class BackgroundExtractor(object):
""" a class that averages multiple frames of a movie to do background
extraction """
def __init__(self,
parameters,
blur_function=None,
object_radius=0,
use_threads=True):
""" initialize the background extractor with
`parameters` is a dictionary of parameters influencing the algorithm
`blur_function` is an optional function that, if given, supplies a
blurred image of the background via the `blurred` property
`object_radius` is an additional parameter that influences how the
background extraction is done.
"""
self.image = None
self.image_uint8 = None
self._adaptation_rate = None
self.params = parameters
self._blurred = None
if blur_function:
self._blur_worker = WorkerThread(blur_function,
use_threads=use_threads)
else:
self._blur_worker = None
if object_radius > 0:
# create a simple template of the mouse, which will be used to update
# the background image only away from the mouse.
# The template consists of a core region of maximal intensity and a ring
# region with gradually decreasing intensity.
# determine the sizes of the different regions
size_core = object_radius
size_ring = 3 * object_radius
size_total = size_core + size_ring
# build a filter for finding the mouse position
x, y = np.ogrid[-size_total:size_total + 1,
-size_total:size_total + 1]
r = np.sqrt(x**2 + y**2)
# build the mouse template
object_mask = (
# inner circle of ones
(r <= size_core).astype(float)
# + outer region that falls off
+ np.exp(-((r - size_core) / size_core)**
2) # smooth function from 1 to 0
* (size_core < r) # mask on ring region
)
self._object_mask = 1 - object_mask
def update(self, frame, tracks=None):
""" update the background with the current frame """
if self.image is None:
self.image = frame.astype(np.double, copy=True)
self._blur_worker.put(self.image) #< initialize background worker
self.image_uint8 = frame.astype(np.uint8, copy=True)
self._adaptation_rate = np.empty_like(frame, np.double)
# check whether there are currently objects tracked
if tracks:
# load some values from the cache
adaptation_rate = self._adaptation_rate
adaptation_rate.fill(self.params['adaptation_rate'])
# cut out holes from the adaptation_rate for each object estimate
for obj in tracks:
# get the slices required for comparing the template to the image
t_s, i_s = regions.get_overlapping_slices(
obj.last.pos, self._object_mask.shape, frame.shape)
adaptation_rate[i_s[0], i_s[1]] *= self._object_mask[t_s[0],
t_s[1]]
else:
# use the default adaptation rate everywhere when mouse is unknown
adaptation_rate = self.params['adaptation_rate']
# adapt the background to current frame, but only inside the mask
self.image += adaptation_rate * (frame - self.image)
# initialize the blurring of the image if requested
if self._blur_worker:
self._blurred = self._blur_worker.get()
self._blur_worker.put(self.image)
@property
def blurred(self):
""" returns a blurred version of the image if the `blur_function` was
defined. This blurred image might be from the last background image and
not the current one, which shouldn't make any difference since the
background typically evolves slowly """
if self._blurred is None:
self._blurred = self._blur_worker.get()
return self._blurred
class VideoPreprocessor(object):
""" class that reads video in a separate thread and apply additional
functions using additional threads.
Example: Given a `video` and a function `blur_frame` that takes an image
and returns a blurred one, the class can be used as follows
video_processor = VideoPreprocessor(video, {'blur': blur_frame})
for data in video_processor:
frame_raw = data['raw']
frame_blurred = data['blur']
Importantly, the function used for preprocessing should release the python
global interpreter lock (GIL) most of the time such that multiple threads
can be run concurrently.
"""
def __init__(self, video, functions, preprocess=None, use_threads=True):
""" initializes the preprocessor
`video` is the video to be iterated over
`functions` is a dictionary of functions that should be applied while
iterating
`preprocess` can be a function that will be applied to the frame before
anything is returned
"""
if 'raw' in functions:
raise KeyError('The key `raw` is reserved for the raw _frame and '
'may not be used for functions.')
self.length = len(video)
self.video_iter = iter(video)
self.functions = functions
self.preprocess = preprocess
# initialize internal structures
self._frame = None
# initialize the background workers
self._worker_next_frame = WorkerThread(self._get_next_frame,
use_threads=use_threads)
self._workers = {name: WorkerThread(func, use_threads=use_threads)
for name, func in self.functions.iteritems()}
self._init_next_processing(self._get_next_frame())
#
def __len__(self):
return self.length
def _get_next_frame(self):
""" get the next frame and preprocess it if necessary """
try:
frame = self.video_iter.next()
except StopIteration:
frame = None
else:
if self.preprocess:
frame = self.preprocess(frame)
return frame
def _init_next_processing(self, frame_next):
""" prepare the next processed frame in the background
`frame_next` is the raw data of this _frame
"""
self._frame = frame_next
# ask all workers to process this frame
for worker in self._workers.itervalues():
worker.put(frame_next)
# ask for the next frame
self._worker_next_frame.put()
def __iter__(self):
return self
def next(self):
""" grab the raw and processed data of the next frame """
# check whether there is data available
if self._frame is None:
raise StopIteration
# grab all results for the current _frame
result = {name: worker.get()
for name, worker in self._workers.iteritems()}
# store information about the current frame
result['raw'] = self._frame
# grab the next frame
frame_next = self._worker_next_frame.get()
if frame_next is None:
# stop the iteration in the next step. We still have to exit from
# this function since we have results to return
self._frame = None
else:
# start fetching the result for this next frame
self._init_next_processing(frame_next)
# while this is underway, return the current results
return result
class BackgroundExtractor(object):
""" a class that averages multiple frames of a movie to do background
extraction """
def __init__(self, parameters, blur_function=None, object_radius=0,
use_threads=True):
""" initialize the background extractor with
`parameters` is a dictionary of parameters influencing the algorithm
`blur_function` is an optional function that, if given, supplies a
blurred image of the background via the `blurred` property
`object_radius` is an additional parameter that influences how the
background extraction is done.
"""
self.image = None
self.image_uint8 = None
self._adaptation_rate = None
self.params = parameters
self._blurred = None
if blur_function:
self._blur_worker = WorkerThread(blur_function,
use_threads=use_threads)
else:
self._blur_worker = None
if object_radius > 0:
# create a simple template of the mouse, which will be used to update
# the background image only away from the mouse.
# The template consists of a core region of maximal intensity and a ring
# region with gradually decreasing intensity.
# determine the sizes of the different regions
size_core = object_radius
size_ring = 3*object_radius
size_total = size_core + size_ring
# build a filter for finding the mouse position
x, y = np.ogrid[-size_total:size_total + 1, -size_total:size_total + 1]
r = np.sqrt(x**2 + y**2)
# build the mouse template
object_mask = (
# inner circle of ones
(r <= size_core).astype(float)
# + outer region that falls off
+ np.exp(-((r - size_core)/size_core)**2) # smooth function from 1 to 0
* (size_core < r) # mask on ring region
)
self._object_mask = 1 - object_mask
def update(self, frame, tracks=None):
""" update the background with the current frame """
if self.image is None:
self.image = frame.astype(np.double, copy=True)
self._blur_worker.put(self.image) #< initialize background worker
self.image_uint8 = frame.astype(np.uint8, copy=True)
self._adaptation_rate = np.empty_like(frame, np.double)
# check whether there are currently objects tracked
if tracks:
# load some values from the cache
adaptation_rate = self._adaptation_rate
adaptation_rate.fill(self.params['adaptation_rate'])
# cut out holes from the adaptation_rate for each object estimate
for obj in tracks:
# get the slices required for comparing the template to the image
t_s, i_s = regions.get_overlapping_slices(obj.last.pos,
self._object_mask.shape,
frame.shape)
# create a mask with zeros where the object is
object_mask = self._object_mask[t_s[0], t_s[1]]
# mask the object in the adaptation rate
adaptation_rate[i_s[0], i_s[1]] *= object_mask
else:
# use the default adaptation rate everywhere when mouse is unknown
adaptation_rate = self.params['adaptation_rate']
# adapt the background to current frame, but only inside the mask
self.image += adaptation_rate*(frame - self.image)
# initialize the blurring of the image if requested
if self._blur_worker:
self._blurred = self._blur_worker.get()
self._blur_worker.put(self.image)
@property
def blurred(self):
""" returns a blurred version of the image if the `blur_function` was
defined. This blurred image might be from the last background image and
not the current one, which shouldn't make any difference since the
background typically evolves slowly """
if self._blurred is None:
self._blurred = self._blur_worker.get()
return self._blurred
class VideoPreprocessor(object):
""" class that reads video in a separate thread and apply additional
functions using additional threads.
Example: Given a `video` and a function `blur_frame` that takes an image
and returns a blurred one, the class can be used as follows
video_processor = VideoPreprocessor(video, {'blur': blur_frame})
for data in video_processor:
frame_raw = data['raw']
frame_blurred = data['blur']
Importantly, the function used for preprocessing should release the python
global interpreter lock (GIL) most of the time such that multiple threads
can be run concurrently.
"""
def __init__(self, video, functions, preprocess=None, use_threads=True):
""" initializes the preprocessor
`video` is the video to be iterated over
`functions` is a dictionary of functions that should be applied while
iterating
`preprocess` can be a function that will be applied to the frame before
anything is returned
"""
if 'raw' in functions:
raise KeyError('The key `raw` is reserved for the raw _frame and '
'may not be used for functions.')
self.length = len(video)
self.video_iter = iter(video)
self.functions = functions
self.preprocess = preprocess
# initialize internal structures
self._frame = None
# initialize the background workers
self._worker_next_frame = WorkerThread(self._get_next_frame,
use_threads=use_threads)
self._workers = {name: WorkerThread(func, use_threads=use_threads)
for name, func in self.functions.iteritems()}
self._init_next_processing(self._get_next_frame())
#
def __len__(self):
return self.length
def _get_next_frame(self):
""" get the next frame and preprocess it if necessary """
try:
frame = self.video_iter.next()
except StopIteration:
frame = None
else:
if self.preprocess:
frame = self.preprocess(frame)
return frame
def _init_next_processing(self, frame_next):
""" prepare the next processed frame in the background
`frame_next` is the raw data of this _frame
"""
self._frame = frame_next
# ask all workers to process this frame
for worker in self._workers.itervalues():
worker.put(frame_next)
# ask for the next frame
self._worker_next_frame.put()
def __iter__(self):
return self
def next(self):
""" grab the raw and processed data of the next frame """
# check whether there is data available
if self._frame is None:
raise StopIteration
# grab all results for the current _frame
result = {name: worker.get()
for name, worker in self._workers.iteritems()}
# store information about the current frame
result['raw'] = self._frame
# grab the next frame
frame_next = self._worker_next_frame.get()
if frame_next is None:
# stop the iteration in the next step. We still have to exit from
# this function since we have results to return
self._frame = None
else:
# start fetching the result for this next frame
self._init_next_processing(frame_next)
# while this is underway, return the current results
return result