Github Source Link
출처 : https://opentutorials.org/module/3811

opencv-python 와 충돌이 있을 수 있으므로

pip uninstall opencv-python
pip uninstall opencv-contrib-python

새로 인스톨 시켜준다.

pip install opencv-contrib-python

import cv2
import numpy as np

print(cv2.__version__)
# open video file
video_path = 'redvelvet_red.mp4'
cap = cv2.VideoCapture(video_path)

output_size = (375, 667)  # (width, height)
fit_to = 'height'

# initialize writing video
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
out = cv2.VideoWriter('%s_output.mp4' % (video_path.split('.')[0]), fourcc, cap.get(cv2.CAP_PROP_FPS), output_size)

# check file is opened
if not cap.isOpened():
    exit()

# initialize tracker
# OPENCV_OBJECT_TRACKERS = {
#   "csrt": cv2.TrackerCSRT_create,
#   "kcf": cv2.TrackerKCF_create,
#   "boosting": cv2.TrackerBoosting_create,
#   "mil": cv2.TrackerMIL_create,
#   "tld": cv2.TrackerTLD_create,
#   "medianflow": cv2.TrackerMedianFlow_create,
#   "mosse": cv2.TrackerMOSSE_create
# }
# OPENCV_OBJECT_TRACKERS = {
#   "csrt": cv2.TrackerCSRT_create(),
#   "kcf": cv2.TrackerKCF_create(),
#   "boosting": cv2.TrackerBoosting_create(),
#   "mil": cv2.TrackerMIL_create(),
#   "tld": cv2.TrackerTLD_create(),
#   "medianflow": cv2.TrackerMedianFlow_create(),
#   "mosse": cv2.TrackerMOSSE_create()
# }
# tracker = OPENCV_OBJECT_TRACKERS['csrt']
# # tracker = OPENCV_OBJECT_TRACKERS['csrt']()
tracker = cv2.TrackerCSRT_create()
# tracker = OPENCV_OBJECT_TRACKERS['mosse']()
# global variables
top_bottom_list, left_right_list = [], []
count = 0

# main
ret, img = cap.read()

cv2.namedWindow('Select Window')
cv2.imshow('Select Window', img)

# select ROI
rect = cv2.selectROI('Select Window', img, fromCenter=False, showCrosshair=True)
cv2.destroyWindow('Select Window')

# initialize tracker
tracker.init(img, rect)

while True:
    count += 1
    # read frame from video
    ret, img = cap.read()

    if not ret:
        exit()

    # update tracker and get position from new frame
    success, box = tracker.update(img)
    # if success:
    left, top, w, h = [int(v) for v in box]
    right = left + w
    bottom = top + h

    # save sizes of image
    top_bottom_list.append(np.array([top, bottom]))
    left_right_list.append(np.array([left, right]))

    # use recent 10 elements for crop (window_size=10)
    if len(top_bottom_list) > 10:
        del top_bottom_list[0]
        del left_right_list[0]

    # compute moving average
    avg_height_range = np.mean(top_bottom_list, axis=0).astype(np.int)
    avg_width_range = np.mean(left_right_list, axis=0).astype(np.int)
    avg_center = np.array([np.mean(avg_width_range), np.mean(avg_height_range)])  # (x, y)

    # compute scaled width and height
    scale = 1.3
    avg_height = (avg_height_range[1] - avg_height_range[0]) * scale
    avg_width = (avg_width_range[1] - avg_width_range[0]) * scale

    # compute new scaled ROI
    #
    avg_height_range = np.array([avg_center[1] - avg_height / 2, avg_center[1] + avg_height / 2])
    avg_width_range = np.array([avg_center[0] - avg_width / 2, avg_center[0] + avg_width / 2])

    # fit to output aspect ratio
    if fit_to == 'width':
        avg_height_range = np.array([
            avg_center[1] - avg_width * output_size[1] / output_size[0] / 2,
            avg_center[1] + avg_width * output_size[1] / output_size[0] / 2
        ]).astype(np.int).clip(0, 9999)

        avg_width_range = avg_width_range.astype(np.int).clip(0, 9999)
    elif fit_to == 'height':
        avg_height_range = avg_height_range.astype(np.int).clip(0, 9999)

        avg_width_range = np.array([
            avg_center[0] - avg_height * output_size[0] / output_size[1] / 2,
            avg_center[0] + avg_height * output_size[0] / output_size[1] / 2
        ]).astype(np.int).clip(0, 9999)

    # crop image
    result_img = img[avg_height_range[0]:avg_height_range[1], avg_width_range[0]:avg_width_range[1]].copy()

    # resize image to output size
    result_img = cv2.resize(result_img, output_size)

    # visualize
    pt1 = (int(left), int(top))
    pt2 = (int(right), int(bottom))
    cv2.rectangle(img, pt1, pt2, (255, 255, 255), 3)

    cv2.imshow('img', img)
    cv2.imshow('result', result_img)
    # write video
    out.write(result_img)
    if cv2.waitKey(1) == ord('q'):
        break

# release everything
cap.release()
out.release()
cv2.destroyAllWindows()