13,257,276 members (55,196 online)
alternative version

#### Stats

354.9K views
124 bookmarked
Posted 3 Jan 2002

# A C++ implementation of Douglas-Peucker Line Approximation Algorithm

, 3 Mar 2003
 Rate this:
DP Line approximation algorithm is a well-known method to approximate 2D lines. It is quite fast, O(nlog_2(n)) for a n-points line and can drastically compress a data curve. Here, a fully OOP implementation is given.

## Introduction

When working with mathematical simulations or engineering problems, it is not unusual to handle curves that contains thousands of points. Usually, displaying all the points is not useful, a number of them will be rendered on the same pixel since the screen precision is finite. Hence, you use a lot of resource for nothing!

This article presents a fast 2D-line approximation algorithm based on the Douglas-Peucker algorithm (see [1]), well-known in the cartography community. It computes a hull, scaled by a tolerance factor, around the curve by choosing a minimum of key points. This algorithm has several advantages:

• It is fast!: For a n point curve, computation time of the approximation is proportional to nlog_2(n),
• You don't need a priori knowledge on the curve,
• The compression ratio can be easily tuned using the tolerance parameter.

The class has been integrated to a plotting library: Plot Graphic Library.

## Douglas-Peucker (DP) Line Simplification Algorithm

The DP line simplification algorithm was originally proposed in [1]. John Hershberger and Jack Snoeyink have implemented it in C in [2] as a package named `DPSimp`:

`DPsimp` is a Douglas-Peucker line simplification algorithm implementation by John Hershberger and Jack Snoeyink. They analyze the line simplification algorithm reported by Douglas and Peucker and show that its worst case is quadratic in n, the number of input points. The algorithm is based on path hulls, that uses the geometric structure of the problem to attain a worst-case running time proportional to nlog_2(n), which is the best case of the Douglas algorithm.

The algorithm is based on a recursive construction of path hull, as depicted in the picture below. They did all the hard work (writing the base code in C), I ported it to C++ templates.

## Modifications to DPSimp

`DPSimp` was using a recursive call in the `DP` method. This could lead to a stack overflow when the algorithm would go deep in recursion. To avoid this problem an internal stack has been added to the class to mimic the recursive function call without stack overflow.

## Concepts and class design

Let the `points` denote all the points of the original curve and the `keys` the points from the original curve that are kept for the approximation.

The idea is that the user provides a container for the `points`, denoted `PointContainer`, and for the `keys`, denoted `KeyContainer`, and the link between those containers will be the line simplification class, denoted `LineApproximator`.

How do we build a class hierarchy without restricting ourselves to particular container? In fact, one user might store it's `points` in `vector< pair<T,T>>` and another one in 2 separate `vector<T>`. Of course, the same argument applies to the `KeyContainer`.

A possible answer is templating. Passing the `PointContainer` and `KeyContainer` as template arguments for `LineApproximator` allows to build the approximation class without specifying the containers, since the class is built at compilation time (We could write interface for those containers but in fact, I'm too lazy for that :) ).

With this idea in mind, here are the specifications of the container:

### PointContainer

Let

• `Point` a structure, template or class that has `x`,`y` of type `T` as member,
• `PointRef` a structure, template or class that has `x`,`y` of type `T&` as member,

`PointContainer` behaves like a vector of `Point`:

• has `clear()`, `size()` and `resize()` methods,
• has random access iterator,
• `const_iterator` points to a structure similar to `Point`,
• `iterator` points to a structure similar to `PointRef`
• `operator[] const` returns a `Point`,
• `operator[]` returns a `PointRef`

A simple example of valid `PointContainer` is

`vector<Point>`

However, a hybrid container has been developed to handle the case where `x` and `y` are in separate containers (See below).

### KeyContainer

`KeyContainer` behaves like a list of `PointContainer::const_iterator`:

• has `size()`, `clear()` methods,
• support `push_back` method

A simple example of valid `KeyContainer` is

`vector<PointContainer::const_iterator>`

Again, a hydrid container to handle the case where the `keys` must be outputted in separate containers is provided.

## Templates

All the classes are templated to support `float` and `double` version of the algorithm.

The template `TDPHull` is the user interface to the DP algorithm. However, it relies on a series of subclasses detailed below:

NameDescriptionUse
`TLine<T, TPointContainer, TKeyContainer>`2D Line templatePoints and keys
`TLineApproximator<T, TPointContainer, TKeyContainer>`2D Line approximator base classDefault interface to approximation algorithms
`TDPHull<T, TPointContainer, TKeyContainer>`Implementing Douglas-Peukler algorithmUser front end
`TPathHull<T, TPointContainer, TKeyContainer>`Path hullInternal in `TDPHull<T>`
`TPoint<T>`A pair of `T`: x, yTemplate for 2D point `TLineApproximator<T>`
`TPointRef<T>`A pair of `T&`: x, yTemplate for 2D point `TLineApproximator<T>`
`SHomog`2D Homogenous point, `T` x,y,wInternal structure to `TLineApproximator<T>`

## How to use TDPHull?

In the following examples, we adopt the following notations

```using namespace hull; // all classes are in the hull namespace,
using namespace std; // using some STL containers

// defining the point container
typedef vector<TPoint<float>> VectorPointContainer;
// defining the key container
typedef vector<MyPointContainer::const_iterator> ListKeyContainer;
// defining the line approximation class
typedef TDPHull<float, VectorPointContainer, ListKeyContainer> CDPHullF;
CDPHullF dp; // a DPHull object, note that you can also work with doubles```

The approximation methods throw exception, so you should always enclose them in a `try`-`catch` statement.

### Normalization

The data points are, by default, normalized before approximation. This is in order to reduce numerical errors in the gradients computations. This happens when the data is badly scaled: using big numbers close together will lead to disastrous loss of significant numbers.

However, if you feel confident about your data, you can disable it by using `SetNormalization`.

### Handling points and keys

Get a reference to the point container and modify it:

```// getting reference to container,
TDPHull<float>::PointContainer& pc = dp.GetPoints();
for (UINT i=0;i<pc.size();i++)
{
pc[i].x=...;
pc[i].y=...;
}```

If you are using normalization (default behavior), do not forget to re-compute the data bounding box after your changes:

`dp.ComputeBoundingBox();`

### Approximation tuning

You can control the compression ratio by different ways:

• Setting the tolerance
```// Setting tolerance of approximation
dp.SetTol(0.5);```
• Setting a compression ratio, an acceptable compression threshold:
```// dp will find the tolerance corresponding to 10 % of
// the original points, with 5% of possible error.
try
{
dp.ShrinkNorm(0.1,0.05);
}
catch(TCHAR* str)
{
// catch and process errors throw by dp
// example: output to cerr
cerr<<str<<endl;
}```

The method uses dichotomy to accelerate convergence.

• Setting the desired number of points, an acceptable number of points threshold:
```// dp will find the tolerance corresponding to 100
// in the approximated curve, with 5 points of possible error.
try
{
dp.Shrink(100,5);
}
catch(TCHAR* str)
{
// catch and process errors throw by dp
...
}```

### Simplifaction

The easiest part of the job:

```try
{
dp.Simplify();
}
catch(TCHAR* str)
{
// catch and process errors throw by dp
...
}```

or by using `ShrinkNorm, Shrink` methods.

### Accessing the approximated curve

The keys are stored as `PointContainer::const_iterator`. You can access the key container by using `GetKeys`:

```// getting conts reference to keys
const TDPHull<float>::KeyContainer& kc = dp.GetKeys();
TDPHull<float>::KeyContainer::const_iterator it; // iterator for keys
for (it = kc.begin(); it != kc.end(); it++)
{
// it is an const_iterator pointing to a PointContainer::const_iterator
xi=(*it)->x;
yi=(*it)->y;
}```

All you have to do is inherit a class from `TLineApproximator` and override the function `ComputeKeys`.

## Hydrid containers

You can implement your own containers for `points` and `keys` as long as they meet the requirements.

### Separate containers for x,y

It is not unusual to have `x`,`y` stored in separate containers and moreover these containers are not of the same type. To tackle this problem, two wrapper templates have been written: `TPointDoubleContainer` and `TKeyDoubleContainer` which serve as an interface between the approximation algorithms and the containers:

```CVectorX vX;    // the vector of x coordinates
CVectorY vY;    // the vector of y coordinates

// defining the hydrid container
typedef TPointDoubleContainer<float, CVectorX,
CVectorY> HybridPointContainer;

// a list of key x coordinates
CListX lKeyX;
// a vector of key y coordinates, any container
// that has push_back will do the job :)
CVectorY vKeyY;
// defining the hydrid container
typedef TKeyDoubleContainer<float, CListX, CVectorY> HybridKeyContainer;

// creating approximator
TDPHull< float, HybridPointContainer, HydridKeyContainer> dp;
// attaching point containers
dp.GetPoints().SetContainers( &vX, &vY);
// attaching key containers
dp.GetKeys().SetContainers( &lKeyX, &vKeyY);
// dp is now ready to work```

## Using the demo

The demo shows a real time approximation of a curve by different algorithms.

• You can stop/start the animation using the toolbar buttons,
• You can modify the shrinking ration with the scroll bar,
• You can load your own data with the menu, File->Load Data Set. The file must be formatted with a pair `x`,`y` per line.

## Using it in your project

Insert the following files in your project and you're done.

`LineApproximator.h,DPHull.h, PathHull.h`

## Known issues

• The original curve must not self-intersect. This means also that the start and end points must be different, no closed curve !
• Sick dataset and stack overflow: solved. The problem was due to recursion leading to stack overflow. It is solved now.

## Update history

• 04-03-2003
• Got rid of DP recursion by adding an internal function call stack. Hence, the stack overflow problem is solved!!!
• Applied stuff I learned in Effective STL to the classes: using algorithms, functors, etc...
• Changed <CODE lang=c++>class T to <CODE lang=c++>typename T
• Better floating point comparison using `boost::close_at_tolerance`
• 15-11-2002
• More and more templating,
• Detecting when curve is closed
• Hybrid containers
• Fixed bug in compute limits
• Added LOTS of `ASSERT`, so Debug version is significantly slower than release build
• 7-11-2002
• Fixed a bug in the `SLimits` structure (`GetCenterY`)
• `TPathHull` working with iterators rather that `SPoint*`
• Added exception to handle errors
• Fixed a bug in `TDPHull::ComputeKeys`. Was using `pc.end()` rather that `pc.end()--`
• 6-11-2002
• Added base class `TLineApproximator`
• Added proposed algorithm by S.Rog: see `TKeyFramer`, `TGlobalKeyFramer`, `TDispKeyFramer`, `TVectKeyFramer`
• Updated demo
• 3-11-2002
• Added data normalization for better numerical behavior. Avoids the algorithm to crash when using badly conditioned data. Problem submitted by Corey W.
• Templated version
• Got rid of macro and rewrote in more OOP style

## References

1. D. H. Douglas and T. K. Peucker. Algorithms for the reduction of the number of points required to represent a line or its caricature. The Canadian Cartographer, 10(2):112--122, 1973.
2. J. Hershberger and J. Snoeyink. Speeding up the Douglas-Peucker line simplification algorithm. In Proc. 5th Intl. Symp. Spatial Data Handling. IGU Commission on GIS, pages 134--143, 1992. (home page).

A list of licenses authors might use can be found here

## Share

 Engineer United States
Jonathan de Halleux is Civil Engineer in Applied Mathematics. He finished his PhD in 2004 in the rainy country of Belgium. After 2 years in the Common Language Runtime (i.e. .net), he is now working at Microsoft Research on Pex (http://research.microsoft.com/pex).

## You may also be interested in...

 FirstPrev Next
 Re: Cool Corey W13-Mar-03 6:23 Corey W 13-Mar-03 6:23
 Credits Jonathan de Halleux13-Mar-03 7:00 Jonathan de Halleux 13-Mar-03 7:00
 Re: Credits Corey W13-Mar-03 7:05 Corey W 13-Mar-03 7:05
 Update done? Corey W3-Mar-03 7:31 Corey W 3-Mar-03 7:31
 Done Jonathan de Halleux4-Mar-03 0:45 Jonathan de Halleux 4-Mar-03 0:45
 Re: Done Corey W4-Mar-03 5:30 Corey W 4-Mar-03 5:30
 Yep Jonathan de Halleux4-Mar-03 8:59 Jonathan de Halleux 4-Mar-03 8:59
 Does this algorithm scale to 3d? Anonymous7-Nov-02 16:08 Anonymous 7-Nov-02 16:08
 Re: Does this algorithm scale to 3d? Dafydd11-Mar-03 1:08 Dafydd 11-Mar-03 1:08
 I guess so Jonathan de Halleux13-Mar-03 3:20 Jonathan de Halleux 13-Mar-03 3:20
 Not sure about the hull, but the core can Nils Haeck23-Aug-03 23:12 Nils Haeck 23-Aug-03 23:12
 Could this take advantage of block_allocator? Joaquín M López Muñoz7-Nov-02 1:45 Joaquín M López Muñoz 7-Nov-02 1:45
 This is a good question Jonathan de Halleux7-Nov-02 4:01 Jonathan de Halleux 7-Nov-02 4:01
 Re: This is a good question Joaquín M López Muñoz7-Nov-02 4:28 Joaquín M López Muñoz 7-Nov-02 4:28
 PathHull Jonathan de Halleux7-Nov-02 9:28 Jonathan de Halleux 7-Nov-02 9:28
 Re: PathHull Joaquín M López Muñoz7-Nov-02 9:55 Joaquín M López Muñoz 7-Nov-02 9:55
 Peaks are shifted Juan Carlos Cobas5-Nov-02 8:08 Juan Carlos Cobas 5-Nov-02 8:08
 Re: Peaks are shifted Jonathan de Halleux5-Nov-02 23:11 Jonathan de Halleux 5-Nov-02 23:11
 linear regression ? .S.Rod.4-Nov-02 23:15 .S.Rod. 4-Nov-02 23:15
 Check the demo Jonathan de Halleux5-Nov-02 23:12 Jonathan de Halleux 5-Nov-02 23:12
 Nice ender! Anatoly Ivasyuk4-Nov-02 11:51 Anatoly Ivasyuk 4-Nov-02 11:51
 Yep Jonathan de Halleux4-Nov-02 23:32 Jonathan de Halleux 4-Nov-02 23:32
 program stuck in the while loop in DP Haitao Huang10-Jun-02 7:11 Haitao Huang 10-Jun-02 7:11
 Re: program stuck in the while loop in DP Jonathan de Halleux10-Jun-02 22:21 Jonathan de Halleux 10-Jun-02 22:21
 Re: program stuck in the while loop in DP haitaohuang22-Jul-02 10:39 haitaohuang 22-Jul-02 10:39
 Better ask real authors of the algorithms Jonathan de Halleux29-Jul-02 6:32 Jonathan de Halleux 29-Jul-02 6:32
 Re: Better ask real authors of the algorithms Corey W9-Sep-02 13:00 Corey W 9-Sep-02 13:00
 Dataset available? Jonathan de Halleux10-Sep-02 23:21 Jonathan de Halleux 10-Sep-02 23:21
 Re: Dataset available? Corey W23-Oct-02 7:17 Corey W 23-Oct-02 7:17
 Fixed Jonathan de Halleux3-Nov-02 23:46 Jonathan de Halleux 3-Nov-02 23:46
 Try new version Jonathan de Halleux12-Mar-03 8:26 Jonathan de Halleux 12-Mar-03 8:26
 Re: Try new version haitaohuang14-Mar-03 7:17 haitaohuang 14-Mar-03 7:17
 Getting rid of recursion Jonathan de Halleux4-Mar-03 0:48 Jonathan de Halleux 4-Mar-03 0:48
 How can i get the original positions of points in approxiamated curve? qudanqudan15-May-02 15:14 qudanqudan 15-May-02 15:14
 Re: How can i get the original positions of points in approxiamated curve? Jonathan de Halleux27-May-02 5:16 Jonathan de Halleux 27-May-02 5:16
 How does this compare to other line interpolation algorithms? Tim Finer17-Jan-02 7:40 Tim Finer 17-Jan-02 7:40
 Re: How does this compare to other line interpolation algorithms? pelikhan17-Jan-02 8:17 pelikhan 17-Jan-02 8:17
 Re: How does this compare to other line interpolation algorithms? Anonymous14-May-02 4:56 Anonymous 14-May-02 4:56
 Combining it with cubic spline interpolation... Jonathan de Halleux17-Jul-02 4:27 Jonathan de Halleux 17-Jul-02 4:27
 If only... Chris Maunder7-Jan-02 11:38 Chris Maunder 7-Jan-02 11:38
 Re: If only... pelikhan15-Jan-02 1:03 pelikhan 15-Jan-02 1:03
 Re: If only... neojoj17-Jan-02 5:04 neojoj 17-Jan-02 5:04
 Re: If only... pelikhan17-Jan-02 5:08 pelikhan 17-Jan-02 5:08
 Last Visit: 31-Dec-99 19:00     Last Update: 23-Nov-17 2:04 Refresh « Prev12