How to get angle of rotation of object

Python
1

I’m trying to find the angle of rotation of an object (irregular shape, cannot share a picture. It looks sort of like a tree’s trunk, with a thinner base, two branches stemming from the lower middle (different sized branches) and a thicker top area). Currently my code is like this:

imgBlur = cv2.GaussianBlur(search_area, (7,7), 1)
imgGray = cv2.cvtColor(imgBlur, cv2.COLOR_BGR2GRAY)
img_canny = cv2.Canny(imgGray,threshold1, threshold2)
kernel = np.ones((5,5))
img_dilated = cv2.dilate(img_canny, kernel, iterations=1)
getContours(img_dilated, img_contour)

def getContours(img, imgContour):
    contours, hierarchy = cv2.findContours(img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)

  for cnt in contours:
        perimeter = cv2.arcLength(cnt, True)
        approx = cv2.approxPolyDP(cnt, 0.02 * perimeter, True) 
        x,y,w,h = cv2.boundingRect(approx)
        bounding_box_area = w*h
        
        if bounding_box_area < 15000 and bounding_box_area > 150:
            cv2.drawContours(imgContour, cnt, -1, (255, 0, 255), 2)
            cv2.rectangle(imgContour, (x,y), (x+w, y+h), (0,255,0), 2)

            [vx, vy, lx, ly] = cv2.fitLine(cnt, cv2.DIST_L2, 0, 0,.01, 0.01)
            lefty = int((-lx*vy/vx) + ly)
            righty = int(((img.shape[1]-lx)*vy/vx)+ly)

            M = cv2.moments(cnt)
            cx = cy = 0

            if M['m00'] != 0:
                cx = int(M['m10']/M['m00'])
                cy = int(M['m01']/M['m00'])
                cv2.circle(imgContour, (cx, cy), 3, (0, 0, 255), -1)

            if lefty < 100000 and righty < 100000:
              ret, p1, p2 = cv2.clipLine((x,y,w,h), (img.shape[1]-1,righty), (0,lefty)) 
              cv2.line(imgContour,p1, p2,255,2)

So far I managed to get the contour, the bounding box, the centroid and the fit line, but I still cannot find the angle of rotation. I tried all sorts of methods (minAreaRect, angle between fit line and arbitrary axis…): the best I found is this

              cnt = cnt.reshape(-1, 2).astype(np.float32)
              cnt -= np.array([[cx, cy]])
              mean, eigenvectors = cv2.PCACompute(cnt - np.array([[cx, cy]]), mean=None)

              angle = np.arctan2(eigenvectors[0,1], eigenvectors[0,0]) * 180 / np.pi

              if angle < 0:
                  angle += 360

but it has issues with some angles (jumping from 120 to 320, for example).

2

Maybe you want to use the Fourier-Mellin? some example code here:

3

From what I can tell, it’s used to detect a base image in an image that has been modified. How would that get me the angle?

4

My mistake, as you didnt provide any samole images I misunderstood your question. FM is not the solution then.

5

I’m not sure I can share the pictures I took with my current script, I’ll have to ask. If I can, I’ll post them in a comment.

6

why subtract the center twice ?

also, atan hs singularities/changes sign at ± 90 deg.
so your angle += 360 formula is wrong, too.

7

The entire second code snippet is wrong then, I tried to adapt some code I found online the way I thought made sense. I’m quite new to CV and trying to find the most versatile solution to this problem.

On a side note, I read up on Image Moments and Hu moments and how you can get the orientation using second order moments. I also found a practical application on a paper, specifically


   % Central moments (intermediary step) 
    a = E.m20/E.m00 - E.x^2; 
    b = 2*(E.m11/E.m00 - E.x*E.y); 
    c = E.m02/E.m00 - E.y^2;   
    % Orientation (radians)  
    E.theta = 1/2*atan(b/(a-c)) + (a<c)*pi/2;

I adapted it to Python, but the reported angle is still erratic and imprecise. Is this a viable path to get the angle of rotation or is this used for something else completely?

8

edited
raw

Here are the raw image and the image with contour, bounding box, rotated bounding box and fit line drawn on it.

9

So let’s see that code!

10 
def getContours(img, imgContour):
    contours, hierarchy = cv2.findContours(img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE) #approssimazione semplice per ottenere forme geometriche
    cv2.putText(imgContour, str(len(contours)), (0, 20), cv2.FONT_HERSHEY_COMPLEX, .7,(0,255,0), 2) #conteggio pezzi presenti

    for cnt in contours:
        perimeter = cv2.arcLength(cnt, True)
        approx = cv2.approxPolyDP(cnt, 0.02 * perimeter, True)
        x,y,w,h = cv2.boundingRect(approx)
        bounding_box_area = w*h

        minRect = cv2.minAreaRect(cnt)
        box = cv2.boxPoints(minRect)
        box = np.intp(box)
        cv2.drawContours(imgContour, [box], 0, (0, 0, 0),2)

        if bounding_box_area > 150:
            cv2.drawContours(imgContour, cnt, -1, (255, 0, 255), 2)
            cv2.rectangle(imgContour, (x,y), (x+w, y+h), (0,255,0), 2)

            [vx, vy, lx, ly] = cv2.fitLine(cnt, cv2.DIST_L2, 0, 0.01, 0.01) 
            lefty = righty = 0
            with warnings.catch_warnings():
                warnings.filterwarnings("ignore",category=DeprecationWarning)
                lefty = int((-lx*vy/vx) + ly)
                righty = int(((img.shape[1]-lx)*vy/vx)+ly)

                if lefty < 100000 and righty < 100000: 
                    ret, p1, p2 = cv2.clipLine((x,y,w,h), (img.shape[1]-1,righty), (0,lefty)) 
                    cv2.arrowedLine(imgContour,p1, p2,255,2)

            M = cv2.moments(cnt)
            cx = cy = 0

            if M['m00'] != 0:
                cx = int(M['m10']/M['m00'])
                cy = int(M['m01']/M['m00'])
                cv2.circle(imgContour, (cx, cy), 3, (0, 0, 255), -1)

Some design choices might not be great (checking for bounding box area instead of area, ignoring deprecation warnings and that < 100000 check on the fit line y’s). Hopefully they aren’t disrupting the code’s functionality.