OpenCV  5.0.0-pre
Open Source Computer Vision
Loading...
Searching...
No Matches
Object detection with Generalized Ballard and Guil Hough Transform

Prev Tutorial: Hough Circle Transform
Next Tutorial: Remapping

Original author Markus Heck
Compatibility OpenCV >= 3.4

Goal

In this tutorial you will learn how to:

Example

What does this program do?

  1. Load the image and template

  1. Instantiate cv::GeneralizedHoughBallard with the help of createGeneralizedHoughBallard()
  2. Instantiate cv::GeneralizedHoughGuil with the help of createGeneralizedHoughGuil()
  3. Set the required parameters for both GeneralizedHough variants
  4. Detect and show found results
Note
  • Both variants can't be instantiated directly. Using the create methods is required.
  • Guil Hough is very slow. Calculating the results for the "mini" files used in this tutorial takes only a few seconds. With image and template in a higher resolution, as shown below, my notebook requires about 5 minutes to calculate a result.

Code

The complete code for this tutorial is shown below.

using namespace cv;
using namespace std;
int main() {
// load source image and grayscale template
samples::addSamplesDataSearchSubDirectory("doc/tutorials/imgproc/generalized_hough_ballard_guil");
Mat image = imread(samples::findFile("images/generalized_hough_mini_image.jpg"));
Mat templ = imread(samples::findFile("images/generalized_hough_mini_template.jpg"), IMREAD_GRAYSCALE);
// create grayscale image
Mat grayImage;
cvtColor(image, grayImage, COLOR_RGB2GRAY);
// create variable for location, scale and rotation of detected templates
vector<Vec4f> positionBallard, positionGuil;
// template width and height
int w = templ.cols;
int h = templ.rows;
// create ballard and set options
Ptr<GeneralizedHoughBallard> ballard = createGeneralizedHoughBallard();
ballard->setMinDist(10);
ballard->setLevels(360);
ballard->setDp(2);
ballard->setMaxBufferSize(1000);
ballard->setVotesThreshold(40);
ballard->setCannyLowThresh(30);
ballard->setCannyHighThresh(110);
ballard->setTemplate(templ);
// create guil and set options
Ptr<GeneralizedHoughGuil> guil = createGeneralizedHoughGuil();
guil->setMinDist(10);
guil->setLevels(360);
guil->setDp(3);
guil->setMaxBufferSize(1000);
guil->setMinAngle(0);
guil->setMaxAngle(360);
guil->setAngleStep(1);
guil->setAngleThresh(1500);
guil->setMinScale(0.5);
guil->setMaxScale(2.0);
guil->setScaleStep(0.05);
guil->setScaleThresh(50);
guil->setPosThresh(10);
guil->setCannyLowThresh(30);
guil->setCannyHighThresh(110);
guil->setTemplate(templ);
// execute ballard detection
ballard->detect(grayImage, positionBallard);
// execute guil detection
guil->detect(grayImage, positionGuil);
// draw ballard
for (vector<Vec4f>::iterator iter = positionBallard.begin(); iter != positionBallard.end(); ++iter) {
RotatedRect rRect = RotatedRect(Point2f((*iter)[0], (*iter)[1]),
Size2f(w * (*iter)[2], h * (*iter)[2]),
(*iter)[3]);
Point2f vertices[4];
rRect.points(vertices);
for (int i = 0; i < 4; i++)
line(image, vertices[i], vertices[(i + 1) % 4], Scalar(255, 0, 0), 6);
}
// draw guil
for (vector<Vec4f>::iterator iter = positionGuil.begin(); iter != positionGuil.end(); ++iter) {
RotatedRect rRect = RotatedRect(Point2f((*iter)[0], (*iter)[1]),
Size2f(w * (*iter)[2], h * (*iter)[2]),
(*iter)[3]);
Point2f vertices[4];
rRect.points(vertices);
for (int i = 0; i < 4; i++)
line(image, vertices[i], vertices[(i + 1) % 4], Scalar(0, 255, 0), 2);
}
imshow("result_img", image);
waitKey();
return EXIT_SUCCESS;
}
n-dimensional dense array class
Definition mat.hpp:950
int cols
Definition mat.hpp:2424
int rows
the number of rows and columns or (-1, -1) when the matrix has more than 2 dimensions
Definition mat.hpp:2424
The class represents rotated (i.e. not up-right) rectangles on a plane.
Definition types.hpp:541
void points(Point2f pts[]) const
std::shared_ptr< _Tp > Ptr
Definition cvstd_wrapper.hpp:23
int main(int argc, char *argv[])
Definition highgui_qt.cpp:3
Definition core.hpp:107
STL namespace.

Explanation

Load image, template and setup variables

// load source image and grayscale template
samples::addSamplesDataSearchSubDirectory("doc/tutorials/imgproc/generalized_hough_ballard_guil");
Mat image = imread(samples::findFile("images/generalized_hough_mini_image.jpg"));
Mat templ = imread(samples::findFile("images/generalized_hough_mini_template.jpg"), IMREAD_GRAYSCALE);
// create grayscale image
Mat grayImage;
cvtColor(image, grayImage, COLOR_RGB2GRAY);
// create variable for location, scale and rotation of detected templates
vector<Vec4f> positionBallard, positionGuil;
// template width and height
int w = templ.cols;
int h = templ.rows;

The position vectors will contain the matches the detectors will find. Every entry contains four floating point values: position vector

  • [0]: x coordinate of center point
  • [1]: y coordinate of center point
  • [2]: scale of detected object compared to template
  • [3]: rotation of detected object in degree in relation to template

An example could look as follows: [200, 100, 0.9, 120]

Setup parameters

// create ballard and set options
Ptr<GeneralizedHoughBallard> ballard = createGeneralizedHoughBallard();
ballard->setMinDist(10);
ballard->setLevels(360);
ballard->setDp(2);
ballard->setMaxBufferSize(1000);
ballard->setVotesThreshold(40);
ballard->setCannyLowThresh(30);
ballard->setCannyHighThresh(110);
ballard->setTemplate(templ);
// create guil and set options
Ptr<GeneralizedHoughGuil> guil = createGeneralizedHoughGuil();
guil->setMinDist(10);
guil->setLevels(360);
guil->setDp(3);
guil->setMaxBufferSize(1000);
guil->setMinAngle(0);
guil->setMaxAngle(360);
guil->setAngleStep(1);
guil->setAngleThresh(1500);
guil->setMinScale(0.5);
guil->setMaxScale(2.0);
guil->setScaleStep(0.05);
guil->setScaleThresh(50);
guil->setPosThresh(10);
guil->setCannyLowThresh(30);
guil->setCannyHighThresh(110);
guil->setTemplate(templ);

Finding the optimal values can end up in trial and error and depends on many factors, such as the image resolution.

Run detection

// execute ballard detection
ballard->detect(grayImage, positionBallard);
// execute guil detection
guil->detect(grayImage, positionGuil);

As mentioned above, this step will take some time, especially with larger images and when using Guil.

Draw results and show image

// draw ballard
for (vector<Vec4f>::iterator iter = positionBallard.begin(); iter != positionBallard.end(); ++iter) {
RotatedRect rRect = RotatedRect(Point2f((*iter)[0], (*iter)[1]),
Size2f(w * (*iter)[2], h * (*iter)[2]),
(*iter)[3]);
Point2f vertices[4];
rRect.points(vertices);
for (int i = 0; i < 4; i++)
line(image, vertices[i], vertices[(i + 1) % 4], Scalar(255, 0, 0), 6);
}
// draw guil
for (vector<Vec4f>::iterator iter = positionGuil.begin(); iter != positionGuil.end(); ++iter) {
RotatedRect rRect = RotatedRect(Point2f((*iter)[0], (*iter)[1]),
Size2f(w * (*iter)[2], h * (*iter)[2]),
(*iter)[3]);
Point2f vertices[4];
rRect.points(vertices);
for (int i = 0; i < 4; i++)
line(image, vertices[i], vertices[(i + 1) % 4], Scalar(0, 255, 0), 2);
}
imshow("result_img", image);
waitKey();

Result

result image

The blue rectangle shows the result of cv::GeneralizedHoughBallard and the green rectangles the results of cv::GeneralizedHoughGuil.

Getting perfect results like in this example is unlikely if the parameters are not perfectly adapted to the sample. An example with less perfect parameters is shown below. For the Ballard variant, only the center of the result is marked as a black dot on this image. The rectangle would be the same as on the previous image.

less perfect result