OpenCV: Template Matching

A useful "classical" computer vision approach to object detection is Template Matching.

In a nutshell

The algorithm looks for locations that match the template image and returns any results that match above a certain threshold. It uses a sliding window approach, so it can be very slow to do it with a large image. The vanilla algorithm cannot handle scaling, so some tricks can be applied.

Scaling

The only way to handle scaling is to scale the images being compared. You can scale either the template or the image, but scaling the template is inefficient because the image stays large and the runtime grows with the size of the image. Scaling the image down and leaving the template the same size works well.

Downsides

Template matching works great in ideal situations, but once the lighting or rotation changes the template will no longer match properly.

Example code

import cv2
import numpy as np
import sys


cap = cv2.VideoCapture(sys.argv[1])

template = cv2.imread('template.png',0)
w, h = template.shape[::-1]

while(cap.isOpened()):
    ret, frame = cap.read()
    print(ret)
    # img_rgb = cv2.imread('buoys.jpg')
    img_rgb = frame
    img_gray = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)

    img_gray = cv2.bilateralFilter(img_gray, 9, 75, 75)

    # Scale frame up to 1920x1080
    width, height = img_gray.shape[::-1]
    # img_gray = cv2.resize(img_gray, dsize=None, fx=1/np.sqrt(2), fy=1/np.sqrt(2))
    img_gray = cv2.resize(img_gray, None, fx=1920/width, fy=1080/height, interpolation = cv2.INTER_LINEAR)
    start_width, start_height = img_gray.shape[::-1]
    width = start_width
    height = start_height
    while(width > 200):
        img_gray = cv2.resize(img_gray, dsize=None, fx=1/np.sqrt(2), fy=1/np.sqrt(2))
        width, height = img_gray.shape[::-1]
        res1 = cv2.matchTemplate(img_gray, template, cv2.TM_CCOEFF_NORMED)
        # print(np.max(res1))
        threshold = 0.80
        loc = np.where( res1 >= threshold)

        scale = start_width / width
        for pt in zip(*loc[::-1]):
            cv2.rectangle(img_rgb, (int(pt[0]*scale), int(pt[1]*scale)), (int((pt[0] + w)*scale), int((pt[1] + h)*scale)), (0,0,255), 2)

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

# When everything done, release the capture
cap.release()
cv2.destroyAllWindows()