Implementation of Planar Point Location using slabs - algorithm

I am looking into a Planar Point Location, and right now thinking about make use of The Slabs approach.
I've read through a couple of articles:
Planar Point Location Using Persistent Search Trees
Persistent Data Structures
Point location on Wikipedia
I understand the concept behind a persistent balanced binary search tree, but let's omit them in this question, I don't really care about extra storage space overhead. The thing with the articles is they are discussing how speed can be improved, but don't explain basic stuff. For example:
Could you please correct me if I am wrong:
We draw a line across all intersections.
Now, we have slabs that are split by the segments at different angles.
Every intersection with any line is considered a vertex.
Slabs are sorted by order in a binary search tree (let's omit a partially persistent bst)
Somehow, sectors are sorted in those respective BST's, even though the segments dividing them are almost always at an angle. Does each node has to carry a definition of area?
Please refer to this example image:
Questions:
How would I actually figure out if the point lies in Node c and not in Node b? Is it somehow via area?
Do I instead need arrange my nodes to contain information about the segments? I could then check if the query point lies above a segment (if that's how I should determine my sectors)? If so, would I then search through a Polygon List after, to see which polygon this particular segment belongs to?
Maybe I need to store BST for each line and not a slab?
Would I then have to look at 2 BST's belonging to line on the left, and the 2nd one of the line to the right from the vertex? I could then sort the vertexes by y coordinate in each tree and return the y coordinate of a vertex (the end of a segment) right below my query point. Having done this for left and right line, I would then do a comparison to see if the names of the segment those vertexes come from actually match.
However, this will not give me the right answer, since even if the names do match, I might be below or above the segment (if I am close to it). Also, this implies I have to do 3 binary searches (1 for lines, 1 for the y-coordinate on left line, 1 for the right one), and the books says I only need to do 2 searches (1 for slab, 2nd for sector).
Could someone please point me in the direction to do it?
I probably just missed some essential thought or something.
Edit:
Here is another good article, that explains the solution to the problem, however, I don't quite understand how I would achieve the following:
"Consider any query point q ∈ R2. To find the face of G that contains q, we first use binary search with the x-coordinate of q to find the vertical slab s that contains q. Given s, we use binary search with the y-coordinate of q to find the edges of Es between which q lies. "
How exactly to find those 2 edges? Is it as simple as checking if the point lies below the segment? However, this seems like a complicated check to do (and expensive), as we descend down the tree, inspecting the other nodes.

Well, this was asked a year ago so I guess you're not going to be around to accept the answer. It's a fun question, though, so here goes anyway...
Yes, you can divide the plane into slabs using vertical lines through every intersection. The important result is that all line segments that touch a slab will go all the way through, and the order of those line segments will be the same across the entire slab.
In each slab, you make a binary search tree of those line segments, ordered by their order in the slab. The areas that you call "nodes" are the leaves of the tree, and the line segments are the internal nodes. I will call the leaves "areas" to avoid ambiguity. Since the order of the line segments is the same across the entire slab, the order in this tree is valid for every x coordinate in the slab.
to determine the area that contains any point, you find the slab that contains the point, and then do a simple search in the BST to find the area.
Lets say you're looking for point C, and you're at the node in slab 2 that contains (x2,y2) - (x3,y4). Determine whether C is above or below this line segment, and then take that branch in the tree.
For example, if C is (cx,cy) then the y coordinate of the segment at x=cx is:
testy = ((cx-x2)/(x3-x2)) * (y4-y2) + y2
If cy < testy, then take the upward branch, otherwise take the downward branch. (assuming positive Y goes downward).
When you get to a leaf, that's the area that contains C. You're done.
Now... Making a whole new tree for every slab takes a lot of space, since you can have N^2 intersection points and therefore N^2 slabs. Storing a separate tree in every slab would take O(N^3) space.
Using, say, persistent red-black trees, however, the slabs can actually share a lot of nodes. Using a simple purely functional implementation like Chris Okasaki's red-black tree takes O(log N) space per intersection for a total of O(N^2*log N) space.
I seem to remember that there is also a constant-space-per-change persistent red-black tree that would give you O(N^2), but I don't have a reference.
Note that for most real-life scenarios there are closer to O(N) intersections, because lines don't cross very much, but it's still nice to save the space by using a persistent tree.

Related

Finding Contiguous Areas of Bits in 2D Bit Array

The Problem
I have a bit array which represents a 2-dimensional "map" of "tiles". This image provides a graphical example of the bits in the bit array:
I need to determine how many contiguous "areas" of bits exist in the array. In the example above, there are two such contiguous "areas", as illustrated here:
Tiles must be located directly N, S, E or W of a tile to be considered "contiguous". Diagonally-touching tiles do not count.
What I've Got So Far
Because these bit arrays can become relatively large (several MB in size), I have intentionally avoided using any sort of recursion in my algorithm.
The pseudo-code is as follows:
LET S BE SOURCE DATA ARRAY
LET C BE ARRAY OF IDENTICAL LENGTH TO SOURCE DATA USED TO TRACK "CHECKED" BITS
FOREACH INDEX I IN S
IF C[I] THEN
CONTINUE
ELSE
SET C[I]
IF S[I] THEN
EXTRACT_AREA(S, C, I)
EXTRACT_AREA(S, C, I):
LET T BE TARGET DATA ARRAY FOR STORING BITS OF THE AREA WE'RE EXTRACTING
LET F BE STACK OF TILES TO SEARCH NEXT
PUSH I UNTO F
SET T[I]
WHILE F IS NOT EMPTY
LET X = POP FROM F
IF C[X] THEN
CONTINUE
ELSE
SET C[X]
IF S[X] THEN
PUSH TILE NORTH OF X TO F
PUSH TILE SOUTH OF X TO F
PUSH TILE WEST OF X TO F
PUSH TILE EAST OF X TO F
SET T[X]
RETURN T
What I Don't Like About My Solution
Just to run, it requires two times the memory of the bitmap array it's processing.
While extracting an "area", it uses three times the memory of the bitmap array.
Duplicates exist in the "tiles to check" stack - which seems ugly, but not worth avoiding the way I have things now.
What I'd Like To See
Better memory profile
Faster handling of large areas
Solution / Follow-Up
I re-wrote the solution to explore the edges only (per #hatchet 's suggestion).
This was very simple to implement - and eliminated the need to keep track of "visited tiles" completely.
Based on three simple rules, I can traverse the edges, track min/max x & y values, and complete when I've arrived at the start again.
Here's the demo with the three rules I used:
One approach would be a perimeter walk.
Given a starting point anywhere along the edge of the shape, remember that point.
Start the bounding box as just that point.
Walk the perimeter using a clockwise rule set - if the point used to get to the current point was above, then first look right, then down, then left to find the next point on the shape perimeter. This is kind of like the simple strategy of solving a maze where you continuously follow a wall and always bear to the right.
Each time you visit a new perimeter point, expand the bounding box if the new point is outside it (i.e. keep track of the min and max x and y.
Continue until the starting point is reached.
Cons: if the shape has lots of single pixel 'filaments', you'll be revisiting them as the walk comes back.
Pros: if the shape has large expanses of internal occupied space, you never have to visit them or record them like you would if you were recording visited pixels in a flood fill.
So, conserves space, but in some cases at the expense of time.
Edit
As is so often the case, this problem is known, named, and has multiple algorithmic solutions. The problem you described is called Minimum Bounding Rectangle. One way to solve this is using Contour Tracing. The method I described above is in that class, and is called Moore-Neighbor Tracing or Radial Sweep. The links I've included for them discuss them in detail and point out a problem I hadn't caught. Sometimes you'll revisit the start point before you have traversed the entire perimeter. If your start point is for instance somewhere along a single pixel 'filament', you will revisit it before you're done, and unless you consider this possibility, you'll stop prematurely. The website I linked to talks about ways to address this stopping problem. Other pages at that website also talk about two other algorithms: Square Tracing, and Theo Pavlidis's Algorithm. One thing to note is that these consider diagonals as contiguous, whereas you don't, but that should just something that can be handled with minor modifications to the basic algorithms.
An alternative approach to your problem is Connected-component labeling. For your needs though, this may be a more time expensive solution than you require.
Additional resource:
Moore Neighbor Contour Tracing Algorithm in C++
I actually got a question like this in an interview once.
You can pretend the array is a graph and the connected nodes are the adjacent ones. My algo would involves going 1 to the right until you find a marked node. When you find one do a breadth first search which runs in O(n) and avoids recursion. When the BFS returns keep searching from where you left off and if the node has already been marked by one of the previous BFS's you obviously don't need to search. I wasn't sure if you wanted to actually return the number of objects found, but it's easy to keep track by just incrementing a counter when you hit the first marked square.
Generally when you do a flood fill type algorithm you are placed in a spot and asked to fill. Since this is finding all the filled regions one way you would want to optimize it is to avoid rechecking the already marked nodes from previous BFS's, unfortunately at the moment I cannot think of a way to do that.
One hacky way to reduce memory consumption would be too store a short[][] instead of a boolean. Then use this scheme to avoid making a whole second 2d-array
unmarked = 0, marked = 1, checked and unmarked = 3, checked and marked = 3
This way you can check the status of an entry by its value and avoid making a second array.

How to find the neighbors of a graph effiiciently

I have a program that create graphs as shown below
The algorithm starts at the green color node and traverses the graph. Assume that a node (Linked list type node with 4 references Left, Right, Up and Down) has been added to the graph depicted by the red dot in the image. Inorder to integrate the newly created node with it neighbors I need to find the four objects and link it so the graph connectivity will be preserved.
Following is what I need to clarify
Assume that all yellow colored nodes are null and I do not keep a another data structure to map nodes what is the most efficient way to find the existence of the neighbors of the newly created node. I know the basic graph search algorithms like DFS, BFS etc and shortest path algorithms but I do not think any of these are efficient enough because the graph can have about 10000 nodes and doing graph search algorithms (starting from the green node) to find the neighbors when a new node is added seems computationally expensive to me.
If the graph search is not avoidable what is the best alternative structure. I thought of a large multi-dimensional array. However, this has memory wastage and also has the issue of not having negative indexes. Since the graph in the image can grow in any directions. My solution to this is to write a separate class that consists of a array based data structure to portray negative indexes. However, before taking this option I would like to know if I could still solve the problem without resolving to a new structure and save a lot of rework.
Thank you for any feedback and reading this question.
I'm not sure if I understand you correctly. Do you want to
Check that there is a path from (0,0) to (x1,y1)
or
Check if any of the neighbors of (x1,y1) are in the graph? (even if there is no path from (0,0) to any of this neighbors).
I assume that you are looking for a path (otherwise you won't use a linked-list), which implies that you can't store points which have no path to (0,0).
Also, you mentioned that you don't want to use any other data structure beside / instead of your 2D linked-list.
You can't avoid full graph search. BFS and DFS are the classic algorithms. I don't think that you care about the shortest path - any path would do.
Another approaches you may consider is A* (simple explanation here) or one of its variants (look here).
An alternative data structure would be a set of nodes (each node is a pair < x,y > of course). You can easily run 4 checks to see if any of its neighbors are already in the set. It would take O(n) space and O(logn) time for both check and add. If your programming language does not support pairs as nodes of a set, you can use a single integer (x*(Ymax+1) + Y) instead.
Your data structure can be made to work, but probably not efficiently. And it will be a lot of work.
With your current data structure you can use an A* search (see https://en.wikipedia.org/wiki/A*_search_algorithm for a basic description) to find a path to the point, which necessarily finds a neighbor. Then pretend that you've got a little guy at that point, put his right hand on the wall, then have him find his way clockwise around the point. When he gets back, he'll have found the rest.
What do I mean by find his way clockwise? For example suppose that you go Down from the neighbor to get to his point. Then your guy should be faced the first of Right, Up, and Left which he has a neighbor. If he can go Right, he will, then he will try the directions Down, Right, Up, and Left. (Just imagine trying to walk through the maze yourself with your right hand on the wall.)
This way lies insanity.
Here are two alternative data structures that are much easier to work with.
You can use a quadtree. See http://en.wikipedia.org/wiki/Quadtree for a description. With this inserting a node is logarithmic in time. Finding neighbors is also logarithmic. And you're only using space for the data you have, so even if your graph is very spread out this is memory efficient.
Alternately you can create a class for a type of array that takes both positive and negative indices. Then one that builds on that to be 2-d class that takes both positive and negative indices. Under the hood that class would be implemented as a regular array and an offset. So an array that can start at some number, positive or negative. If ever you try to insert a piece of data that is before the offset, you create a new offset that is below that piece by a fixed fraction of the length of the array, create a new array, and copy data from the old to the new. Now insert/finding neighbors are usually O(1) but it can be very wasteful of memory.
You can use a spatial index like a quad tree or a r-tree.

Algorithm for falling grid items

I do not know how to describe the goal succinctly, which may be why I haven't been able to find an applicable algorithm despite ample searching, but a picture shows it clearly:
Given the state of items in the grid at the left, does anyone know of an algorithm for efficiently finding the ending positions shown in the grid at right? In this case all the items have "fallen" "down", but the direction of course is arbitrary. The point is just that:
There are a collection of items of arbitrary shapes, but all composed of contiguous squares
Items cannot overlap
All items should move the maximum distance in a given direction until they are touching a wall, or they are touching another item which [...is touching another item ad infinitum...] is touching a wall.
This is not homework, I'm not a student. This is for my own interest in geometry and programming. I haven't mentioned the language because it doesn't matter. I can implement whatever algorithm in the language I'm using for the specific project I'm working on. A useful answer could be described in words or code; it's the ideas that matter.
This problem could probably be abstracted into some kind of graph (in the mathematical sense) of dependencies and slack space, so perhaps an algorithm aimed at minimizing lag time could be adapted.
If you don't know the answer but are about to try to make up an algorithm on the spot, just remember that there can be circular dependencies, such as with the interlocking pink (backwards) "C" and blue "T" shapes. Parts of T are below C, and parts of C are below T. This would be even more tricky if interlocking items were locked through a "loop" of several pieces.
Some notes for an applicable algorithm: All the following are very easy and fast to do because of the way I've built the grid object management framework:
Enumerate the individual squares within a piece
Enumerate all pieces
Find the piece, if any, occupying a specific square in the overall grid
A note on the answer:
maniek hinted it first, but bloops has provided a brilliant explanation. I think the absolute key is the insight that all pieces moving the same amount maintain their relationship to each other, and therefore those relationships don't have to be considered.
An additional speed-up for a sparsely populated board would be to shift all pieces to eliminate rows that are completely empty. It is very easy to count empty rows and to identify pieces on one side ("above") an empty row.
Last note: I did in fact implement the algorithm described by bloops, with a few implementation-specific modifications. It works beautifully.
The Idea
Define the set of frozen objects inductively as follows:
An object touching the bottom is frozen.
An object lying on a frozen object is frozen.
Intuitively, exactly the frozen objects have reached their final place. Call the non-frozen objects active.
Claim: All active objects can fall one unit downwards simultaneously.
Proof: Of course, an active object will not hit another active object, since their relative position with respect to each other does not change. An active object will also not hit a frozen object. If that was so, then the active object was, in fact, frozen, because it was lying on a a frozen object. This contradicts our assumption.
Our algorithm's very high-level pseudo-code would be as follows:
while (there are active objects):
move active objects downwards simultaneously until one of them hits a frozen object
update the status (active/frozen) of each object
Notice that at least one object becomes frozen in each iteration of the while loop. Also, every object becomes frozen exactly once. These observations would be used while analyzing the run-time complexity of the actual algorithm.
The Algorithm
We use the concept of time to improve the efficiency of most operations. The time is measured starting from 0, and every unit movement of the active objects take 1 unit time. Observe that, when we are at time t, the displacement of all the objects currently active at time t, is exactly t units downward.
Note that in each column, the relative ordering of each cell is fixed. One of the implications of this is that each cell can directly stop at most one other cell from falling. This observation could be used to efficiently predict the time of the next collision. We can also get away with 'processing' every cell at most once.
We index the columns starting from 1 and increasing from left to right; and the rows with height starting from 1. For ease of implementation, introduce a new object called bottom - which is the only object which is initially frozen and consists of all the cells at height 0.
Data Structures
For an efficient implementation, we maintain the following data structures:
An associative array A containing the final displacement of each cell. If a cell is active, its entry should be, say, -1.
For each column k, we maintain the set S_k of the initial row numbers of active cells in column k. We need to be able to support successor queries and deletions on this set. We could use a Van Emde Boas tree, and answer every query in O(log log H); where H is the height of the grid. Or, we could use a balanced binary search tree which can perform these operations in O(log N); where N is the number of cells in column k.
A priority queue Q which will store the the active cells with its key as the expected time of its future collision. Again, we could go for either a vEB tree for a query time of O(log log H) or a priority queue with O(log N) time per operation.
Implementation
The detailed pseudo-code of the algorithm follows:
Populate the S_k's with active cells
Initialize Q to be an empty priority queue
For every cell b in bottom:
Push Q[b] = 0
while (Q is not empty):
(x,t) = Q.extract_min() // the active cell x collides at time t
Object O = parent_object(x)
For every cell y in O:
A[y] = t // freeze cell y at displacement t
k = column(y)
S_k.delete(y)
a = S_k.successor(y) // find the only active cell that can collide with y
if(a != nil):
// find the expected time of the collision between a and y
// note that both their positions are currently t + (their original height)
coll_t = t + height(a) - height(y) - 1
Push/update Q[a] = coll_t
The final position of any object can be obtained by querying A for the displacement of any cell belonging to that object.
Running Time
We process and freeze every cell exactly once. We perform a constant number of lookups while freezing every cell. We assume parent_object lookup can be done in constant time. The complexity of the whole algorithm is O(N log N) or O(N log log H) depending on the data structures we use. Here, N is the total number of cells across all objects.
And now something completely different :)
Each piece that rests on the ground is fixed. Each piece that rests on a fixed piece is fixed. The rest can move. Move the unfixed pieces 1 square down, repeat until nothing can move.
Okay so this appears to be as follows -
we proceed in step
in each step we build a directed graph whose vertices are the object set and whose edges are as follows =>
1) if x and y are two objects then add an edge x->y if x cannot move until y moves. Note that we can have both x->y and y->x edges.
2) further there are objects which can no longer move since they are at the bottom so we colour their vertices as blue. The remaining vertices are red.
2) in the directed graph we find all the strongly connected components using Kosaraju's algorithm/Tarjan's algorithm etc. (In case you are not familiar with SCC then they are extremely powerful technique and you can refer to Kosaraju's algorithm.) Once we get the SCCs we shrink them to a small vertex i.e. we replace the SCC by a single vertex while preserving all external(to SCC) edges. Now if any of the vertex in SCC is blue then we color the new vertex as blue else it is red. This implies that if one object cannot move in SCC then none can.
3) the graph you get will be a directed acyclic graph and you can do a topological sort. Traverse the vertex in increasing order of top numbering and as long as you see a red colour vertex and move the objects represented by the vertex.
continue this step until you can no longer move any vertex in step 3.
If two objects A and B overlap then we say that they are inconsistent relative to each other. For proof of correctness argue the following lemmas
1) "if I move an SCC then none of the object in it cause inconsistency among themselves."
2) "when I move an object in step 3 then I do not cause inconsistencies"
The challenge for you now will be to formally prove the correctness and find suitable data structures to solve it in efficient. Let me know if you need any help.
I haven't cleared all the details, but I think the following seems like a somewhat systematic approach:
Looking at the whole picture as a graph, what you need is a topologic sort of all the vertices - i.e. the items. Item A should be before item B in the sorting if any part of A is below any part of B. Then, when you have the items sorted topologically, you can just iterate through them in this order and determine the positions - as all the items below the current one already have fixed positions.
In order to be able to do the topologic sort, you need an acyclic graph. You can then use some of the algorithms for Strongly Connected Components to find all the cycles and compress them into single vertices. Then you can perform a topsort on the resulting graph.
To find the positions of the pieces within a SCC: first consider it as one big piece and determine where it will end up. This will determine some fixed pieces that cannot move anymore. Remove them and repeat the same procedure for the rest of the pieces in this SCC (if any) to find out their final positions.
The third part is the only one that seems computationally intensive - if you have a very complicated structure for the pieces, but it should still be more optimal than trying to move the pieces one square of the grid at a time.
EDITED several times. I think this is all you need to do:
Find all the pieces that can only fall mutually dependent on each other and combine them into an equivalent larger piece (e.g. the T and backwards C in your picture.)
Iterate through all the pieces, moving them the maximum direction down before they hit something. Repeat until nothing moves.

Nearest neighbour search in a constantly changing set of line segments

I have a set of line segments. I want to perform the following operations on them:
Insert a new line segment.
Find all line segments within radius R of a given point.
Find all points within radium R1 of a given point.
Given a line segment, find if it intersects any of the existing line segments. Exact intersection point is not necessary(though that probably doesnt simplify anything.)
The problem is algorithms like kd/bd tree or BSP trees is that they assume a static set of points, and in my case the points will constantly get augmented with new points, necessitating rebuilding the tree.
What data-structure/algorithms will be most suited for this situation ?
Edit: Accepted the answer that is the simplest and solves my problem. Thanks everyone!
Maintaining a dynamic tree might not be as bad as you think.
As you insert new points/lines etc into the collection it's clear that you'd need to refine the current tree, but I can't see why you'd have to re-build the whole tree from scratch every time a new entity was added, as you're suggesting.
With a dynamic tree approach you'd have a valid tree at all times, so you can then just use the fast range searches supported by your tree type to find the geometric entities you've mentioned.
For your particular problem:
You could setup a dynamic geometric tree, where the leaf elements
maintain a list of geometric entities (points and lines) associated
with that leaf.
When a line is inserted into the collection it should be pushed onto
the lists of all leaf elements that it intersects with. You can do
this efficiently by traversing the subset of the tree from the root
that intersects with the line.
To find all points/lines within a specified radial halo you first need
to find all leaves in this region. Again, this can be done by traversing
the subset of the tree from the root that is enclosed by, or that
intersects with the halo. Since there maybe some overlap, you need to
check that the entities associated with this set of leaf elements
actually lie within the halo.
Once you've inserted a line into a set of leaf elements, you can find
whether it intersects with another line by scanning all of the lines
associated with the subset of leaf boxes you've just found. You can then
do line-line intersection tests on this subset.
A potential dynamic tree refinement algorithm, based on an upper limit to the number of entities associated with each leaf in the tree, might work along these lines:
function insert(x, y)
find the tree leaf element enclosing the new entitiy at (x,y) based on whatever
fast search algorithm your tree supports
if (number of entities per leaf > max allowable) do
refine current leaf element (would typically either be a bisection
or quadrisection based on a 2D tree type)
push all entities from the old leaf element list onto the new child element
lists, based on the enclosing child element
else
push new entity onto list for leaf element
endif
This type of refinement strategy only makes local changes to the tree and is thus generally pretty fast in practice. If you're also deleting entities from the collection you can also support dynamic aggregation by imposing a minimum number of entities per leaf, and collapsing leaf elements to their parents when necessary.
I've used this type of approach a number of times with quadtrees/octrees, and I can't at this stage see why a similar approach wouldn't work with kd-trees etc.
Hope this helps.
One possibility is dividing your space into a grid of boxes - perhaps 10 in the y-axis and 10 in the x-axis for a total of 100.
Store these boxes in an array, so it's very easy/fast to determine neighboring boxes. Each box will hold a list vector of line segments that live in that box.
When you calculate line segments within R of one segment, you can check only the relevant neighboring boxes.
Of course, you can create multiple maps of differing granularities, maybe 100 by 100 smaller boxes. Simply consider space vs time and maintenance trade-offs in your design.
updating segment positions is cheap: just integer-divide by box sizes in the x and y directions. For example, if box-size is 20 in both directions and your new coordinate is 145,30. 145/20==7 and 30/20==1, so it goes into box(7,1), for a 0-based system.
While items 2 & 3 are relatively easy, using a simple linear search with distance checks as each line is inserted, item 4 is a bit more involved.
I'd tend to use a constrained triangulation to solve this, where all the input lines are treated as constraints, and the triangulation is balanced using a nearest neighbour rather than Delaunay criterion. This is covered pretty well in Triangulations and applications by Øyvind Hjelle, Morten Dæhlen and Joseph O'Rourkes Computational Geometry in C Both have source available, including getting sets of all intersections.
The approach I've taken to do this dynamically in the past is as follows;
Create a arbitrary triangulation (TIN) comprising of two triangles
surrounding the extents (current + future) of your data.
For each new line
Insert both points into the TIN. This can be done very quickly by
traversing triangles to find the insertion point, and replacing the
triangle found with three new triangles based on the new point and
old triangle.
Cut a section through the TIN based on the two end points, keeping a
list of points where the section cuts any previously inserted lined.
Add the intersection point details to a list stored against both
lines, and insert them into the TIN.
Force the inserted line as a constraint
Balance all pairs of adjacent triangles modified in the above process
using a nearest neighbour criterion, and repeat until all triangles
have been balanced.
This works better than a grid based method for poorly distributed data, but is more difficult to implement. Grouping end-point and lines into overlapping grids will probably be a good optimization for 2 & 3.
Note that I think using the term 'nearest neighbour' in your question is misleading, as this is not the same as 'all points within a given distance of a line', or 'all points within a given radius of another point'. Nearest neighbour typically implies a single result, and does not equate to 'within a given point to point or point to line distance'.
Instead of inserting and deleting into a tree you can calculate a curve that completely fills the plane. Such a curve reduce the 2d complexity to a 1d complexity and you would be able to find the nearest neighbor. You can find some example like z curve and hilbert curve. Here is a better description of my problem http://en.wikipedia.org/wiki/Closest_pair_of_points_problem.

How to implement nearest neighbor search using KDTrees?

So, I'm implementing a KD-Tree to do a nearest neighbor search. I've got the building the tree part working, but I don't think I understand the search part completely.
About traversing the tree to search for the neighbor, the Wikipedia article says the following:
Starting with the root node, the algorithm moves down the tree recursively, in the same
way that it would if the search point were being inserted (i.e. it goes right or left
depending on whether the point is greater or less than the current node in the split
dimension).
What does "greater or less than the current node in the spit dimension mean? Do we compare the points based on distance from the query or do we compare the points by the split dimension?
Also, could someone explain the part about hyperspace and hyperplane? I feel like I understand it, but since I'm not sure I'd like some more explanation.
Thanks!
Each node splits the space into 2 half-spaces along one axis. You look at where the point in question is relative to that split plane to decide which side of the tree to go down. For example if your point is (4,7,12) and you have a splitting plane that cuts the y axis at 9, you'd compare the 7 to the 9 and decide to go down the left (less than) side of the k-d tree first. After finding the nearest neighbor on the left side, you then check if it is closer than 2 (the distance to the split plane: 9-7). If it is closer than the split-plane you do not need to traverse the other half-tree at all. This is why it works so well, most of the time you will only have to traverse one sub-tree.
Hope that helps.

Resources