Detection of position/angle of landing platform to steer a drone - opencv

Question

Bacisally the task is:
Implement automatic landing/takeoff of a quadcopter, which can carry one FlyCam/GoPro camera. Orientation should happen relative to altitude and position of the landing platform on 2D plane including rotation. That means the drone has a "head" and "tail" and should land in a specific position.
The landing platform looks like this
The corner shapes are for orientation on big distances, and small repetitive shapes in the center circle are for exact landing.
What approach would you take to solve this task?

Answer 1

Here is a pseudo code, assuming you already have full access to the motor control API; i.e. you have successfully defined what is needed for changing altitude, rotating left etc.
loop
{
if(landing board detected)
{
if(circle including the center point detected)
{
find orientation from corner circles' center
change device's orientation accordingly
}
else
{
lose altitude & move towards the center point
}
}
else
{
move around
}
}
Landing board & its center:
Assumption: It is the biggest & nearly perfect square.
1- Threshold
2- Extract contours
3- Apply shape (square) filter to contours
4- Find the biggest contour
5- Find its center
6- Crop the image with the bounding rect of this contour
Mat image = imread("~\\image.jpg");
// scale down for faster processing
pyrDown(image, image);
pyrDown(image, image);
// safe copy
Mat temp = image.clone();
// noise reduction & thresholding
GaussianBlur(image, image, Size(5,5), 3);
cvtColor(image, image, CV_BGR2GRAY);
threshold(image, image, 127, 255, CV_THRESH_OTSU);
// extract all contours
vector<vector<Point> > contours;
findContours(image, contours, CV_RETR_LIST, CV_CHAIN_APPROX_NONE);
// define a perfect square
vector<Point> square;
square.push_back(Point(0,0));
square.push_back(Point(0,10));
square.push_back(Point(10,10));
square.push_back(Point(10,0));
// filter out contours that are not square
bool erased;
for(unsigned int i = 0; i<contours.size(); i++)
{
erased = false;
double x = matchShapes(contours[i], square, CV_CONTOURS_MATCH_I2, 0);
if(x > 0.005)
{
contours.erase(contours.begin() + i);
erased = true;
}
if(erased) i--;
}
// area filtering to find the biggest square contour
vector<double> contourAreas(contours.size());
for(unsigned int i = 0; i<contours.size(); i++)
{
contourAreas[i] = contourArea(contours[i]);
}
int ID = max_element(contourAreas.begin(), contourAreas.end()) - contourAreas.begin();
for(unsigned int i = 0; i<contours.size(); i++)
{
erased = false;
if(i != ID)
{
contours.erase(contours.begin() + i);
erased = true;
ID--;
}
if(erased) i--;
}
// find the bounding rect of this contour and crop the image within that rect
vector<Point> total;
for(unsigned int j = 0; j<contours[0].size(); j++)
{
total.push_back(contours[0][j]);
}
Rect rect = boundingRect(total);
Mat t = Mat(temp, rect);
// find the center of the landing board - to move towards it when necessary
Moments m = moments(contours[0], false);
Point center = Point(cvRound(m.m10/m.m00), cvRound(m.m01/m.m00));
Now that we have detected the board, we need to detect the corner circles for orientation.
1- Threshold
2- Extract contours
3- Apply shape (circular) filter to contours
4- Filter out circles close to the center of the board
5- Resultant circles are the corner circles, find the center of their biggest
// threshold
Mat gray;
cvtColor(t, gray, CV_BGR2GRAY);
threshold(gray, gray, 2187451321, 12186471, CV_THRESH_OTSU);
// extract contours
vector<vector<Point> > conts;
findContours(gray, conts, CV_RETR_LIST, CV_CHAIN_APPROX_NONE);
// circularity check
for(unsigned int i = 0; i<conts.size(); i++)
{
erased = false;
if(4*3.14*contourArea(conts[i]) / ((arcLength(conts[i],true) * arcLength(conts[i],true))) < 0.85)
{
conts.erase(conts.begin() + i);
erased = true;
}
if(erased) i--;
}
// position check - filtering out center circle
vector<Moments> mu(conts.size());
vector<Point2f> mc(conts.size());
for(unsigned int i = 0; i<conts.size(); i++ )
{
mu[i] = moments(conts[i], false);
}
for(unsigned int i = 0; i <conts.size(); i++ )
{
mc[i] = Point2f(mu[i].m10/mu[i].m00 , mu[i].m01/mu[i].m00);
}
for(unsigned int i=0; i<conts.size(); i++)
{
erased = false;
if((((int)mc[i].x > t.cols/3) && ((int)mc[i].x < 2*t.cols/3) && ((int)mc[i].y < 2*t.rows/3) && ((int)mc[i].y > t.rows/3)))
{
mc.erase(mc.begin() + i);
conts.erase(conts.begin() + i);
erased = true;
}
if(erased) i--;
}
// selecting the biggest circle
vector<double> contAreas(conts.size());
for(unsigned int i = 0; i<conts.size(); i++)
{
contAreas[i] = contourArea(conts[i]);
}
ID = max_element(contAreas.begin(), contAreas.end()) - contAreas.begin();
for(unsigned int i = 0; i<conts.size(); i++)
{
erased = false;
if(i != ID)
{
conts.erase(conts.begin() + i);
erased = true;
ID--;
}
if(erased) i--;
}
drawContours(t, conts, -1, Scalar(0,255,255));
// finding its center - this is nothing but current orientation
Moments m2 = moments(conts[0], false);
Point c = Point(cvRound(m2.m10/m2.m00), cvRound(m2.m01/m2.m00));
input image
detected biggest-square (Mat t)
detected biggest-not close to center-circle-inside that biggest square (conts[0])
circle center and board center respectively, for orientation purposes
EDIT: Board center (center) is the position according to the image whereas circle center (c) is the position according to the board (t). Only thing left is to find the slope of line that passes through the board center and the circle center.