Space partitioning for flocking

Recommended Posts

Hi,
I'm looking for space partitioning structure which will be good for moving points and will enable fast fixed radius near neighbors search.

Share on other sites
Well, you could use binary space partitions or quadtree's, which are probably easier for 2d things, or even octrees which are easier to use with 3d things.
Or, you can just use a grid, when each cell in the grid has a pointer to the objects in it, the objects only have to check for collisions with objects that are in the same cell.

Share on other sites
A simple grid is what I would try first, they're the easiest thing to implement and often very effective (especially for uniform size/fixed radius).

See Ericson's Real-Time Collision Detection for a terrific chapter on various ways to implement and use grids.

Share on other sites
I'd go for grid. Simple, plus no re-partitioning as the points move.

Share on other sites
I second the grid, since presumably your objects are all the same size and you can tune the size of the grid cells.

Share on other sites
But with grid I would need to search 9 cells right? And what if I would need different radius in future? Anyway I'll test the grid thing. I should also mention that this its in 2d and I'm just using points if that helps in some case.

Share on other sites
9 is pretty arbitrary. Whatever size grid works well. The key is that the grid is only for reducing the candidates for comparison, not doing the actual comparison. Say you have a flock of 64 objects. If you use a grid of size 8 by 8, on average there will be 1 object per cell (obviously for flocking some cells would have more). Based on this average you would need to do distance checks for the objects in a cell with the contents of that cell [i]and the directly neighbouring cells too[/i]. So around 8 distance comparisons per cell. Which may seem like a lot, but still miles better than comparing every object to every other object.

Excuse my late-night mathematics skills if there are errors. ;)

Share on other sites
Alternatively, skip the partitioning and instead only do a fixed number of distance checks at all. There's research that real flockers (i.e. birds, fish) do that same thing too. Not sursprising if you think about it, a fish brain isn't so terribly huge.

Looking at half a dozen randomly selected (not necessarily nearest) neighbours pretty accurately simulates real flocking.

Share on other sites
I don't think that 3*3 = 9 is an arbitrary number any more than 3*3*3 = 27 is... As long as the sphere's radius is less than or equal to the cell size.

Share on other sites
[quote name='ardmax1' timestamp='1341488866' post='4955932']
But with grid I would need to search 9 cells right? And what if I would need different radius in future? Anyway I'll test the grid thing. I should also mention that this its in 2d and I'm just using points if that helps in some case.
[/quote]

No, you only need to check for collision in the 4 neighbouring cells with a higher number. That is the neighbour to the right and the three neighbours below (assuming cell 0 is top left corner and last cell is bottom right corner).

Share on other sites
[quote name='taby' timestamp='1341518377' post='4956073']
I don't think that 3*3 = 9 is an arbitrary number any more than 3*3*3 = 27 is... As long as the sphere's radius is less than or equal to the cell size.
[/quote]

As stated, late at night. ;) I thought he meant a 3 x 3 grid (which is pretty coarse and not very helpful), but he probably meant the same thing that I said.

Share on other sites
[quote name='h4tt3n' timestamp='1341521355' post='4956094']
No, you only need to check for collision in the 4 neighbouring cells with a higher number. That is the neighbour to the right and the three neighbours below (assuming cell 0 is top left corner and last cell is bottom right corner).
[/quote]
If you need to find objects whose center is within a radius R around the center of each object, the best grid size is a 2R by 2R square: every disc straddles up to 4 cells, so if you assign objects to the cell containing the upper left corner of the 2R by 2R AABB of their region of interest you only need to check the objects in the same cell and the three adjacent cells to the right, to the bottom and diagonally to the bottom right.

Share on other sites
First try with grid i got 4 times more (1k to 4k) birds without fps drop, but I'm sure it can be faster. Any idea how to optimize it?
[source]Grid::Grid( Flock* flock, float cellsize ) {
min = flock->min - flock->max / 2;
max = flock->max * 1.5f;

this->cellsize = cellsize;
ccx = (int)((max.x - min.x) / cellsize) + 1;
ccy = (int)((max.y - min.y) / cellsize) + 1;

cells = vector< vector< vector< Bird* > > >(ccy);
for( int i = 0; i < ccy; i++ ){
cells[i] = vector< vector< Bird* > >(ccx);
for( int j = 0; j < ccx; j++ ){
cells[i][j] = vector< Bird* >();
}
}

for( auto b : flock->birds ){
int cx = (int)((b->pos.x - min.x) / cellsize);
int cy = (int)((b->pos.y - min.y) / cellsize);
b->cx = cx; b->cy = cy;
cells[cy][cx].push_back( b );
}
}

Grid::~Grid() {}

vector< Bird* > Grid::getNeighbors( float x, float y, float r ){
int cx = (x - min.x) / cellsize;
int cy = (y - min.y) / cellsize;

vector< Bird* > ret;
int m = ceil( r / cellsize );
for( int i = cy-m; i <= cy+m; i++ ){
for( int j = cx-m; j <= cx+m; j++ ){
if( j < 0 || i < 0 || j >= ccx || i >= ccy )
continue;
ret.insert( ret.end(), cells[i][j].begin(), cells[i][j].end());
}
}

return ret;
}

void Grid::update( Bird* bird ){
int cx = (bird->pos.x - min.x) / cellsize;
int cy = (bird->pos.y - min.y) / cellsize;

if( bird->cx != cx || bird->cy != cy ){
auto cell = &cells[bird->cy][bird->cx];
cell->erase( remove( cell->begin(), cell->end(), bird ) );
cells[cy][cx].push_back( bird );

bird->cx = cx; bird->cy = cy;
}
}
[/source]

Share on other sites
[quote name='LorenzoGatti' timestamp='1341593992' post='4956375']
If you need to find objects whose center is within a radius R around the center of each object, the best grid size is a 2R by 2R square: every disc straddles up to 4 cells, so if you assign objects to the cell containing the upper left corner of the 2R by 2R AABB of their region of interest you only need to check the objects in the same cell and the three adjacent cells to the right, to the bottom and diagonally to the bottom right.
[/quote]

Neat trick, will implement this in my sph simulations.

Edit: Are you sure about this? After doing some pen-and-paper experiments it was easy to construct a scenario where a particle collision is not detected. If sphere A's top left AABB corner is in cell(x, y) and Sphere B's AABB top left corner is in cell(x-1, y+1) then a collision can never be detected.

cheers,
Mike Edited by h4tt3n

Share on other sites
[quote name='h4tt3n' timestamp='1341599285' post='4956419']
[quote name='LorenzoGatti' timestamp='1341593992' post='4956375']
If you need to find objects whose center is within a radius R around the center of each object, the best grid size is a 2R by 2R square: every disc straddles up to 4 cells, so if you assign objects to the cell containing the upper left corner of the 2R by 2R AABB of their region of interest you only need to check the objects in the same cell and the three adjacent cells to the right, to the bottom and diagonally to the bottom right.
[/quote]

Neat trick, will implement this in my sph simulations.

Edit: Are you sure about this? After doing some pen-and-paper experiments it was easy to construct a scenario where a particle collision is not detected. If sphere A's top left AABB corner is in cell(x, y) and Sphere B's AABB top left corner is in cell(x-1, y+1) then a collision can never be detected.

cheers,
Mike
[/quote]

I *really* recommend Ericson's "Real-Time Collision Detection", seriously: the chapter on grids covers several different aspects of implementation, including different ways to approach querying and storing objects in cells (including the 4- vs 9-way search), different ways to approach updating the cell occupancy, etc., it's great.

Seriously... I thought I knew something about different ways to use grids, then I read the chapter on grids, and I realized how little I knew [img]http://public.gamedev.net//public/style_emoticons/default/smile.png[/img] Edited by raigan

Share on other sites
Okay, I'll see if I can pick up a copy somewhere.

Create an account

Register a new account

• Forum Statistics

• Total Topics
628378
• Total Posts
2982334

• 10
• 9
• 15
• 24
• 11