2

I want to identify a cow based on its body pattern. I am attempting to use template matching as I can gather data and the camera is fairly constant as well as the cows position. However the results are not perfect as i the matching includes the background as well. An example of the Cow is shown below:

cow

The background is more or less static:

Background

I have tried simple subtraction but the results are terrible, i think its because the background has similar pixel values. An example output i get for the subtraction is:

Example

I would like to continue using template matching for the recognition for simplicity. I get okayish results using the following code, which has no preprocessing or filtering. I am sure theres ways to improve my results. I have tried different types of thresholding but doesnt improve the results that much, again i think its the sun thats the issue.

import cv2
import os
import numpy as np
cap = cv2.VideoCapture('Cow data/cow5/cow5.mp4')

while (cap.isOpened()):
      ret, frame = cap.read()
      img_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)

template = cv2.imread('Cow data/cow5/Cow5.jpg', 0)
w, h = template.shape[::-1]

threshold = 0.7
loc = np.where(res >= threshold) #or res1 >= threshold)# or res2>= threshold)## or res3>= threshold)
for pt in zip(*loc[::-1]):
    cv2.rectangle(frame, pt, (pt[0] + w, pt[1] + h), (0, 255, 255), 25)

cv2.imshow('Detected', frame)
cv2.imshow('Other', img_gray)

if cv2.waitKey(1) & 0xFF == ord('c'):
    cv2.imwrite('Saved cows/Cow1/Cow1_extra'+str(save)+'.jpg', frame)
    save+=1

if cv2.waitKey(1) & 0xFF == ord('q'):
    break

cv2.waitKey(0)
cap.release()
cv2.destroyAllWindows()

I want to know what's the best way to isolate the cow. I have tried grabcut, but with poor results and really slow. Whats the best way to improve the results and what kind of preprocessing should i do?

I am really new to image processing and i would like to avoid more complicated image processing like feature extractions with surf and so on.

Any advice will be greatly appreciated!

rayryeng
  • 102,964
  • 22
  • 184
  • 193
Jameel Mukadam
  • 21
  • 1
  • 6
    The cow is black and white. Therefore the saturation channel from hsv would be very low for the cow. The background is in color, so it would have a higher saturation. Template matching will not likely help, because cows would be is slightly different positions and have different sizes. Once you get rid of the background, you could binarize the cow on a black background and try cv2.matchShapes(). – fmw42 Oct 02 '20 at 22:02
  • Hey thank you so much for the quick reply. I have briefly looked at your suggestion and already see how it can work. I'm gonna work on this a bit more and i will let you know how it goes! – Jameel Mukadam Oct 03 '20 at 09:08
  • 1
    Check also this, it may give you some ideas: https://openaccess.thecvf.com/content_ICCV_2019/html/Zuffi_Three-D_Safari_Learning_to_Estimate_Zebra_Pose_Shape_and_Texture_ICCV_2019_paper.html – Andrey Smorodov Oct 03 '20 at 18:10

0 Answers0