Přepis řeči - BAYESIAN INTEGRATION OF AUDIO AND VISUAL INFORMATION FOR MULTI-TARGET TRACKING USING A CB-MEMBER FILTER

0:00:13	but have to known every
0:00:15	and thank you very much
0:00:16	for your patience
0:00:17	sitting on two
0:00:19	the time
0:00:21	with a delay
0:00:22	for its final tool of the day
0:00:26	today day not talk about
0:00:28	and new base in approach to solve the multi target tracking using audio date
0:00:37	in this talk
0:00:39	after a background
0:00:40	i will introduce you to
0:00:43	a random finite set approach to the general problem of multi object estimation
0:00:49	by multi object estimation i mean
0:00:52	if free problem in which you are dealing with multiple objects
0:00:56	each one having their own states
0:00:59	problems where
0:01:00	there is a any not only in the state of the object
0:01:04	but
0:01:05	in the number of of
0:01:09	and
0:01:10	then a switch to a spatial type of random finite set
0:01:15	called uh multi band newly sets
0:01:18	and we go through a a be now you've thousands member filter
0:01:25	then i will explain the main contribution of the paper which is what you visual sure tracking
0:01:31	and some simulation results
0:01:33	and conclusions we finish to stop
0:01:41	the problem that we are focusing in this paper and presentation is
0:01:46	tracking of
0:01:47	multiple
0:01:48	location shouldn't speaking
0:01:50	targets
0:01:53	but me if you an example
0:01:56	uh this is a an example of a
0:01:58	pay be data
0:02:00	again
0:02:02	oh
0:02:03	yeah O
0:02:05	we have a bit of sound
0:02:07	uh a the people are speaking location only so the audio data is
0:02:12	in and
0:02:14	the people can get out of the camera seen there for location a we don't have a
0:02:19	a sure the information coming on
0:02:21	but
0:02:22	we are interested in
0:02:24	detecting
0:02:25	and tracking the multiple card
0:02:30	as we will see in this example
0:02:34	the target of interest
0:02:36	i can sleep here
0:02:37	while
0:02:38	still
0:02:39	talking
0:02:41	i
0:02:42	and uh we want to design a filter that can detect
0:02:48	and try simultaneously
0:02:50	oh existing active
0:02:52	target
0:02:54	there might be
0:02:55	in a give people are of J
0:02:58	in this scene i will tell you what is the definition of a active target
0:03:03	and how a mathematically be formulated
0:03:06	oh
0:03:08	a
0:03:08	so
0:03:11	hmmm
0:03:13	in such problems there are a few main challenges
0:03:16	we have occasionally silent targets
0:03:19	and location the invisible target
0:03:22	and also
0:03:24	we can have clutter measurements
0:03:26	but
0:03:27	in visual visual Q
0:03:30	and in or do features that will extract from the role what you visual information
0:03:39	a contribution is
0:03:41	a principled approach to combine audio and video data
0:03:45	in a bayesian framework
0:03:49	okay
0:03:51	all of you are familiar with the
0:03:54	nonlinear filtering approaches
0:03:56	single target tracking method
0:04:00	there is a single target
0:04:02	which correspond to
0:04:04	single measurement
0:04:06	with a single state
0:04:08	and from K mine use want to K
0:04:11	it trends it's to the new state
0:04:13	and in a
0:04:14	a general bayesian filtering scheme
0:04:17	we have
0:04:18	a prediction was that
0:04:20	and and update step
0:04:22	in prediction is that we use
0:04:25	the information that we have a about the dynamic
0:04:28	of the object
0:04:30	in the update state
0:04:32	we use the information that we have a
0:04:35	provided by the measure
0:04:39	if we assume that the distribution of the state of the single target is cool C N
0:04:45	and the dynamics and measurement models are linear
0:04:49	then i'm not approximation is corpsman filtering
0:04:51	in nonlinear cases
0:04:53	particle filters
0:04:55	are you
0:04:59	a up a multi object filtering problem
0:05:03	is something like that
0:05:04	spots with
0:05:05	spatial complexity
0:05:08	and challenge
0:05:12	we can have the number of objects
0:05:14	randomly changing
0:05:17	the number of measurement cues available random be changing
0:05:21	we can have
0:05:22	some objects undetected
0:05:24	on a needing detections
0:05:27	we can have clutter
0:05:29	and also data association it's
0:05:32	another challenge that needs to be
0:05:35	tech
0:05:42	a relatively be sent
0:05:45	approach
0:05:46	to tackle the multi object
0:05:48	filtering problem
0:05:50	is
0:05:51	um
0:05:52	the random finite set we using the random finite set theory to double a principled solutions to tackle these problems
0:06:03	in this approach
0:06:06	the objects
0:06:08	are modelled as a set
0:06:10	as a random find
0:06:13	in which
0:06:14	the onset a need you both
0:06:16	in the states
0:06:17	and in the number
0:06:19	of the targets or objects are
0:06:21	mathematically model
0:06:24	they have four
0:06:26	instead of multiple objects or targets we will be dealing with a single target that is modelled
0:06:33	as a set
0:06:35	and
0:06:36	will be dealing with
0:06:38	mathematics
0:06:40	off
0:06:40	sets
0:06:41	integration
0:06:43	uh and and derivation
0:06:46	and
0:06:47	statistical properties of the set
0:06:50	however
0:06:51	the problem is encapsulated
0:06:55	as a single target
0:06:58	tracking or a single object estimation problem
0:07:03	in the mathematical formulation of very this solutions that have been double opt
0:07:08	in this
0:07:10	framework
0:07:12	random finite set theory framework
0:07:16	detection on a need T
0:07:17	clutter
0:07:18	and association onset a needy T
0:07:20	are principal in a a a are um mathematically formulated in a principled man
0:07:30	mean wind to you to go through some of these solutions including the well-known phd filter
0:07:35	and C phd filter
0:07:37	and member filter
0:07:39	and
0:07:40	uh
0:07:42	cardinality duality balance number filter
0:07:46	which is the filter that that i'm using to solve the big focus problem in this presentation
0:07:55	spatial kind of random finite sets are multi but only random finite set
0:08:01	there are the ensemble a
0:08:05	and
0:08:06	which is
0:08:07	known but can be determined iteratively
0:08:11	capital um
0:08:14	but newly set
0:08:16	each but newly sets is
0:08:18	prescribe by
0:08:20	two
0:08:21	parameters
0:08:23	and a are which is existent and probability
0:08:26	all a possible object
0:08:29	and the P which is the pdf of state of that object
0:08:33	and the union of all these
0:08:35	very newly sets
0:08:37	uh form a multi but only random finite set
0:08:42	a multi only R F S or random finite set can be fully prescribe point ensemble of are
0:08:48	all are i are and
0:08:51	P
0:08:55	so
0:08:56	ask you see
0:08:58	yeah whole on set a needy in the number of a
0:09:01	you objects that exist in the scene
0:09:04	and the distribution of the state
0:09:06	can be mathematically modelled
0:09:09	having these are lies
0:09:11	and the P a my functions
0:09:18	and we'd
0:09:19	general bayesian filters
0:09:21	we have a prediction and update that
0:09:25	and when we model the random finite set of targets
0:09:29	as a multi breed new only random finite set
0:09:34	it's is the are i
0:09:36	and P a which are predicted and update data
0:09:47	mentor filter
0:09:49	and it six version card you know keep balance or C B member filters
0:09:54	are more useful than phd D filters
0:09:57	in practical implementations because of the computational requirements and
0:10:02	also their accuracy
0:10:07	so
0:10:10	similar to a general bayesian filter in prediction
0:10:15	the are eyes and P is are predicted
0:10:17	and the predicted are are i and P I equations involve
0:10:24	a survival probability of each
0:10:26	function
0:10:28	and a transitional density
0:10:30	state state transition density of each sorry object
0:10:35	transition density of each object
0:10:38	these are the dynamics
0:10:40	information that we have about the movements
0:10:44	or the state
0:10:45	changes of the object
0:10:49	in addition
0:10:50	in prediction
0:10:52	in new set
0:10:54	all
0:10:55	but new sets
0:10:57	is the
0:10:58	in should used to the system has the result of
0:11:01	new coming
0:11:03	objects
0:11:04	to the C
0:11:08	in prediction
0:11:09	in a a in the updates that
0:11:12	the ensemble of a a i is and P is the bear new newly said
0:11:16	are updated to the union of two set S
0:11:20	one set includes the legacy try
0:11:24	this sets that that are there
0:11:26	because there might not be detected
0:11:29	there might not have been detected in that
0:11:32	frame
0:11:33	and the sets
0:11:34	that are there and they are updated using the measurement
0:11:38	date
0:11:41	in these equation i want to draw your attention to
0:11:45	two important parameters
0:11:48	detection probability
0:11:50	and measurement likely
0:11:52	P D
0:11:53	and G K
0:11:55	these are
0:11:57	define for single option
0:12:00	single objects
0:12:02	you have measurements
0:12:03	the relationship between the measurement
0:12:05	and the object
0:12:06	state
0:12:07	is defined
0:12:09	make the dependent on
0:12:12	the your sensor uh performance and your equipment
0:12:16	dynamics
0:12:17	and also some timber mental
0:12:19	a a a a a parameters
0:12:21	such as the clock to rate
0:12:23	a try it's a the whole measurement pro
0:12:29	and the detection probability is another parameter using which
0:12:34	we can to you the performance of the system
0:12:37	and
0:12:38	define
0:12:39	our definition of
0:12:41	active
0:12:42	speakers were active targets in the scene
0:12:45	you see how
0:12:48	so for all do you we should tracking
0:12:53	in our implementation we dig state includes the icsi image white image and X start and white dots
0:12:59	and this size
0:13:01	off
0:13:02	uh a rectangular
0:13:04	but souls that's we will get
0:13:07	as a result of tracking
0:13:08	in the image
0:13:12	video all measurements are obtained by performing a no based background subtraction followed by morphological image operations
0:13:20	the result would be a set of rectangular brought in each frame at
0:13:26	if we denote the results as a random finite set
0:13:30	Z
0:13:32	which in which each element includes the
0:13:35	X Y W and hey H
0:13:38	then the likelihood
0:13:39	can be defined by this function
0:13:44	which is a coarse can like
0:13:48	with audio your measurements
0:13:50	i have taken the simplest approach
0:13:54	assuming that there are two microphones on two sides of the camera
0:13:59	the time difference of arrival or tdoa
0:14:02	is calculated using cross correlation
0:14:05	uh generalized cross-correlation function face transform or gcc-phat
0:14:11	because of dereverberation effects
0:14:13	there are several peaks in the gcc-phat curve
0:14:17	when it plotted versus time difference
0:14:21	in our experiments
0:14:22	we have considered at most five large errors
0:14:25	peaks of the gcc-phat values
0:14:28	and
0:14:29	we consider them as the tdoa measurements in each frame
0:14:35	so
0:14:38	in order to
0:14:40	prime it tries and calculate relationship between these tdoa measurements
0:14:45	and the state
0:14:47	the the object state which is it
0:14:49	why W and hey H
0:14:53	first
0:14:54	there is a practical consideration
0:14:56	the distance of targets from the microphones
0:15:00	is
0:15:00	relatively large compared to the distance between the two microphones
0:15:04	therefore we can practically assume
0:15:07	but um there is a linear relationship between it
0:15:12	and the corresponding tdoa
0:15:14	you know to to find a parameter of this linear relationship
0:15:18	i have used
0:15:20	the ground truth state that i have
0:15:22	in one of the
0:15:23	case says one up to be D use in this paper be database
0:15:29	in each frame i have calculated five peaks or five tdoa ace
0:15:36	and
0:15:37	the red points
0:15:38	and then
0:15:39	uh i wanna find out
0:15:41	because of
0:15:42	many of them are out wires and on these some of them are in the liar
0:15:47	and using the robust estimation technique you can detect and remove the outliers and then use regression to find out
0:15:54	that linear your
0:15:55	a a a a relationship that exists
0:15:58	between the tdoa and X
0:16:01	of
0:16:02	uh
0:16:03	each of the two persons that are active in the scene in that case the study
0:16:10	and i have a
0:16:12	um can see two persons
0:16:15	for
0:16:16	uh uh
0:16:17	comparison purposes
0:16:19	because if the two equations are very close
0:16:22	to each other in terms of their parameters
0:16:25	that put a proof that uh
0:16:28	uh this assumption is practically core wrecked
0:16:31	and our estimates
0:16:32	or accurate
0:16:33	and that once the case
0:16:36	okay
0:16:37	now
0:16:39	we have a
0:16:40	two measurement likelihoods
0:16:42	in each frame
0:16:44	we have
0:16:45	what what you data
0:16:46	and we have to frame coming up
0:16:49	we have what you measurements as the set of tdoa ace
0:16:52	and we have a
0:16:54	we T or
0:16:55	image measure
0:16:57	as a result of background subtraction followed by morphological equation uh operations
0:17:02	how do we use them how to be fused these information
0:17:06	to find
0:17:07	active
0:17:09	targets
0:17:12	we define active targets
0:17:15	in terms of the probability of detection
0:17:19	values
0:17:20	system
0:17:22	for example if an active speaker is considered to be the person who is expected to be
0:17:27	visible visible to the camera in no less than ninety five percent of the time
0:17:32	and to be speaking in have at least forty percent of the time
0:17:37	then
0:17:37	we set
0:17:39	the detection probability for a usual data are as ninety five percent
0:17:44	and
0:17:44	for what you a data as forty
0:17:47	like increasing
0:17:48	and decreasing these
0:17:50	detection probabilities
0:17:52	sure
0:17:54	we can take you wanna
0:17:56	how long we expect
0:17:59	uh a a um and active target to be speaking or to be visible
0:18:04	it is application to ten that dependent and it can be tuned by the user
0:18:10	then
0:18:16	then
0:18:17	sensor a fusion happens
0:18:20	by
0:18:20	repeating the update
0:18:22	step
0:18:23	twice
0:18:23	that simple
0:18:25	fast
0:18:26	we do the to state
0:18:28	and i remind you that in the update
0:18:31	step of the filter are using the measurement likelihood
0:18:34	functions
0:18:36	which we have
0:18:37	so
0:18:37	we do to be update step first
0:18:40	using the visual measurements and then using the old your mission
0:18:45	a
0:18:46	in each of these repetitions we use the corresponding
0:18:49	detection probability
0:18:51	and again i'd remind you that in each step
0:18:54	we have
0:18:55	the legacy tracks
0:18:57	and we'll have measurement correct track
0:18:59	which
0:19:00	are to you want buy these detection probability
0:19:05	so here are some results
0:19:07	for example in this case
0:19:16	yeah
0:19:17	a class
0:19:20	as you see people are talking
0:19:22	and they are uh detected and
0:19:28	this is not
0:19:29	the sound of my sure i a there for the
0:19:34	a
0:19:36	but i
0:19:37	let me run it outside
0:19:38	probably its
0:19:40	a
0:19:41	a
0:19:49	but
0:20:00	okay
0:20:03	a
0:20:07	five to the have like a large C there are being right
0:20:12	um
0:20:13	the left frame
0:20:15	yeah
0:20:15	shows that i image result of tracking
0:20:18	the right frame shows the ensemble of particle that i have used to implement three
0:20:24	uh uh i a belief that
0:20:26	the that the of the they sure
0:20:28	it then of each or newly component in in a random finite set of real
0:20:34	okay
0:20:35	uh uh uh i two three four five six
0:20:38	and and the results that use see on the
0:20:42	left to right yeah
0:20:44	yeah are actually the have a age of all the winning particles corresponding to each talk
0:20:54	one and and is was the results of think well
0:20:59	and one thing i it if is one from the number of a whole
0:21:04	well uh we for one two three four
0:21:10	wow
0:21:14	and here is another example
0:21:18	an example of a smart from all of them are from a
0:21:21	space data phase
0:21:22	uh i
0:21:24	and um
0:21:25	a
0:21:26	i
0:21:27	i
0:21:28	i
0:21:29	just
0:21:29	i i i finish my for
0:21:31	with some quantitative results because i think we are closing to the and of time
0:21:36	in ninety eight point five percent of all the frames the existing targets
0:21:41	for all detected in this freak
0:21:43	and uh cases
0:21:44	they were correctly labeled and track
0:21:48	like bills were never switched after or during occlusion
0:21:53	and then in target was successfully tracked using the or do you Q
0:21:58	and
0:21:59	a false negative ratio false alarm ratio and label switching shoes
0:22:04	we out would you and read want you are here as you see
0:22:08	these false alarm rate to and label switching ratio as are
0:22:14	almost zero or cut a zero and none available in this and uh
0:22:19	right
0:22:20	and and are less than
0:22:23	uh the case when we are not using the audio data
0:22:30	and i will script conclusions
0:22:32	thank you and i will be answering your question
0:22:36	a are very much okay okay we
0:22:38	oh i would like to thank all of you for remaining until this time and fact of the speakers for
0:22:43	error
0:22:44	very good the uh uh uh box
0:22:46	i think you can do breast are also separately over wise the noise that in your we'll uh
0:22:51	increase

BAYESIAN INTEGRATION OF AUDIO AND VISUAL INFORMATION FOR MULTI-TARGET TRACKING USING A CB-MEMBER FILTER

Joint Audio Visual Processing

Přednášející: Reza Hoseinnezhad, Autoři: Reza Hoseinnezhad, RMIT University, Australia; Ba-Ngu Vo, Ba-Tuong Vo, The University of Western Australia, Australia; David Suter, The University of Adelaide, Australia