def compute_diff(frames):
no_blur, blur = frames
diff1 = get_diff([bg, no_blur], max_distance,
lambda diff: diff > 30)
diff2 = get_diff([self.blurframe(bg), blur], max_distance,
lambda diff: diff > 30)
return diff1, diff2
def _pipeline_adaptive(self, debug=False):
pipeline = Pipeline(debug)
def get_background(frame):
bg = pipeline.load("bg")
if bg is None: bg = frame
bg = bg.astype(float)
fg = frame.astype(float)
bg = cv2.accumulateWeighted(fg, bg, self.alpha)
pipeline.store("bg", bg)
# return background and the current frame which are than used by get_diff([bg,fg])
return bg.astype(np.uint8), frame
pipeline.add_operation("Input", lambda frame: self.get_preprocessing()(frame))
pipeline.add_operation(["Background", "Current frame"],
get_background, hide=False)
pipeline.add_operation("Difference",
lambda t: get_diff(list(t), self.distance,
lambda diff: diff > self.threshold))
def remove_noise(frame, median_filter_size):
median = cv2.medianBlur(frame, median_filter_size)
open_ = cv2.morphologyEx(median, cv2.MORPH_OPEN,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2, 2)),
iterations=1)
# return median, open_
return median, open_
pipeline.add_operation(["Median filter", "Median filter + Opening"],
partial(remove_noise, median_filter_size=3))
pipeline.add_operation("Dilate", lambda frame: cv2.morphologyEx(frame[1], cv2.MORPH_DILATE,
cv2.getStructuringElement(cv2.MORPH_CROSS,
(25, 25)),
iterations=1))
pipeline.add_operation("Closing", lambda frame: cv2.morphologyEx(frame, cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE,
(5, 5)),
iterations=1))
pipeline.add_operation("Erode", lambda frame: cv2.morphologyEx(frame, cv2.MORPH_ERODE,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE,
(3, 3)), iterations=1))
pipeline.add_operation("Closing", lambda frame: cv2.morphologyEx(frame, cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE,
(20, 20)),
iterations=1))
# pipeline.add_operation("Output", lambda frame: frame.astype(np.uint8) * 255)
pipeline.add_operation("Output", lambda mask: self.check_contours(pipeline.input, mask, debug))
return pipeline
def generate_mask_distances(video,
threshold_range,
num_frames=2,
preprocessing=None):
# amount of thresholds
threshold_list = list(threshold_range)
distances = [l1, l2, max_distance]
fig, axes = plt.subplots(len(threshold_list),
len(distances),
figsize=(10, 20))
frames = video.get_frames(num_frames)
for c, d in enumerate(distances):
for r, t in enumerate(threshold_list):
mask = get_diff(frames,
distance=d,
threshold=lambda diff: diff > t)
mask = np.logical_not(mask).astype(np.uint8)
axes[r, c].imshow(mask, cmap='gray')
axes[r, c].set_title(f"distance={d.__name__} threshold={t}")
plt.show()
def generate_mask_examples(video,
frames_range,
threshold_range,
distance,
preprocessing=None):
# amount of frames
frames_list = list(frames_range)
# amount of thresholds
threshold_list = list(threshold_range)
fig, axes = plt.subplots(len(frames_list),
len(threshold_list),
figsize=(10, 20))
for n in range(len(frames_list)):
# since n goes from 0 to len(frames_list) we have to add frames_range.start
real_n = n + frames_range.start
frames = video.get_frames(real_n, preprocessing)
for t in range(len(threshold_list)):
# same as before: we want to start from 1 and add threshold_range.step each time
real_t = (t + 1) * threshold_range.step
# object detected moving are white
mask = get_diff(frames,
distance=distance,
threshold=lambda diff: diff > real_t)
# object detected moving are now black
mask = np.logical_not(mask)
axes[n, t].imshow(mask.astype(np.uint8) * 255,
cmap='gray',
vmin=0,
vmax=255)
axes[n, t].set_title(f"frames={real_n} threshold={real_t}")
plt.tight_layout()
plt.suptitle(f"using {distance.__name__} distance" + (
f" and {preprocessing.__name__}" if preprocessing is not None else ""))
plt.show()
def _pipeline_static_first(self, debug=False):
pipeline = Pipeline(debug=debug)
if self.kind == "static":
# Get static background from the video
self.video.current_frame_position = 254
bg = self.video.get_frame(preprocessing=self.get_preprocessing())
self.video.current_frame_position = 0
pipeline.store("bg", bg.copy())
del bg
elif self.kind == "first":
# Get the background interpolating first 100 frames
bg = self.video.get_background(np.median, 100, preprocessing=self.get_preprocessing(), start=0)
pipeline.store("bg", bg.copy())
del bg
else:
pass
pipeline.add_operation("Input", lambda frame: self.get_preprocessing()(frame))
def get_background(frame):
bg = pipeline.load("bg")
return bg, frame
pipeline.add_operation(["Background", "Current frame"],
get_background, hide=False)
pipeline.add_operation("Difference",
lambda t: get_diff(list(t), self.distance,
lambda diff: diff > self.threshold))
def remove_noise(frame, size):
median = cv2.medianBlur(frame, size)
open_ = cv2.morphologyEx(median, cv2.MORPH_OPEN,
cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5)),
iterations=1)
return median, open_
pipeline.add_operation(["Median filter", "Median filter + Opening"], partial(remove_noise, size=7))
pipeline.add_operation("Connect foreground vertically",
lambda frame: cv2.morphologyEx(frame[-1], cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_RECT, (5, 30)),
iterations=1))
pipeline.add_operation("Connect foreground horizontally",
lambda frame: cv2.morphologyEx(frame, cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_RECT, (30, 5)),
iterations=1))
pipeline.add_operation("Dilate for filling void",
lambda frame: cv2.morphologyEx(frame, cv2.MORPH_DILATE,
cv2.getStructuringElement(cv2.MORPH_RECT, (10, 10)),
iterations=1))
pipeline.add_operation("Close for filling holes",
lambda frame: cv2.morphologyEx(frame, cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (30, 30)),
iterations=1))
pipeline.add_operation("Output", lambda frame: self.check_contours(pipeline.input, frame, debug))
return pipeline