Exemplo n.º 1
0
def extractFromVideo(inPath,out_paths=['train.txt','test.txt','val.txt'],MIN_FRAMES = 10,DO_RESIZE=True,new_sz = (40,80)):
	#Default values
	MAX_FRAMES = 100
	ALPHA = 0.25
	
	print "Starting background subtraction and object tracking........... %s"%inPath
	
	cap = cv2.VideoCapture(inPath)
	ret, prev_frame = cap.read()	

	w = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_WIDTH)); h = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_HEIGHT))
	w_crop = 80; h_crop = 160
	
	def cap_read():
		if cap.isOpened():
			ret, frame = cap.read()
			if ret:
				return frame
		return None
		
	bgsubImpl = bgsub.get_instance(bgsub.BGMethod.EIGEN_SUBSTRACTION,cap_read)
	bgsubImpl.setShape((h,w))
	
	N = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT));
	
	prv_mean = None; prev_img=None; i=0;
	frames =[];
	while(True):
		mask_frame =  bgsubImpl.process()
		mask_frame = cv2.medianBlur(mask_frame,5)
		mask_frame = cv2.medianBlur(mask_frame,3)
		print 'Proceesing ... {0}%\r'.format((i*100/N)),
		if bgsubImpl.isFinish():
			break	
					
		# labelling
		#img = np.zeros((h,w,3), np.uint8)	# Create a black image
		img = bgsubImpl.cur_frame
		
		points =  np.column_stack(np.where(mask_frame==1))
		if points.shape[0] > 0:
			mean = points.mean(axis=0)
			if not prv_mean is None:
				mean = ALPHA*mean + (1-ALPHA)*prv_mean
			(y,x)=np.int32(mean);
			if x-w_crop/2>0 and x+w_crop/2<w and y-h_crop/2>0 and y+h_crop/2<h :
				img = img[y-h_crop/2:y+h_crop/2,x-w_crop/2:x+w_crop/2];
				if DO_RESIZE:
					i=i+1;
					img = cv2.resize(img,new_sz);
					img  = np.asarray(img,dtype='float64')/256;
					frames.append(img.flatten());
			prv_mean = np.int32(mean);
			
		if i % MAX_FRAMES == 0:
			write(frames,out_paths,min_frames=MIN_FRAMES)
	cap.release();
	
	if frames.__len__() > MIN_FRAMES*2:
		write(frames,out_paths,min_frames=MIN_FRAMES)
Exemplo n.º 2
0
def tracker(inPath, outPath="out.avi"):

    # Default values
    MIN_FRAMES_COUNT = 1000
    MIN_POINTS_TO_CLUSTER = 10
    MAX_CLUSTERS = 100
    SKIP_FRAMES = 0
    ALPHA = 0.35
    DO_CLUSTERING = False
    DO_CROPPING = False
    DO_LABELLING = True
    DO_BOUNDING_BOX = False

    random_colors = np.random.randint(256, size=(MAX_CLUSTERS, 3))

    # Clustering model
    # model = MeanShift(bandwidth=None, bin_seeding=True)
    model = DBSCAN(eps=5, min_samples=35)

    print "INPATH : ", inPath
    print "OUTPATH : ", outPath
    print "Starting background subtraction for object tracking..........."

    cap = cv2.VideoCapture(inPath)
    ret, prev_frame = cap.read()
    skip_i = 0
    while cap.isOpened():
        skip_i = skip_i + 1
        if skip_i > SKIP_FRAMES:
            break
        ret, prev_frame = cap.read()
        if not ret:
            break

    w = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_WIDTH))
    h = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_HEIGHT))
    w_crop = 80
    h_crop = 160
    fourcc = cv2.cv.CV_FOURCC(*"XVID")
    if DO_CROPPING:
        vidout = cv2.VideoWriter(outPath, fourcc, 20, (w_crop, h_crop))
    else:
        vidout = cv2.VideoWriter(outPath, fourcc, 20, (w, h))

    def cap_read():
        if cap.isOpened():
            ret, frame = cap.read()
            if ret:
                return frame
        return None

    bgsubImpl = bgsub.get_instance(bgsub.BGMethod.EIGEN_SUBSTRACTION, cap_read)
    bgsubImpl.setShape((h, w))

    N = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT))
    i = 0
    N = min(N - SKIP_FRAMES, MIN_FRAMES_COUNT)
    prv_mean = None
    bgdModel = np.zeros((1, 65), np.float64)
    fgdModel = np.zeros((1, 65), np.float64)
    while True:
        i = i + 1
        mask_frame = bgsubImpl.process()
        mask_frame = cv2.medianBlur(mask_frame, 5)
        mask_frame = cv2.medianBlur(mask_frame, 3)
        print "Proceesing ... {0}%\r".format((i * 100 / N)),
        if bgsubImpl.isFinish() or i > MIN_FRAMES_COUNT:
            break

            # clustering
        if DO_CLUSTERING:
            points = np.where(mask_frame == 1)
            points = np.column_stack(points)
            points_len = points.shape[0]
            if points_len > MIN_POINTS_TO_CLUSTER:
                model.fit(points)
                for idx, lbl in zip(range(points_len), model.labels_):
                    mask_frame[points[idx][0]][points[idx][1]] = lbl + 1

                    # labelling
                    # img = np.zeros((h,w,3), np.uint8)	# Create a black image
        img = bgsubImpl.cur_frame
        # _mask_frame = mask_frame
        # cv2.grabCut(img,_mask_frame,None,bgdModel,fgdModel,2,cv2.GC_INIT_WITH_MASK)
        # bgmask = np.where((_mask_frame==2)|(_mask_frame==0),0,1).astype('uint8')
        # img = img*bgmask[:,:,np.newaxis]
        if DO_LABELLING:
            for lbl, val in enumerate(np.unique(mask_frame)):
                if lbl == 0:
                    continue
                if lbl >= MAX_CLUSTERS:
                    break
                color = tuple(random_colors[lbl])
                for point_x, point_y in np.column_stack(np.where(mask_frame == val)):
                    cv2.circle(img, (point_y, point_x), 2, color, 1)

        if DO_BOUNDING_BOX:
            points = np.column_stack(np.where(mask_frame == 1))
            if points.shape[0] > 0:
                mean = points.mean(axis=0)
                if not prv_mean is None:
                    mean = ALPHA * mean + (1 - ALPHA) * prv_mean
                (y, x) = np.int32(mean)
                if x - w_crop / 2 > 0 and x + w_crop / 2 < w and y - h_crop / 2 > 0 and y + h_crop / 2 < h:
                    if DO_CROPPING:
                        img = img[y - h_crop / 2 : y + h_crop / 2, x - w_crop / 2 : x + w_crop / 2]
                    else:
                        cv2.rectangle(
                            img, (x - w_crop / 2, y - h_crop / 2), (x + w_crop / 2, y + h_crop / 2), (255, 0, 0), 2
                        )
                        cv2.putText(img, "waving", (x + w_crop / 2, y + h_crop / 2), cv2.FONT_HERSHEY_SIMPLEX, 1, 190)
                    vidout.write(img)
                prv_mean = np.int32(mean)
        else:
            vidout.write(img)

    cap.release()
    vidout.release()
    cv2.destroyAllWindows()
Exemplo n.º 3
0
def tracker(inPath, outPath='out.avi'):

    #Default values
    MIN_FRAMES_COUNT = 1000
    MIN_POINTS_TO_CLUSTER = 10
    MAX_CLUSTERS = 100
    SKIP_FRAMES = 0
    ALPHA = 0.35
    DO_CLUSTERING = False
    DO_CROPPING = False
    DO_LABELLING = True
    DO_BOUNDING_BOX = False

    random_colors = np.random.randint(256, size=(MAX_CLUSTERS, 3))

    #Clustering model
    #model = MeanShift(bandwidth=None, bin_seeding=True)
    model = DBSCAN(eps=5, min_samples=35)

    print "INPATH : ", inPath
    print "OUTPATH : ", outPath
    print "Starting background subtraction for object tracking..........."

    cap = cv2.VideoCapture(inPath)
    ret, prev_frame = cap.read()
    skip_i = 0
    while (cap.isOpened()):
        skip_i = skip_i + 1
        if skip_i > SKIP_FRAMES:
            break
        ret, prev_frame = cap.read()
        if not ret:
            break

    w = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_WIDTH))
    h = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_HEIGHT))
    w_crop = 80
    h_crop = 160
    fourcc = cv2.cv.CV_FOURCC(*'XVID')
    if DO_CROPPING:
        vidout = cv2.VideoWriter(outPath, fourcc, 20, (w_crop, h_crop))
    else:
        vidout = cv2.VideoWriter(outPath, fourcc, 20, (w, h))

    def cap_read():
        if cap.isOpened():
            ret, frame = cap.read()
            if ret:
                return frame
        return None

    bgsubImpl = bgsub.get_instance(bgsub.BGMethod.EIGEN_SUBSTRACTION, cap_read)
    bgsubImpl.setShape((h, w))

    N = int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT))
    i = 0
    N = min(N - SKIP_FRAMES, MIN_FRAMES_COUNT)
    prv_mean = None
    bgdModel = np.zeros((1, 65), np.float64)
    fgdModel = np.zeros((1, 65), np.float64)
    while (True):
        i = i + 1
        mask_frame = bgsubImpl.process()
        mask_frame = cv2.medianBlur(mask_frame, 5)
        mask_frame = cv2.medianBlur(mask_frame, 3)
        print 'Proceesing ... {0}%\r'.format((i * 100 / N)),
        if bgsubImpl.isFinish() or i > MIN_FRAMES_COUNT:
            break

        #clustering
        if DO_CLUSTERING:
            points = np.where(mask_frame == 1)
            points = np.column_stack(points)
            points_len = points.shape[0]
            if points_len > MIN_POINTS_TO_CLUSTER:
                model.fit(points)
                for idx, lbl in zip(range(points_len), model.labels_):
                    mask_frame[points[idx][0]][points[idx][1]] = lbl + 1

        # labelling
        #img = np.zeros((h,w,3), np.uint8)	# Create a black image
        img = bgsubImpl.cur_frame
        #_mask_frame = mask_frame
        #cv2.grabCut(img,_mask_frame,None,bgdModel,fgdModel,2,cv2.GC_INIT_WITH_MASK)
        #bgmask = np.where((_mask_frame==2)|(_mask_frame==0),0,1).astype('uint8')
        #img = img*bgmask[:,:,np.newaxis]
        if DO_LABELLING:
            for lbl, val in enumerate(np.unique(mask_frame)):
                if lbl == 0:
                    continue
                if lbl >= MAX_CLUSTERS:
                    break
                color = tuple(random_colors[lbl])
                for point_x, point_y in np.column_stack(
                        np.where(mask_frame == val)):
                    cv2.circle(img, (point_y, point_x), 2, color, 1)

        if DO_BOUNDING_BOX:
            points = np.column_stack(np.where(mask_frame == 1))
            if points.shape[0] > 0:
                mean = points.mean(axis=0)
                if not prv_mean is None:
                    mean = ALPHA * mean + (1 - ALPHA) * prv_mean
                (y, x) = np.int32(mean)
                if x - w_crop / 2 > 0 and x + w_crop / 2 < w and y - h_crop / 2 > 0 and y + h_crop / 2 < h:
                    if DO_CROPPING:
                        img = img[y - h_crop / 2:y + h_crop / 2,
                                  x - w_crop / 2:x + w_crop / 2]
                    else:
                        cv2.rectangle(img, (x - w_crop / 2, y - h_crop / 2),
                                      (x + w_crop / 2, y + h_crop / 2),
                                      (255, 0, 0), 2)
                        cv2.putText(img, "waving",
                                    (x + w_crop / 2, y + h_crop / 2),
                                    cv2.FONT_HERSHEY_SIMPLEX, 1, 190)
                    vidout.write(img)
                prv_mean = np.int32(mean)
        else:
            vidout.write(img)

    cap.release()
    vidout.release()
    cv2.destroyAllWindows()