the shortest arc on a 4D unit sphere (this is called SLERP). The 3D alternative is to store orientations in matrices and we all know you cannot interpolate between two matrices. Don't worry though, we'll stay in 3D today - the above just goes as a motivation for the use of "non-standard" vector spaces :)

    OK, so here are the rules or axioms every vector space must follow :
  • It must be closed under addition and scalar multiplication. This means that given two vectors from the vector space, say X and Y, then Z = X + Y must also be in the same space. For any real number r (1, 3, 0.23423 etc.) if X is in the space then so must r*X. Informally said we cannot ever "leave" the vector space using these operations.
  • It must contain a zero-vector V0 so for all vectors X in the space V0 + X = X. For all vectors X is must also be true that X*0 = V0.
  • it must follow standard mathematical rules such as the associative law, commutative law etc.
  • Now I'll introduce to you the concept of a linear combination of vectors from a vector space. Said in words it's what you get if you combine vectors from the space using addition and multiplication with scalars (real numbers). For example if you have vectors X, Y and Z you could combine them like this : V = 2*X + 4*Y + (-3)*Z

    OK, what can we say about V? Well, from rule number 1 above we can see that V must belong to the same vector space as X, Y and Z. Cool, that's what we need to really define a basis for a given vector space.

    A basis for a vector space is a collection of vectors belonging to the space for which the following holds :
  • Every vector in the vector space is a linear combination of vectors from the basis.
  • Each of the vectors in the basis can NOT be made from a linear combination of the other basis vectors.
  • The minimal number of vectors necessary to create a basis for a vector space is equal to the vector space's dimension. So how many vectors are in R3s basis ?

    I think it's time to take a break from the abstract definitions above and look at a familiar example. After reading and understanding this example you should understand the stuff above good enough for this tutorial. Lets look at the standard 3D vector space we know as R3. Is it a vector space ? Well, we check the rules (very informally) :
  • Is it closed under addition ? If we have two vectors from R3, say x=(2,5,4) and y=(6,3,8), is z=x+y in R3 ?
    Well z = (2,5,4)+(6,3,8) = (8,8,12) - we believe this is in R3. For any real number r, say r=2, is r*x in R3 ?
    r*x = 2*(2,5,4) = (4,10,8). Check! First rule holds.
  • Is there a zero vector ? Yes, we use V0 = (0,0,0) and the rule holds. (You can check)
  • Third rule involves a lot of checks which are tedious so I'll just mention the commutative law saying that x + y = y + x. Obviously that holds..
  • So R3 is a vector space! What is its basis ? The basis for R3 is called the 'standard basis' and contains 3 vectors (e1,e2,e3). What are those vectors ? They are :
  • e1 = (1,0,0)
  • e2 = (0,1,0)
  • e3 = (0,0,1).
  • Well, if that is indeed the basis for R3 then we should be able to derive ANY vector from R3 as a linear combination of those vectors right ? Well, we'll have to convince ourselves about this so given any vector from R3, say v=(5,2,4), can we derive this from a linear combination of (e1,e2,e3)?
    Yes, because 5*e1 + 2*e2 + 4*e3 = 5*(1,0,0) + 2*(0,1,0) + 4*(0,0,1) = (5,2,4) = v
    It's clear that e1 cannot be made from e2 and e3 because no matter how you combine e2 and e3 the first vector component would still be zero right? Same can be said for the other two basis vectors.

    Now for an important definition! The coordinate for a point in a vector space is the scalars multiplied to the basis vectors in the linear combination for the point!! In the example above the point v has the coordinate (5,2,4). This might seem obvious because we're used to working with R3 but make sure you see WHY that is indeed its coordinate :
    v = 5*e1 + 2*e2 + 4*e3 so the scalars are (5,2,4) - get it ?

    By now you should be convinced that R3 is a vector space following the axioms for general vector spaces. It should also be clear that other vector spaces could exist - even other spaces of 3 dimensions. For example what about the vector space made from the basis v1=(2,0,0) v2=(0,2,0) and v3=(0,0,2). Which vectors does this space contain ? The same ones as R3 of course but with different coordinates!!
    Given a point (2,2,2) in R3 - which coordinate would that same point have in our new space ? It would have the coordinate (1,1,1) because:
    1*v1 + 1*v2 + 1*v3 = 1*(2,0,0) + 1*(0,2,0) + 1*(0,0,2) = (2,2,2).

    This is exactly the trick we'll use in this tutorial to create a very special vector space which has the property that the ellipsoid we're using to model our player with becomes a unit-sphere in that space!! A unit sphere is a sphere with radius = 1 but it's also an ellipsoid with radius VECTOR = (1,1,1), right?

    So how do we create such a space ? Well, we'll have to come up with a basis for the space. If the ellipsoid is defined by a radius vector of (x,y,z) we'll just chose the basis to be :
  • v1 = (x,0,0)
  • v2 = (0,y,0)
  • v3 = (0,0,z)
  • Now in this space the ellipsoids radius vector will be (1,1,1), right? And you can check if that choice really is a valid vector space if you want - or you can take my word for it (generally not recommended ;)

    The LAST thing we need to know is how to translate a point from R3 into our new vector space. We're in luck, because it turns out that because linear combination is a linear function we can get away with just describing what to do with the 3 basis vectors (e1,e2,e3) (remember these from above?). We find that if we use the basis (v1,v2,v3) from above, then :
  • e1 = 1/x * v1 + 0 * v2 + 0 * v3
  • e2 = 0 * v1 + 1/y* v2 + 0 * v3
  • e3 = 0 * v1 + 0 * v2 + 1/z* v3.
  • We list the coordinates (remember what are those??) for e1,e2 and e3 as column vectors in a 3x3 matrix and get a 'change of basis' matrix which has the effect of translating any vector from R3 into our special vector space. The matrix looks like this :


    As long as we stick to defining new spaces for which the change of basis matrix looks like above (only has entries on the diagonal) then we can simplify the multiplication of a vector to the matrix. To do that we'll need to be able to multiply vectors.
    Of course you cannot multiply or divide two vectors but we CAN define what we mean if we write so anyway ;) We'll define v1 * v2 to mean component-wise multiplication so :
    v1*v2 = (v1.x*v2.x, v1.y*v2.y, v1.z*v2.z) - division is defined in a similar way.

    If we do that, and create spaces like described above, then we can avoid having to use a 3x3 matrix for space translations and just say, that to translate a vector from R3 into our space we just multiply it with a vector describing the basis of our new space. For example we translate x = (3,4,3) into our example space like this :
    x2 = x * (1/x, 1/y, 1/z) - or to simplify even further : x2 = x / (x,y,z).

    This ONLY holds for change of basis matrices which only has entries on the diagonal (sorry, I know I said that before) and those spaces just happens to be those you get from R3 when you multiply each basis vector (e1,e2,e3) with a constant. If you check back, that's exactly what we do to achieve the 'scaling trick' above.



    Collision detection :

    OK, enough talk - lets get started!
    So, we have our guy we wish to move around the world. He's represented by an ellipsoid with a center (which we will consider his position) and a radius vector defining the ellipsoids size along the three axis. Like this :


    We move our guy though the world by applying a force to him in some direction. This is represented by a velocity vector. We wish to move the ellipsoid through the world so its new position equals its current position plus the velocity vector.


    We don't know if we can do that though as there might be something in his way - one or more of the triangles making up our world might be obstructing his way. We have no chance of knowing which triangle he might hit, so we'll have to check them all (another motivation to split geometry up in smaller bits). Also we cannot stop once we find ONE triangle he'll hit - we'll have to check ALL potential colliders and prepare to collide with the closest one.
    Below I have tried to illustrate this by showing what happens if we detect a collision with triangle A and then stops without checking the rest, including triangle B :


    First of all we'll reduce the complexity of the problem greatly by translating everything into something we'll call ellipsoid space, which is a vector space where the ellipsoid is really a unit sphere (See section on vector spaces). If we believe we can do that, then for the rest of this tutorial we can assume our guy is a unit sphere. I believe that..

    OK, so we have to check every face. The check for each face can be split into 5 smaller steps which will give us the following pseudo-code for collision detection :
    For every triangle do:
  • Calculate the plane the triangle lies on.
  • Calculate the intersection point on the unit sphere where it will hit the plane.
  • Calculate the point on the plane where the unit sphere collide.
  • Check if this plane intersection point is in the triangle. If not calculate the polygon intersection point (if any).
  • If we hit the polygon and the distance to the intersection is less or equal to the length of the distance we wish to move we check if this is the closest hit so far. If it is we write down info about this collision.
  • When this algorithm terminates we'll have written down the closest collision (if any) and we can then handle this information as we'll discuss in the section on collision response.
    When the entire process of collision detection and response ends we'll translate the final result back into standard R3 space to get our final, updated player position.

    Below each step is explained in greater detail but make sure you understand the ideas in the pseudo-code above before you read on. From now on I'll present theory and then implement that in real C++ code. When finished reading you should study the code for collision detection in its whole form in the file 'collision.cpp' (using vectormath.cpp and vectormath.h) found in the downloadable package.
    Note that the function in collision.cpp is called from the code taking care of collision response (in response.cpp)


    1) Calculate the plane of a triangle :

    This is the easy step! You should all be familiar with planes but let me summarize real quick what they are in case you forgot. A plane is infinite! That's probably the most important fact you should remember about planes when we're discussing collision detection. Collision with a triangles PLANE is NOT necessarily a bad thing and it will most likely happen all the time. A triangle is a closed subset of the plane it lies on. For example a triangle in our world in which all its points has a z-value of 0 lies on a plane - namely the infinite xy-plane which consists of all points in the set A = { (x,y,z) | z=0 } (which is read " all points (x,y,z) for which z is 0. )

    For our use we'll define a plane as a point of origin and a normal vector perpendicular to all vectors which lies on the plane.
    Given a point on the plane and any two (non parallel) vectors lying on the plane we can calculate the plane normal by using the wedge product (also known as the cross product).
    What we want is the plane on which our triangle lies so we just pick any of the triangles points as the origin of the plane (they all lies on it). Let say our triangle are defined in a clockwise order like this :


    then we chose point A as origin and get two non-parallel vectors on the plane by making each segment outgoing from A a vector like this :
  • v1 = B - A;
  • v2 = C - A;
  • The plane normal is then given by the wedge product of those two vectors :
    inline D3DVECTOR wedge(D3DVECTOR v1, D3DVECTOR v2) {
     D3DVECTOR result;
     
     result.x = (v1.y * v2.z) - (v2.y * v1.z);
     result.y = (v1.z * v2.x) - (v2.z * v1.x);
     result.z = (v1.x * v2.y) - (v2.x * v1.y); 	
     
     return (result);	
    }
    
    D3DVECTOR pOrigin = A;
    D3DVECTOR pNormal = wedge(v1,v2);
    
    NOTE : The order of the parameters for this function with is important! It defines the direction of the normal which basically decides if we're looking at a back face or not! Actually it holds as a general rule that wedge(v1,v2) = -wedge(v2,v1).
    Calling it like above is the correct way if your triangles are defined in clockwise order. We need the normal normalized (its length should be 1) so we also call this function on it :
    inline void normalizeVector(D3DVECTOR& v) {
     double len = sqrt(v.x*v.x + v.y*v.y + v.z*v.z);	
     v.x /= len;
     v.y /= len;
     v.z /= len;
    }
    
    normalizeVector(pNormal);
    
    That's it, we now have the plane on which the triangle lies.

    Remember that because we work in this special ellipsoid space we need to scale our triangle into that same space as described in the section on vector spaces (by dividing all its points with the ellipsoid radius vector)


    2) Calculate the theoretical unit sphere intersection point:

    In this step we try to predict which point on the unit sphere which will eventually hit the triangle we're checking against now. And we'll calculate that based on only the position of the sphere and the plane the triangle lies on. Actually that's not really possible so what we will calculate is the point on the sphere which will hit the PLANE the triangle lies on.

    That is not easy to do with an general ellipsoid but here we find the first benefit of using our special ellipsoid space because in that space the problem reduces to figure out which point on a unit sphere that will hit the plane! This is very easy to do and as you can see from the picture below it boils down to adding the reversed plane normal to the center of the sphere. As the normals length is 1 (it's normalized) and the sphere is a unit sphere (its radius is 1) the length is even correct.

    If for some reason you can't work with collision in ellipsoid space you can email me and ask about how to calculate the intersection point for an ellipsoid but that's a lot of hairy code so I'll not include it here.


    There is a reason why the drawing might seem a little complex! What I have tried to illustrate is that we calculate the sphere intersection point before we move the sphere! So it is the point where the sphere will hit the plane if it is moved along its velocity. The dotted line represent how the plane would be if we moved the sphere close to it.

    So the code for this step is simple :
    eIPoint = source - pNormal;     
    


    3) Calculating plane intersection point :

    OK, next step is to figure out where on the plane the sphere will make contact when moved along the velocity vector. This is simply done by casting a "ray" from the point on the unit sphere calculated above (the sphere intersection point) in the direction of the velocity vector.


    For that we'll need a few new functions :
    inline float dot(D3DVECTOR& v1, D3DVECTOR& v2) {
     return ( v1.x * v2.x + v1.y * v2.y + v1.z * v2.z );	
    }
    
    double intersectRayPlane(D3DVECTOR rOrigin, D3DVECTOR rVector, D3DVECTOR pOrigin, D3DVECTOR pNormal) {
      
      double d = - (dot(pNormal,pOrigin));
      double numer = dot(pNormal,rOrigin) + d;
      double denom = dot(pNormal,rVector);
      
      if (denom == 0)  // normal is orthogonal to vector, cant intersect
       return (-1.0f);
      return -(numer / denom);	
    }
    

    The parameters should be obvious. One thing to note though is that rVector, the direction of the ray, must be normalized! What this function returns is the the distance from the ray origin to the plane along the ray so to get the plane intersection point we do :
    (eIPoint = ellipsoid intersection point from step 2, pIPoint = plane intersection point)
    // shoot ray along the velocity vector
    distToPlaneIntersection = intersectRayPlane(eIPoint, normalizedVelocity, pOrigin, pNormal);
              
    // calculate plane intersection point
    pIPoint.x = eIPoint.x + distToPlaneIntersection * normalizedVelocity.x; 
    pIPoint.y = eIPoint.y + distToPlaneIntersection * normalizedVelocity.y; 
    pIPoint.z = eIPoint.z + distToPlaneIntersection * normalizedVelocity.z; 	
    
    But there is a catch! The code above assumes that the plane is in front of the sphere. That might not be the case. The plane can span the sphere like shown below. If that happens this step isn't really very useful. Still we need some point to use as plane intersection point and we choose to use the point we get by casting a ray from the sphere intersection point along the plane normal.


    We only want to do this if the plane spans the sphere and how do we check this ?
    We need a few more functions :
    #define PLANE_BACKSIDE 0x000001
    #define PLANE_FRONT    0x000010
    #define ON_PLANE       0x000100
    
    DWORD classifyPoint(D3DVECTOR point, D3DVECTOR pO, D3DVECTOR pN) {
    
     D3DVECTOR dir = pO - point;
     double d = dot(dir, pN);
     
     if (d<-0.001f)
      return PLANE_FRONT;	
     else
     if (d>0.001f)
      return PLANE_BACKSIDE;	
    
    return ON_PLANE;	
    }
    
    The function above checks whether a point is in front of, in back of, or on a plane. If the sphere intersection point is on the backside of the plane then the plane spans the sphere. (Because we'll only looks at faces where our source point is in front)

    So our code to handle spanning planes becomes :
    DWORD pClass = classifyPoint(eIPoint, pOrigin, pNormal);
         
    // find the plane intersection point
    if (pClass == PLANE_BACKSIDE) { // plane spans sphere
         
      // find plane intersection point by shooting a ray from the 
      // sphere intersection point along the planes normal.
      distToPlaneIntersection = intersectRayPlane(eIPoint, pNormal, pOrigin, pNormal);
                     
      // calculate plane intersection point
      pIPoint.x = eIPoint.x + distToPlaneIntersection * pNormal.x; 
      pIPoint.y = eIPoint.y + distToPlaneIntersection * pNormal.y; 
      pIPoint.z = eIPoint.z + distToPlaneIntersection * pNormal.z; 	
    } 
    
    Now we have calculated a plane intersection point no matter if the sphere spans the plane or not. Let us move on to the next step...

    4) Calculate polygon intersection point (if any) :

    OK, almost there! Now it's time to calculate the point we're really interested in - the point where the current triangle will be hit. There are two cases : either we're in luck and the plane intersection point we calculated above is contained in the triangle. Or it's not and we'll have to do a little more work.

    If the plane intersection point is in the triangle then that point is actually where the triangle will be hit and we got our polygon intersection point. If not we have the situation below and both plane intersection point AND sphere intersection point is wrong and will have to be recalculated.

    In this case the polygon intersection point is the point on the edge of the triangle closest to the plane intersection point and to get the correct sphere intersection point we have to cast a ray from that point along the reversed velocity vector and see if that ray hits the sphere! The above three lines might be hard to understand but read them until you do! The drawing might help you some.


    So, it boils down to figuring out if the point is in the triangle or not. What do we know that can help us do that. Well, at least we know the point lies on the same plane as the triangle, so we can reduce the problem to a 2d problem. Imagine you draw a triangle on a piece of paper, closes your eyes and draws a point randomly on the paper. How can you tell if the point is within the triangle ? Well, you could look but that's cheating ;)
    No, lets say you are allowed to draw 3 lines from the point to the 3 points in the triangle. You are also told the angle between any 2 of those lines. Then you would know the answer because if the point lies within the triangle then the sum of those angles would be 360 degrees - or 2*PI. If the point is outside the triangle then the sum will be less. Check the drawing below :


    Here is the code to calculate this. (remember that the dot product of two unit vector equals the cosine to the angle between them)
    BOOL CheckPointInTriangle(D3DVECTOR point ,D3DVECTOR a, D3DVECTOR b, D3DVECTOR c) {
      
      double total_angles = 0.0f;
           
      // make the 3 vectors
      D3DVECTOR v1 = point-a;
      D3DVECTOR v2 = point-b;
      D3DVECTOR v3 = point-c;
      
      normalizeVector(v1);
      normalizeVector(v2);
      normalizeVector(v3);
    
      total_angles += acos(dot(v1,v2));   
      total_angles += acos(dot(v2,v3));
      total_angles += acos(dot(v3,v1)); 
        
      // allow a small margin because of the limited precision of
      // floating point math.
      if (fabs(total_angles-2*PI) <= 0.005)
       return (TRUE);
         
      return(FALSE);
    }
    
    But as mentioned above this function only solves ONE of two possible cases. If the plane intersection point is NOT contained in the triangle then we have to calculate the closest point on the edge of triangle. This is done by taking the smallest value of three 'shortest distance to line' calls which you should all know from standard school math. So :
    D3DVECTOR closestPointOnLine(D3DVECTOR& a, D3DVECTOR& b, D3DVECTOR& p) {
    	
       // a-b is the line, p the point in question
       D3DVECTOR c = p-a;
       D3DVECTOR V = b-a; 
       double d = lengthOfVector(V);
          
       normalizeVector(V);  
       double t = dot(V,c);
       
       // Check to see if ‘t’ is beyond the extents of the line segment
       if (t < 0.0f) return (a);
       if (t > d) return (b);
       
       // Return the point between ‘a’ and ‘b’
       //set length of V to t. V is normalized so this is easy
       V.x = V.x * t;
       V.y = V.y * t;
       V.z = V.z * t;
               
       return (a+V);	
    }
    
    
    D3DVECTOR closestPointOnTriangle(D3DVECTOR a, D3DVECTOR b, D3DVECTOR c, D3DVECTOR p) {
    	
      D3DVECTOR Rab = closestPointOnLine(a, b, p);
      D3DVECTOR Rbc = closestPointOnLine(b, c, p);
      D3DVECTOR Rca = closestPointOnLine(c, a, p);
        
      double dAB = lengthOfVector(p-Rab);
      double dBC = lengthOfVector(p-Rbc);
      double dCA = lengthOfVector(p-Rca);
        
      double min = dAB;
      D3DVECTOR result = Rab;
      
      if (dBC < min) {
        min = dBC;
        result = Rbc;
        }
     
      if (dCA < min)
       result = Rca;
       
      return (result);	
    }
    
    Now all we need is the function which checks if a ray intersects with a sphere :
    double intersectRaySphere(D3DVECTOR rO, D3DVECTOR rV, D3DVECTOR sO, double sR) {
    	
       D3DVECTOR Q = sO-rO;
       double c = lengthOfVector(Q);
       double v = dot(Q,rV);
       double d = sR*sR - (c*c - v*v);
    
       // If there was no intersection, return -1
       if (d < 0.0) return (-1.0f);
    
       // Return the distance to the [first] intersecting point
       return (v - sqrt(d));
    }
    
    The whole process of finding the final sphere intersection point and polygon intersection finally boils down to a few lines of code :
    // find polygon intersection point. By default we assume its equal to the 
    // plane intersection point. In that case we already know the distance to it.
    polyIPoint = pIPoint;
    distToSphereIntersection = distToPlaneIntersection;
         
    if (!CheckPointInTriangle(pIPoint,p1,p2,p3)) { // if not in triangle
       polyIPoint = closestPointOnTriangle(p1, p2, p3, pIPoint);	
         
       distToSphereIntersection = intersectRaySphere(polyIPoint, -normalizedVelocity, source, 1.0f);  
                      
       // we cannot know if the ray will actually hit the sphere so we need this check
       if (distToSphereIntersection > 0) { 	
          // calculate true ellipsoid intersection point
          eIPoint.x = polyIPoint.x + distToSphereIntersection * -normalizedVelocity.x;
          eIPoint.y = polyIPoint.y + distToSphereIntersection * -normalizedVelocity.y;
          eIPoint.z = polyIPoint.z + distToSphereIntersection * -normalizedVelocity.z;
       }
    } 
    


    5) OK, yes or no - is this triangle hit or not?

    This final section is easy. Through the 4 steps above we have obtained three important things :
  • The distance along the velocity vector we must move to hit the triangle. -1 if no hit.
  • The point on the triangle which will be hit
  • The point on the sphere which will hit the triangle.
  • First of all the distance to the hit must be within reach of our current velocity vector. In other words we only hit the triangle if distToSphereIntersection is smaller than the length of the velocity vector.
    But even if we do hit the triangle we might have hit another triangle which is closer so we only save information about this hit if the distance is the smallest one yet. The code looks somewhat like this (depends on your data structures)
    // any chance of hit ?
    if ((distToSphereIntersection > 0) && (distToSphereIntersection <= lengthOfVector(velocity)) { 
         
         // if first hit, or closest hit so far
         if ((foundCollision == FALSE) || (distToSphereIntersection < nearestDistance))  {
               
           nearestDistance = distToSphereIntersection;
           nearestSphereIntersectionPoint = eIPoint;
           nearestPolygonIntersectionPoint = polyIPoint;
           foundCollision = TRUE;
         }
    } 
    
    That was ONE triangle, now all that remains is to repeat ;)


    Collision Response :

    So now you know how to check if your sphere collides with any of the triangles in your world given a certain velocity vector. But what do we do when a collision actually occurs? The easiest way to handle a collision is to simply stop and not allow the move when a collision is detected. On the other hand - that is not how the professionals do is it? No, we want fancy collision response like sliding along the walls, automatically climbing stairs and automatically bumping over smaller obstacles on the ground.

    Actually we turn out to be in luck since all these things come automatically from our implementation of sliding!

    So what exactly do we mean by sliding? Well, what we want to do if we collide with a triangle in our world is to move close to it, change the direction of our velocity vector and then continue movement in a new direction.


    I'll now introduce the idea of the sliding plane. The sliding plane is, as the name implies, the plane along which we will continue our movement. From the picture above you might get the idea that the sliding plane is identical with the plane on which the colliding triangle lies. In some cases this is also correct but as the picture below demonstrates it is not always the case :


    Lets generalize the idea of the sliding plane. What does the two planes above have in common ? They are both the tangent plane to the sphere in the sphere intersection point from the section on collision detection. So we have to find a way to calculate the tangent plane for the sphere, given a point on its surface. Remember how we represent planes in this tutorial ? We define them by a point and a normal to the plane, right ? We already have the point - namely the sphere intersection point - so our task at hand is really to calculate the normal to the tangent plane of the sphere in the sphere intersection point, right ?

    It turns out that once again our choice of working in a special ellipsoid space saves us from a lot of computations because the normal to the tangent plane at any point on a unit sphere is just the vector going from the point on the surface to the center of the sphere! This is not the case with the more general ellipsoid shape !!


    So calculating the sliding plane is easy in ellipsoid space and boils down two lines of code :
        D3DVECTOR slidePlaneOrigin = nearestPolygonIntersectionPoint;
        D3DVECTOR slidePlaneNormal = newSourcePoint - nearestPolygonIntersectionPoint;
    
    Well, this isn't really a satisfying explanation is it ? What if we hadn't worked in this fancy ellipsoid space ? Couldn't we have computed this sliding plane anyway ? If you couldn't care less you can skip the next section and continue reading at the section called : 'Collision Response - continued'. The rest of you, get ready for some math because of course we can calculate the sliding plane for any spherical or ellipsoid shape. The next section will also make it clear why the calculations are so simple for the unit sphere.


    The tangent plane of any spherical or ellipsoid shape:

    Anyone still with me ? Cool, lets get enlightened then :) Again this will have to be a little quick and I cannot cover every detail because then this would become a math tutorial. What I try to do is introduce exactly enough concepts for you to understand the important part of this section - the calculation of the general tangent plane.

    The answer lies in the theory of differential geometry of regular surfaces. I will not go into the definition of regular surfaces but think of them as certain 'smooth' surfaces in 3D - like the surface of an ellipsoid or sphere.

    Ok, first lets think about how a tangent plane is defined. Most of you will have an intuitive understanding of what it is but not all have been introduced to a precise definition before. Lets go one dimension down and look at what we know about tangents in 2D. From school we know that the tangent to a curve looks like on the picture below and is calculated by differencing its function:


    On the drawing above the curve is in 2D but we could actually consider it as lying on a plane in 3D right ? Namely the xy plane in standard 3D. What if I told you that we could have the curve lying on a regular surface like a sphere or ellipsoid! It would run over the surface and be parameterized by one single variable t. This is just like in 2D where the curves we know are parameterized by a single variable x (we consider a curve in 2D as a functionf(x) = y, right ? ). To do that we will have to change the definition of the function for the curve but it would still be a function of one variable and we would know how to differentiate it.

    In 3D a curve is defined by 3 functions, each one describing the value of one of the 3 coordinates given a parameter t. So if we call the curve A then its function becomes :
  • A(t) = (x(t), y(t), z(t))
    and to differentiate it we differentiate its 3 coordinate functions, so :
  • A'(t) = (x'(t), y'(t), z'(t)).

    OK, back to the task at hand - to determine the tangent plane of the regular surface. Like described in the section about calculating the plane of a triangle we'll need two vectors lying on the plane as well as a point. We already have the point because that's the point on the surface we're trying to get the tangent plane for, so what we need are the two vectors. Here is the trick : Say the point is called p. We make 2 distinct curves A and B running on the surface and both passing through p. Now if we differentiate both A and B we'll get to vectors A'(p) and B'(p). Those two vectors are both tangents to the surface and they span a plane - exactly the tangent plane !!


    Now the only problem is that we don't have those two curves A and B. And there really is no easy way of finding two curves running though any point on a sphere or ellipsoid. It might seem we have a problem but first lets take a look at how the equations for a sphere and an ellipsoid looks like:
  • A sphere of radius r is given by the points (x,y,z) where x^2 + y^2 + z^2 = r^2
  • An ellipsoid with radius vector (a,b,c) is given by the points (x,y,z) where x^2/a^2 + y^2/b^2 + z^2/c^2 = 1
  • That makes sense because if you insert a vector (r,r,r) in the equation for the ellipsoid you get exactly the equation for the sphere with radius r.

    Now we are saved by a theorem which (ironically enough) says something about a totally different thing than what we are looking for. Here is what it says :
  • Theorem : If a regular surface is given by S = {(x,y,z), f(x,y,z) = a} where f is a differentiable function from R3 to R and a is a constant, regular value of f - then S is orientable.
  • The formula we need is in the proof of this theorem but first we have to make sure our ellipsoid fits the theorems requirements. OK, we know that an ellipsoid is a regular surface. What do we use as the function f ? We let f(x,y,z) = x^2/a^2 + y^2/b^2 + z^2/c^2 and as described above our ellipsoid is exactly the set A given by :
  • A = {(x,y,z) | f(x,y,z) = 1}.
    From school we know that a function like f is differentiable and 1 sure is a constant. You'll have to trust me when I say that 1 is a regular value of f :)

    So, our ellipsoid fits the requirements of the theorem, so lets look at the proof :
  • Given any point p = (x0,y0,z0) on the surface we consider any parameterized curve A(t) = (x(t),y(t),z(t)) passing through p for some value of t, say for t=t0. So A(t0) = p, right ?
  • Because the curve lies on the surface we know that f(x(t), y(t), z(t)) = a, for all t. (Because the theorem talks about a surface where for all points on it we have f(p) = a)
  • If we differentiate the expression f(x(t),y(t),z(t)) = a on both sides with respect to t, then for t=t0 we get :
    f'x(p)*(dx/dt) + f'y(p)*(dy/dt) + f'z(p)*(dz/dt) = 0, where dx,dy and dz all are evaluated at t0.
    But this is the same as : (f'x(p), f'y(p), f'z(p)) dot (dx/dt, dy/dt, dz/dt) = 0.
  • What do we know about the dot product ? We know that if the dot product is 0 then the two vectors are perpendicular to each other.
    So we conclude that (f'x(p), f'y(p), f'z(p)) is perpendicular to (dx/dt, dy/dt, dz/dt) in particular.
  • But what is (dx/dt, dy/dt, dz/dt) ? That is exactly the tangent vector to the curve A, and as it were evaluated at t=t0 it is exactly the tangent vector to the curve A at the point p!!
    Thus it must lie on the tangent plane for the ellipsoid at the point p.
  • As the above holds for any curve on the surface we conclude that (f'x(p), f'y(p), f'z(p)) must be perpendicular to any vector in the tangent plane, and thus it must be the normal to the tangent plane at p.
  • Cool, now we have exactly what we need!! For any point p on the ellipsoid we now know that the normal to the tangent plane is (f'x(p), f'y(p), f'z(p)).

    So now we just have to differentiate f(x,y,z) = x^2/a^2 + y^2/b^2 + z^2/c^2 with respect to x,y and z.
  • f'x = 2*x / a^2 + 0 + 0 = 2*x/a^2
  • f'y = 0 + 2*y / b^2 + 0 = 2*y/b^2
  • f'z = 0 + 0 + 2*z / c^2 = 2*z/c^2
  • Voila, what you got is a normal-function which looks like this : N(x,y,z) = (2x/a^2, 2y/b^2, 2z/c^2), where (x,y,z) is a point on the sphere/ellipsoid and (a,b,c) is the radius vector. You can then normalize the result if you want.

    As a last interesting point we can now see why the calculations can be simplified in the case of a unit sphere. A unit sphere is really an ellipsoid with a radius vector (1,1,1). So the general formula can be simplified so given any point on a unit sphere the normal to the tangent plane becomes N(x,y,z) = (2x, 2y, 2z). But that is really just 2*(x,y,z). What is the length of that ? Well, as p=(x,y,z) is on the unit sphere we know that |p|=1 so the length of 2*(x,y,z) must be 2. We want to normalize the normal, so we divide each vector component by the length and get that the normalized normal is (2x/2, 2y/2, 2z/2) = (x,y,z).

    So the normal to the tangent plane for a unit sphere at the point p is the point itself, so if we reverse it (which will only affect which side of the plane is front) then we have that ALL normals to the sphere go from the point on the sphere to the center of the sphere. That is exactly what we used in the section above this one (check back).


    Check in vectormath.cpp to see how the entire function for calculating the tangent plane normal looks like.


    Collision response - continued :

    Ok, now we know how to calculate the sliding plane. Now we need to figure out what to do with it. Here is the idea :
  • Move our sphere as close as possible to the triangle we're colliding with.
  • Calculate the sliding plane based on this new position.
  • Project the original velocity vector to the sliding plane to get a new destination.
  • Make a NEW velocity vector by subtracting polygon intersection point from the new destination
  • Recursively call the entire collision detection routine with the new position and the new velocity vector!
  • Ok, the last step might come as a surprise but yes, collision detection is a recursive function and for each recursion we check all triangles all over again! We continue to recurse until either :
  • we do not hit anything, so we just update the position
  • the velocity vector gets very small
    Ok, what about this projecting thing ? Well, it just so happens that it does exactly what we want. It generates a vector on the sliding plane we can move along and the more directly we hit the triangle the less do we slide. Check the drawing :


    To accomplish all this we need to be able to project the velocity vector to a plane, but as it turns out we already have all the functions we need. We'll use the intersectRayPlane from the collision detection section and calculate the projection as the intersection we get by casting a ray from the destination point, along the sliding plane normal, to the sliding plane. So the code becomes :
      // Determine the sliding plane (we do this now, because we're about to
      // change sourcePoint)
      
      D3DVECTOR slidePlaneOrigin = nearestPolygonIntersectionPoint;
      D3DVECTOR slidePlaneNormal = newSourcePoint - nearestPolygonIntersectionPoint;;
      
          
      // We now project the destination point onto the sliding plane
      double l = intersectRayPlane(destinationPoint, slidePlaneNormal, 
                                   slidePlaneOrigin, slidePlaneNormal); 
        
      D3DVECTOR newDestinationPoint;
      newDestinationPoint.x = destinationPoint.x + l * slidePlaneNormal.x;
      newDestinationPoint.y = destinationPoint.y + l * slidePlaneNormal.y;
      newDestinationPoint.z = destinationPoint.z + l * slidePlaneNormal.z;
       
        
      // Generate the slide vector, which will become our new velocity vector
      // for the next iteration
      D3DVECTOR newVelocityVector = newDestinationPoint - collision->nearestPolygonIntersectionPoint;
    
      // Recurse. Pos is where we expect our final position
      finalPos = collideWithWorld(newSourcePoint, newVelocityVector);  
    
      // return the final result
      return finalPos;
    


    Gravity :

    What if we want gravity in our world ? Will that be hard ?
    No not at all, each frame you just make two calls to the collision detection routines - one time with the player velocity and another time with a gravity vector. You can combine the two vectors and then just call the routines once with a vector v = velocity + gravity but that will make climbing stairs become more difficult as the velocity vector must be much longer for the projected vector to slide upwards enough when you hit the edge of a stair (think about it).

    In most cases making two separate calls wont even affect the total numbers of calls to the functions because when moving along flat ground if you combine velocity and gravity you will have two recursions as shown on the drawing below. If you first move along the ground with the velocity you will not hit anything, and the second call with the gravity vector will collide and not recurse because the hit is perpendicular to the surface.




    Error handling :

    You're working with floats and floats only have a limited precision! Thus you are bound to get errors now and then. By errors I mean moving into solid geometry because the routines return the wrong results or falling though the ground. Still, in my humble opinions collision detection is one of the places where you should be extra careful not to have errors! It should never happen that the player falls through the ground and end up outside your world. And it should never happen that the player is stuck because you allowed him to move inside a solid block of stone!

    Now, these statements might seem self contradictable - first I say errors are bound to occur and the next second I say they should never happen. Well, the solution is to be ready for the errors and handle them so the player never notice they are there! There is no shame in errors due to the limited precision of floats - every engine has them, even the Quake engine! In the original Quake John Carmack introduced several ugly hacks to get a player "unstuck" if the engine found out he was.

    There is no single solution so what I'm going to do here is present you my idea of how to avoid getting stuck. The idea is based on the fact that even if we might not be able to check before a move if we are going to get stuck we will know it in the next iteration of the collision routines. When we have calculated the polygon intersection point as described in the section on collision detection we check if that point is inside the sphere. Of course it never should be but my experiences tells me it happens now and then.

    It's easy to check if a point is inside a sphere :
      BOOL CheckPointInSphere(D3DVECTOR point, D3DVECTOR sO, double sR) {
        float d = lengthOfVector(point-sO);
     
       if(d<= sR) return TRUE;
       return FALSE;	
      } 
    
    If we find out that our player (sphere) is stuck we set a certain flag in the structure containing the collision data and then we can handle that in our collision response routine. The way we handle it is to introduce the idea of the last safe position. Each time we receive data from our collision routines we check if we are stuck. If not, we update the last safe position and set it to the position we called the routines with this time. If we are stuck we stop recursing and set the player position back to the last safe position.


    Multiple meshes :

    What if our world is split up into many small collections of triangles ? Well, first of all, this is actually a very good idea! The fewer triangles you check in each call the faster the engine will run. What you do is to keep a structure of collision data which you pass along to each mesh within reach! The structure should contain information about the closest hit so far etc. etc. and must be reset for each new call to the main collision detection function!

    Before checking a mesh you can then first check if it's within reach. You can do this in any way you like - I use a bounding sphere around the mesh and check if my current velocity will bring me inside that sphere. If it will I check the mesh.

    After all meshes has been checked you retrieve the collision data from the structure and react like described in the section on collision response.


    Dynamic geometry :

    So far we have only looked at geometry positioned in world space - i.e. the world matrices for all meshes has been the identity matrix. In a real game you will have a lot of moving geometry. As an example you can check the door or lift in the demo program that comes with this tutorial. The fact that you cannot walk through the door when it is closed seems obvious but when you think about it, it really isn't.
    Of course I don't change all coordinates in the mesh making up the door when it opens - I just change its world matrix. But that means that the data the collision routines receive is the same no matter if the door is open or closed! (think about it!).

    The solution is to move the sphere from world space into object space before you check the triangles in the object. This is done like this :
      D3DMATRIX  invWorldMatrix;
      
      // invert the objects world matrix  
      if (FAILED(D3DMath_MatrixInvert( invWorldMatrix, savedWorld )))
       Log("could not invert world matrix");
     
      D3DVECTOR translatedDestination;
      D3DVECTOR originalSource, originalVelocity;
      
      // Remember - this scales from ellipsoid space to R3
      originalSource = sourcePoint * ellipsoidRadius;
      originalVelocity = velocity * ellipsoidRadius;
          
          
      // translate after animation    
      D3DMath_VectorMatrixMultiply( translatedSourcePoint, originalSource, invWorldMatrix);
      D3DVECTOR originalDestination = originalSource + originalVelocity;
      D3DMath_VectorMatrixMultiply( translatedDestination, originalDestination, invWorldMatrix); 
      translatedVelocity = translatedDestination - translatedSourcePoint;
      
      // convert data back to ellipsoid space
      translatedVelocity /= ellipsoidRadius;
      translatedSourcePoint /= ellipsoidRadius;
    
    Then check all triangles with the translated source and velocity.

    It takes a little more to implement lifts. Not because you don't check the right position of the lift mesh - but because the lift moves! We have to check if it hits the player and if so the player should be pushed upwards with the lift! Lifts going down is automatic because of gravity.

    To handle lifts you should for each object keep a vector called objectVelocity and if the player collides with an object where this vector is not the zero-vector then the player should become sort of a child-object to the lift and the player movement becomes the objectVelocity plus the player velocity plus gravity!

    I will not provide any code for this because such code is quite engine specific and it depends very much on how you handle your objects animations etc. So consider this last bit a source of inspiration for implementing the stuff yourself ;)


    Last remarks :

    Well, that's about all for now.
    Hope you found this doc useful - and btw : If you do make anything public using these techniques please mention me in your greets or where ever you se fit. I do love to see my name in a greeting :=)

    Thanks to Paul Nettle for sharing his ideas about the original algorithm and to Soren Seeberg a.k.a. ZeeGate who went though all the pain of creating those nifty pictures you saw though the document.

    Keep on codin'

    Kasper Fauerby a.k.a. Telemachos
    November, 2000