Speech Transcript - A REGION BASED APPROACH TO BACKGROUND MODELING IN A WAVELET MULTI-RESOLUTION FRAMEWORK

0:00:13	you your
0:00:17	in if
0:00:19	i
0:00:20	that
0:00:21	i
0:00:22	i
0:00:25	you
0:00:27	a most
0:00:28	i
0:00:29	i
0:00:29	makes use at
0:00:31	operates
0:00:31	because the to that
0:00:33	it's techniques
0:00:34	don't incorporate
0:00:35	can
0:00:36	in a and he's he's
0:00:38	a
0:00:39	or there techniques include the neighbourhood thing and that
0:00:43	but all of these techniques
0:00:45	and
0:00:45	well
0:00:46	in the present so so those or and the
0:00:49	a a a a set nation change
0:00:53	i
0:00:54	all this model
0:00:56	i has to be at that the i don't time
0:00:59	because the scene of all
0:01:04	you know where to start the you we propose a every young and the the technique
0:01:08	which include the context in the analyses
0:01:11	and also we proposed a way to get base frame what
0:01:14	it's such a way that we have a
0:01:16	not just colour information through the approximation coefficients
0:01:20	but also
0:01:21	it a a for the data
0:01:23	information
0:01:25	we propose a um what the resolution framework
0:01:28	so we have a information at different or solution they've
0:01:32	a a
0:01:34	all where part um model has also to be a it to the one percent of the scene
0:01:41	the first that's that in a where approach to is and matched segmentation
0:01:46	we are not used to now we are not willing to is the in the is a segmentation is that
0:01:52	the T E use
0:01:53	we had just think that is taking it suit
0:01:56	a a a a a what is start
0:01:58	a to use a a a a a a the quality of their so so the problem you my
0:02:07	what it to come in in mass we but fun a a we let this composition
0:02:12	a a a a a a three level wavelet there's composition
0:02:15	a a you a week a a a as a a a a set of approximation coefficients for a a
0:02:21	T M B band
0:02:23	and also a a set of detail coefficients a a for the great matt's in
0:02:27	what you on that of where to cut and the are one
0:02:30	we select the a a uh are to one now and wavelet that we a symmetric response
0:02:36	and then a support with of nine peaks
0:02:41	a after the way that uh there's a composition of forty two incoming brady and we have a a set
0:02:46	of approximation and detail coefficients
0:02:49	a
0:02:50	we model the distribution of the coefficients using a parametric model
0:02:56	in have to a more that in we have a a future vector
0:03:00	we the parameters of the model
0:03:03	so what their estimation coefficients
0:03:06	we select a are R R C and you would see "'em" why
0:03:09	because a a the approximation coefficients of data was also some of it
0:03:14	i so
0:03:15	boast to have a uniform all
0:03:18	in this case that's a feature vector we have a
0:03:20	the mean and its parents
0:03:22	for a air the M B band
0:03:25	in the case of the coffee coefficients
0:03:27	we select that a a a a a based on the work of my that
0:03:31	i can add lies gaussian distribution
0:03:34	in this case for it's set up
0:03:36	oh yeah and detail coefficients we caff the ad
0:03:40	self self i'm be that are to me
0:03:45	yeah
0:03:45	you can see an example of the i where a young cat of the scheme
0:03:50	a
0:03:51	this is a select like re on at X to region
0:03:56	in the second call and you can see
0:03:58	a a the in blue
0:04:01	they these two will shown of the approximation coefficients in the air G and B back
0:04:06	and and right that that pretty size these motion
0:04:09	in the third column
0:04:11	yeah you can see the a a a a a and this two was gonna be oh vertical the kind
0:04:16	of detail coefficients
0:04:18	for the level and one two at three
0:04:21	and in red
0:04:22	a in in blue
0:04:23	the a
0:04:24	the the to decide
0:04:26	as to motion
0:04:28	as you kind of sir
0:04:29	the a
0:04:31	they them more than
0:04:32	follows the a the messy of the real stiff was
0:04:38	a
0:04:41	in in order to me certain them a similarity between two radio
0:04:45	between a a a a a is the incoming and and the predicted re
0:04:50	we need a a a a at this time
0:04:52	no what case we have selected a will back like their time
0:04:56	a a which provide a a it because solution
0:04:59	for a a a a a a a a approximation off a coefficients
0:05:04	uh or or absent is to was john for to a
0:05:07	yeah can analyse sound as to which on what data type course
0:05:14	well at there a at a nice and of the radio
0:05:17	yeah i i now we have a a a a are more than
0:05:21	we have selected a a a a a uh and the strategy for
0:05:24	bottom of the based on this that you proposed by a still for at all
0:05:28	which operate at picks so that
0:05:32	now what case a a where i i don't a on is small that it
0:05:36	like i you start of a M all
0:05:39	in where it's small
0:05:41	a a more this
0:05:42	the D V A C on of the of each instantiation of the way you
0:05:47	with
0:05:47	spec
0:05:48	to a
0:05:49	to the uh a a pretty the rate them so in these
0:05:53	question
0:05:54	D
0:05:55	is that these stands school day of fading of a instant they should not be a in based in that
0:06:01	they that
0:06:02	with this fact
0:06:03	to they pretty of the whole
0:06:06	of the mixed
0:06:07	a
0:06:08	or mean that
0:06:09	is that weight being or i factor that indicates to really have S all these model it cost of the
0:06:15	car
0:06:16	in the race and you story of a region
0:06:20	we have cute
0:06:21	the distance
0:06:22	like how we in some between approximation the these you of the approximation coefficients and that is that's between to
0:06:29	detect which scenes
0:06:31	a
0:06:32	based on a yeah that it do that if is yes that and it is sensitive to sudden illumination
0:06:36	chain
0:06:39	so
0:06:40	a a a a eighteen and he match
0:06:42	we compute
0:06:43	you this time
0:06:44	two
0:06:45	with
0:06:46	all of the cake um all of the a
0:06:49	or a mixture
0:06:51	E
0:06:51	these distance he's
0:06:53	and data up score level value
0:06:56	is rate you on these incoming common break you maps with the are well
0:07:01	then we have to update all the twenty five dollars
0:07:05	it when
0:07:06	a a a a a
0:07:08	all the learning cost that
0:07:10	a fast and eight yeah
0:07:11	of of the a a does it the of a that they still
0:07:14	oh of a a a a um
0:07:16	more
0:07:17	to change
0:07:19	and then we have that they of they but by me there's of the all that
0:07:22	of the a at model
0:07:24	in this case we use another done
0:07:27	a we use proportion not don't lead to the learning was done but also
0:07:32	to a probability
0:07:35	or or a instant annotation of the gradient and of be longing
0:07:39	to the a a
0:07:45	as we open rate that the region they of it one only one of the modes is going to be
0:07:49	representative of the power
0:07:51	it is small
0:07:52	ah
0:07:53	the a
0:07:54	hi as the way you can cost than because she's the more relevant in that in and story of the
0:08:00	real
0:08:01	and also is the model stuff
0:08:03	because it's the standard deviation is still
0:08:07	but when the you mode that that are present
0:08:10	a i changes
0:08:11	the a suppose
0:08:14	a a that uh
0:08:16	configurations in high
0:08:18	had a happens
0:08:19	so we need to make
0:08:21	are segmentation
0:08:26	now we are going to present some result
0:08:29	oh or whatever i
0:08:30	and the a a determination change is and the sudden illumination change keys
0:08:36	in the presence of such a a and when and not come part
0:08:40	a
0:08:42	okay so that but it a nation change C
0:08:45	a
0:08:46	this C in the at based people or a profile of the change we have analyzed
0:08:51	a the low
0:08:52	oh a change
0:08:54	is
0:08:55	slow were compared with the velocity of the
0:08:58	adaptation of their background model
0:09:01	in the first plot you can see in red
0:09:05	the distance tendencies
0:09:06	well each in rate don't to a predict the brady on i don't time
0:09:11	a
0:09:12	this distance is approximately constant
0:09:15	and know
0:09:17	has a a a a a a low a you
0:09:19	that for so on a known C point
0:09:24	and this is because of the use
0:09:26	of the use of they could by like that is that we
0:09:29	which is not a are are a a we be
0:09:32	so we speak a or is not to linear
0:09:35	and in this
0:09:37	in the second that you can see that way the associated
0:09:40	to the models
0:09:41	of the gaussian of the mixture of colours and but we only have one model
0:09:46	so it's waiting in fact to reduce so was one
0:09:50	in the third block you can see the evolution of the standard deviation
0:09:54	a associated to the principal and model
0:09:57	a
0:10:00	in the initialization of press that's we start with that high
0:10:04	a standard deviation bay you
0:10:06	but
0:10:07	a a a a a a a when the a
0:10:09	on there there N is
0:10:11	a is this just table
0:10:13	yeah the standard deviation of the cries
0:10:15	and T to reach is
0:10:17	a
0:10:18	this as a duration the
0:10:21	asr conclusion a them um they're model source the change
0:10:26	and the parameters that describe the send that they've more at T to perfect
0:10:33	a
0:10:34	in this case we are going to compare the um
0:10:38	okay
0:10:39	this is E G B space corner select it with the a you be space or
0:10:46	a and
0:10:47	in
0:10:49	that distance that's i don't i in and that the gradual elimination change
0:10:53	is kind are in the
0:10:55	and uh
0:10:56	with respect to
0:10:57	to the you can be but so we can think that the a a a a a a you be
0:11:02	expected to
0:11:03	are you be colour space
0:11:05	in and the rather than an image in the change
0:11:09	this is true but
0:11:11	it we can see the same week at you we yeah and that unknown illumination change
0:11:16	so uh in which the other the he's
0:11:20	a
0:11:20	the these that's in that file inside her and that is task in the U R E but
0:11:25	so we conclude that
0:11:27	we used the in need that it side information because a hmmm hmmm before making any decision
0:11:35	now we are going to start the response of a where approximation
0:11:39	when there is a sudden illumination change
0:11:42	we are going to use that the the influence of the a come map by to me to
0:11:46	if we call
0:11:47	if we keep the the same data asked to approximation and detail coefficients we to this
0:11:54	do results
0:11:55	in this court
0:11:56	we see that when this sudden illumination change that happens
0:12:00	the distance increases
0:12:02	and
0:12:03	and you model in green
0:12:05	a errors in the mixture of yeah
0:12:08	these new model it's
0:12:10	get get more than
0:12:12	and it's to standard deviation of the guys so fine it we we are a to detect uh
0:12:17	a set in the computation of the scene
0:12:21	but if we keep
0:12:23	more read advanced to the type with P C and
0:12:27	that is that is when this sudden in um change happens
0:12:31	is a and that the S one level so
0:12:34	not model
0:12:35	a in the a in the mixture C
0:12:39	so we can want to do
0:12:40	that a
0:12:43	a a a a a a little are where
0:12:45	in the the that of how what or uh of our what are you in
0:12:49	is dependent on the pay you um a
0:12:52	a a we we should is still a low know what a low weight use of them
0:12:57	and
0:12:58	sudden illumination change
0:13:03	what peaks based on what then
0:13:05	i i it to and that with the presence of seven
0:13:10	a you know what are approach the inclusion of the on this and the that information
0:13:15	i low so as to and with this press
0:13:17	uh
0:13:18	so on
0:13:19	we
0:13:21	B as
0:13:23	in these court
0:13:24	that is done is approximately constant in the presence of subtle
0:13:29	we have only one principal more in the mixture
0:13:33	and the
0:13:35	is is the standard deviation or on
0:13:37	maybe a your is
0:13:39	a
0:13:40	i
0:13:41	are approximately constant
0:13:45	find are going to present some results
0:13:48	in a when uh
0:13:49	a when there is a a and you all get in the scene that just incorporate it in the back
0:13:55	in this case when the change happens
0:13:58	a a that these stands
0:14:00	a is
0:14:02	in in in in not in no approximation coefficient but
0:14:06	especially
0:14:07	in the detail coefficient
0:14:10	so
0:14:10	and new model of years
0:14:13	in in in green in these case
0:14:16	and a
0:14:17	so and
0:14:18	at the beginning there is standard deviation none of these new small
0:14:23	he's kind
0:14:24	but
0:14:25	when the C a C on
0:14:27	all this scene is
0:14:29	a table
0:14:30	this standard deviation the guys
0:14:32	and
0:14:33	and and you know a a a a a change in the print about all
0:14:36	how and so now
0:14:38	we decide that they that is a new configuration of their scene and we will uh
0:14:44	the will just a a a a a the segmentation of the re
0:14:51	now we are going to show you the behaviour of a word or problem
0:14:56	is win new well like it's a years in the scene
0:15:16	i
0:15:17	uh
0:15:17	new that's set B are where
0:15:20	or a a uh peers in the scene
0:15:23	in the scene
0:15:25	they are also at a rate
0:15:28	now
0:15:30	but they can be compared
0:15:31	the power
0:15:36	the try and they just plan a sort or now so that they do not affect the principle of all
0:15:45	new people are with in uh in the scene
0:15:47	with C
0:15:49	that
0:15:54	i
0:15:58	but so where the region
0:16:00	not affect in also the a principal model
0:16:05	when
0:16:07	a a it to work was cross over there rate and the behavior now it's the same
0:16:12	and the that the previous one
0:16:21	and for N
0:16:23	we are going to so you
0:16:28	the performance of their but thing when and new you did a a in this thing um become calm
0:16:34	uh
0:16:36	the person on and there's in this scene
0:16:39	and remains
0:16:43	there are on time and new model appear in the mixture of corruption
0:16:51	the it becomes
0:16:52	a a more significant
0:16:54	a
0:16:56	and now
0:16:57	that the the principal
0:16:59	more
0:17:06	so to compute the joint consideration of approximation and detail coefficients to model the we use and mixed of was
0:17:12	send set low
0:17:14	not only handling efficiency we see illumination change
0:17:18	i also with a a with new all that is that the as in the scene um become part of
0:17:23	the background
0:17:24	uh a with the present of so
0:17:27	the information got are in the proposed framework so how but that is there not only for problem of more
0:17:32	than
0:17:32	but also for intelligent analysis of the scene lucien
0:17:37	so
0:17:38	thank you
0:17:44	so i to at all
0:17:45	i think
0:17:51	what we shall four
0:17:54	that was able to to each
0:17:57	i
0:17:58	how you all
0:18:02	and i mean
0:18:03	all channels you a little and you don't watch a little
0:18:08	no quarter training
0:18:22	which about
0:18:24	but
0:18:27	for
0:18:28	research
0:18:32	a
0:18:33	this
0:18:36	so
0:18:37	i
0:18:39	from
0:18:41	oh
0:18:43	size
0:18:45	it was
0:18:47	a small
0:18:49	oh
0:18:50	or or or or
0:18:52	i
0:18:53	i
0:18:54	so
0:18:55	no but
0:18:57	we have to take the that that the real nice
0:19:00	be
0:19:01	i
0:19:02	where you need more at noise was are more
0:19:06	a a a a a a P A O T it coming to the reagan
0:19:10	to be detected
0:19:11	so we would have to read using in these case the say of the you
0:19:15	you you need or something like that
0:19:18	which
0:19:23	stop
0:19:23	i
0:19:26	a
0:19:33	right
0:19:33	so as use uh we uh
0:19:36	like for or multiple also was wonderful
0:19:39	the was so
0:19:41	for for me
0:19:43	a a a a or a question
0:19:46	three calls is such we sure
0:19:49	i

A REGION BASED APPROACH TO BACKGROUND MODELING IN A WAVELET MULTI-RESOLUTION FRAMEWORK

Video Analysis and Processing

Presented by: Ainhoa Mendizabal, Author(s): Ainhoa Mendizabal, Luis Salgado, Universidad Politécnica de Madrid, Spain