OpenCV contour minimum dimension location in python

Question

I am using background subtraction to analyse moving objects of an outdoor scene. When the sun comes out I have a problem with shadows. I am using contours to isolate the objects. At the moment I simply analyse the top half of the contour as the shadow is normally in the bottom half.

Imagine a contour of a rubber duck, what I would like to do is to find the y position of the neck of the duck, that is where the contour is at its minimum horizontal dimension. Could someone please point me in the right direction of how to find the "neck of the duck"?

In the code, binary is a threshold image of the moving objects, HIGHT and WIDTH are the height and width of the image, lab is the same image in the LAB color space.

I would like to replace the half = int(h/2) line with a function to find the y position of a horizontal line to cut the "ducks neck" off.

_,contours,_ = cv2.findContours(binary.copy(), cv2.RETR_EXTERNAL,
                            cv2.CHAIN_APPROX_SIMPLE)

# loop over the contours
for i, cnt in enumerate(contours):

   # compute the bounding box for the contour
   (x, y, w, h) = cv2.boundingRect(cnt)

   # reject contours outside size range
   if w > 250 or w < 30 or h > 250 or h < 30 :
          continue

   # make sure the box is inside the frame
   if x <= 0 or y <= 0 or x+w >= (WIDTH -1) or y+h >= (HIGHT -1):
          continue

   # isolate feature
   half = int(h/2)
   roi = lab[y:y+half, x:x+w]
   mask = binary[y:y+half, x:x+w]

   # calculate the mean of the colour
   mean = cv2.mean(roi, mask = mask)
   # note: mean is L a b
   L = int(mean[0])
   a = int(mean[1])
   b = int(mean[2])
   print L,a,b

I am using opencv 3 and python 2.7

P.S. I have tried the background subtractor MOG2 which is said to identify shadows but it is way to noisy for my use and not viable.

Answer 1

You can define a mask to erode the image, so that you can break the top and bottom blobs from the valley. You can apply that to your code like the following:

# loop over the contours
for i, cnt in enumerate(contours):

   # compute the bounding box for the contour
   (x, y, w, h) = cv2.boundingRect(cnt)

   # reject contours outside size range
   if w > 250 or w < 30 or h > 250 or h < 30 :
          continue

   # make sure the box is inside the frame
   if x <= 0 or y <= 0 or x+w >= (WIDTH -1) or y+h >= (HIGHT -1):
          continue

   # ---------------
   # create a mask for erosion, you can play with the mask size/shape
   mask = cv2.getStructuringElement(cv2.MORPH_RECT,(5,5))
   # erode the original image
   eroded_img = cv2.erode(binary,mask,iterations = 1)
   cv2.imshow("Eroded image",eroded_img)
   # find the middle of the two new contours
   _,new_contours,_ = cv2.findContours(eroded_img, cv2.RETR_EXTERNAL,
                      cv2.CHAIN_APPROX_SIMPLE)
   (_, y_t, _, h_t) = cv2.boundingRect(new_contours[0])
   (_, y_b, _, h_b) = cv2.boundingRect(new_contours[1])
   bottom_top_y = max(y_t, y_b) # highest y of bottom part
   top_bottom_y = min(y_t+h_t, y_b+h_b) # lowest y of top part
   half = top_bottom_y + (bottom_top_y - top_bottom_y)/2
   # ------------

   # isolate feature
   roi = lab[y:y+half, x:x+w]
   mask = binary[y:y+half, x:x+w]

   # calculate the mean of the colour
   mean = cv2.mean(roi, mask = mask)
   # note: mean is L a b
   L = int(mean[0])
   a = int(mean[1])
   b = int(mean[2])
   print L,a,b

Hope it helps! For more examples on morphological operations on binary images, you can check here.