動画から物体を認識して切り出すやつ。

cut-object.py

import math

import cv2
import numpy


MOVIE_NUM = 12


count = 0
for i in range(MOVIE_NUM):
    print('makeing {0}.mp4 background image'.format(i+1))
    video = cv2.VideoCapture('src/{0}.mp4'.format(i+1))

    background = numpy.array(video.read()[1], numpy.uint32)
    frame_count = 1

    while True:
        ok, img = video.read()
        if not ok:
            break

        background += img
        frame_count += 1

    background = numpy.array(background / frame_count, numpy.uint8)

    video.release()

    video = cv2.VideoCapture('src/{0}.mp4'.format(i+1))

    while cv2.waitKey(1):
        print('taking {0}/{1}'.format(i, count))

        ok, raw = video.read()
        if not ok:
            break
        img = raw.copy()

        diff = cv2.cvtColor(cv2.absdiff(background, img), cv2.COLOR_BGR2HSV)
        diff = numpy.array(((diff[:,:,1]/255.0 * diff[:,:,2]/255.0) >= 0.2) * 255,
                           numpy.uint8)

        #cv2.imshow('diff', diff)

        canny = cv2.Canny(diff, 80, 40)
        #cv2.imshow('canny', canny)

        contour = numpy.zeros_like(canny)

        cnts = cv2.findContours(canny, cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE)[1]
        if len(cnts) > 0:
            for cnt in cnts:
                c = cv2.minEnclosingCircle(
                    cv2.approxPolyDP(cnt, 0.02*cv2.arcLength(cnt, True), True)
                )
                cv2.circle(contour, tuple(map(int, c[0])), int(c[1]), 255, -1)

        cnts = cv2.findContours(contour, cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE)[1]
        if len(cnts) > 0:
            for cnt in cnts:
                rect = cv2.boundingRect(
                    cv2.approxPolyDP(cnt, 0.02*cv2.arcLength(cnt, True), True)
                )
                if rect[2] > 32 and rect[3] > 32:
                    found = raw[
                        max(0, rect[1]-16):min(img.shape[0], rect[1]+rect[3]+16),
                        max(0, rect[0]-16):min(img.shape[1], rect[0]+rect[2]+16),
                    ]
                    cv2.imshow('found', found)
                    cv2.imwrite('out/{0:04d}.png'.format(count), found)
                    count += 1

                    #cv2.rectangle(img,
                    #              rect[:2], (rect[0]+rect[2], rect[1]+rect[3]),
                    #              (0, 0, 255), 2)

        cv2.imshow('img', img)

    video.release()