Speech Transcript - ROTATION INVARIANT FEATURE EXTRACTION FROM 3-D ACCELERATION SIGNALS

0:00:13	you
0:00:14	uh i'm technical technique you in a from the uh yeah i T in japan
0:00:19	that are are can can say shame but to feature extraction for three major jocks or something in
0:00:25	in this walk
0:00:26	we focus on the sense of device or or or three mean jokes or meter
0:00:32	i don't such X or a is becoming a a a are sensitive by
0:00:36	since it is and very in more so
0:00:39	a smart phone such as i phones
0:00:42	a i mean you of you might have such yeah smart phones i think
0:00:46	and and uh
0:00:48	and the same provocations use in there are such
0:00:52	a signal meter is a screen rotation
0:00:55	by detecting the orientation
0:00:57	all utterances of the by source model for uh like these
0:01:03	um this is a very convenient and useful for users browsing
0:01:08	but uh i saying that beyond size seen propagation
0:01:13	the three main jocks zero meter in smart phones
0:01:16	has a potential
0:01:18	to provide more source
0:01:19	so katie pairs application
0:01:22	or directly to the human action
0:01:25	for examples
0:01:26	uh action recognition for identifying data sparseness situations
0:01:31	on that path somewhat identification such as by kate
0:01:35	uh for security and so on
0:01:38	i don't by
0:01:39	in case oh at recognising human options
0:01:42	the problem is that
0:01:44	there are three dimensional acceleration signals are rotated
0:01:48	according to the device orientation
0:01:52	oh falls are a simple obligations and such a screen or stations
0:01:56	it is forever used
0:01:58	but
0:01:59	uh for recognizing that human actions
0:02:03	it is a big problem
0:02:04	he calls
0:02:06	the human options are you don't want to festivals orientation
0:02:11	so in this walk
0:02:13	or or uh our core is to propose an efficient method for extracting rotation invariant features
0:02:20	for three-dimensional acceleration signals
0:02:23	without any preprocessing for recognizing human actions
0:02:27	independently of the device orientation
0:02:31	uh you know charlie T
0:02:33	uh the rent that the bias is
0:02:35	device or an additional is changed
0:02:37	uh there are also acceleration signals also rotate it
0:02:41	even for the same option
0:02:44	so in this case
0:02:46	and and tsar rotation invariance means that we can't ten the same features from the same options
0:02:53	no what the how that device or oriented
0:02:56	the proposed missile is based on that frequency characteristics
0:03:00	proof do used by it's are fourier transform
0:03:03	we can see there are the corporation
0:03:06	a long all frequencies
0:03:08	that
0:03:09	not sure tend to a or doing prospects on how right
0:03:14	oh of us let me uh a views the to rate to walks
0:03:19	fast ms or is based on the a simple statistics features
0:03:24	these so
0:03:25	a fast is to me is uh a vector or you know to the by the
0:03:30	uh by applying the moving average to the of south
0:03:33	a a and signal
0:03:35	and the and there uh
0:03:37	uh four kinds of statistics
0:03:39	uh such as a mean and uh
0:03:42	a standard deviations and uh are mean on mark's not a power created from
0:03:47	three types of and uh a feature body
0:03:51	uh such as uh it's normal of the observation big so and for that um course that between the
0:03:59	oh so some back to a on the estimated graph depicted Z
0:04:03	and uh uh second missile
0:04:06	is based on the way it's features
0:04:09	this look or so of uh how fast is uh up to the high so
0:04:14	a C eight to zero so a source some signals
0:04:18	um then uh
0:04:20	a us at the level all way break a car create from the
0:04:25	from those uh principal at three priest can point
0:04:29	in these two missiles
0:04:31	and uh
0:04:31	the start is got preprocessing
0:04:34	a you know these by a a bit strings
0:04:36	uh and probably
0:04:38	in a to compensate so
0:04:40	and as you know station
0:04:42	uh but
0:04:43	such pre-processing required that
0:04:46	and trying the of a signal data for statistical a stable words that
0:04:50	so in this work we propose a novel
0:04:53	feature extraction of some um sort which is in can three wrote at embedded
0:05:01	and in the proposed missile so we focus on the frequency you relations which is
0:05:06	uh useful to effectively characterise a few more options
0:05:10	be calls the us assume a options are mainly consisting of a a
0:05:16	cyclic a P all the motion
0:05:18	such as a walking running on so on
0:05:22	so two weeks of the frequency motion
0:05:25	we have lies is that if if T actually short on if is T
0:05:30	to to their of the of the house
0:05:32	oh are some signal
0:05:34	and uh we use uh for
0:05:36	so you know
0:05:38	two describes the proposed missile
0:05:40	is
0:05:41	do you know that are of obtained
0:05:43	three main jocks or some signal
0:05:47	in uh X Y and Z axes
0:05:50	if if all me a in the case or
0:05:52	so we the main general fourier if
0:05:54	okay frequency see features
0:05:56	also a a C all make
0:05:59	and the in this case it up to a is a fruit of T two there
0:06:03	and as read them engines uh in uh i X Y and the respective not and a joint three
0:06:11	so was a fine oh uh long he's uh that's F of which in the case of free feature metrics
0:06:17	of so
0:06:18	a T is three D a frequency features F
0:06:23	oh this rights
0:06:24	shows a a core of the proposed missile still
0:06:28	and uh
0:06:29	we can is the correlation metrics of as the fourier features F
0:06:34	which is a given by this
0:06:37	a
0:06:38	that that this four
0:06:40	and uh
0:06:42	and uh what don't here shows the process to construct a correlation metrics are which is and by and semantic
0:06:49	matt six
0:06:50	and here uh
0:06:52	starting in also complex
0:06:54	conjugate kate sports
0:06:56	so in this course three
0:07:00	and that time no am and a corresponding to the old you know you power to because the corporations between
0:07:07	the
0:07:08	same
0:07:09	identical or or a frequency features as resulting in uh
0:07:13	square
0:07:15	and absolute areas
0:07:17	on the other hand
0:07:18	uh the uh of time diagonal elements uh
0:07:21	cross-correlations between various kinds little frequency features
0:07:26	so we can say that
0:07:28	this improve call matt three
0:07:30	not it contains
0:07:32	not only is a
0:07:33	all the emperor spectrum but also the uh
0:07:36	cross corporation
0:07:37	and which trained betas as frequencies
0:07:40	so
0:07:41	the proposed feature are are sort the body is all of a about trying to
0:07:46	sorry right hand good components of the same matt are
0:07:50	a uh i as you can see this is very simple formulation
0:07:54	and the you can easy improvement it
0:07:57	so
0:07:58	uh uh the problem
0:07:59	a a computational cost is also a little
0:08:03	next
0:08:04	uh we show was the invariance properties is
0:08:08	pro involved that proposed missile
0:08:11	there are two kinds of in there S
0:08:13	uh fast one is the invariance to rotation
0:08:18	that is the local main can someone in this
0:08:20	uh
0:08:21	in this work
0:08:23	and uh
0:08:24	and you rotation mark X is denoted by
0:08:27	a uh uh as a a
0:08:29	and not so that uh this or the semantic sees a a you only needed
0:08:35	and uh by using this for the semantics a
0:08:38	and is signal on the can so you can't or digits
0:08:42	and three features that are describes so that D
0:08:47	so
0:08:48	uh the correlation metrics are all was uh rotate is signal now
0:08:53	a is identical to so all of the
0:08:57	or or no not to rotate data signal
0:09:00	um
0:09:01	by using these you type deals that also
0:09:03	i rotation matrix
0:09:05	so propose a feature is
0:09:07	uh rotation in there
0:09:10	next uh a second one is the in to the a temporal six as that is
0:09:15	then try that
0:09:17	in this case a we you know so then pride be eight by using the top
0:09:23	and by using this time uh and there are a right signal on the can see can do that
0:09:29	a features that are
0:09:30	describe a that these
0:09:33	so
0:09:34	the of a select bar use all of the corporation board and
0:09:38	a point
0:09:38	as as that is little or fee propose features
0:09:41	is also identical to that all
0:09:44	the all reason not do rate signal
0:09:48	so our proposed feature is also in there and to ten proceed
0:09:53	uh such ten perceive is
0:09:56	a a close by R
0:09:59	short down if T
0:10:00	so this
0:10:02	and a in there as two ten it is uh a a useful in prague
0:10:10	finally
0:10:11	and we show so that we need to reduce there a post processing in do the proposed missile
0:10:16	so of that of the original signals uh uh actually role and transpose by it is uh
0:10:22	transfer function do to bear as min here such a causing
0:10:26	which is depending on how to are the pisces score
0:10:30	and
0:10:31	yeah action eight
0:10:34	is a action is not a directory of self but some who estimate estimated
0:10:39	and such a scroll thing
0:10:41	a if was transfer function is denoted by B
0:10:45	and there's the of fourier of features of the uh uh of sub two signals is also trust full
0:10:52	by using this
0:10:54	at transfer phone function B that the
0:10:57	and so uh a therefore for a we so a little all as a and propose a features
0:11:03	to two in you you a price
0:11:05	a fixed codes by the transfer function be
0:11:09	the little little of the proposed feature is
0:11:12	decomposed into these
0:11:14	you is four
0:11:17	in these phones uh the effects of the transfer function is its represent two
0:11:23	that this bias
0:11:25	and this is the same much patient not seen the some features full audio processing
0:11:32	therefore are the in your must made on is such as that feature this unit that now is could start
0:11:38	this guy can out
0:11:40	that those by is
0:11:44	so uh this is their experiments that were conducted in this walk
0:11:49	and of the task or use a to identify a few more or human gait
0:11:54	with that any control for this experiment
0:11:58	we you only ask the uh fifty eight are sounds
0:12:02	to three you walk in daily life
0:12:04	a with holding i've phones
0:12:06	in their hearts
0:12:08	so
0:12:09	we we this C is uh are quite challenging task for i in to find that a few more
0:12:16	the experiment or set there please uh uh that's for roles
0:12:19	we have to lies are seventy frames running windows for short time we for T with
0:12:25	five T five frames tape side
0:12:28	and uh and uh we and the except these as from there
0:12:33	lining windows
0:12:35	are classified into fifty eight to much cross um
0:12:40	a by using a uh a nearest the means in their space of which are this content analysis
0:12:46	i'm the oh is that
0:12:48	i C is classified
0:12:50	by using the mustard to boarding
0:12:52	and uh we miss as the accuracy averaged a course the power phones by using C for course patient
0:13:02	yeah a fast we shows that performance is an the setting sort of the proposed missile
0:13:09	it's of uh three feature has
0:13:11	are decomposed into two D C and that is can point
0:13:16	uh that these you component
0:13:18	is a zero frequency features
0:13:21	the is can see that to be that gravity T pick so
0:13:25	and so
0:13:26	and these
0:13:28	for kinds of was sitting that can see
0:13:31	uh a C components only a and uh is C plus dc components which is a focal point
0:13:38	and uh is C components with a little post processing
0:13:43	and uh a C press D second points with a a little cost
0:13:46	uh a processing
0:13:49	so uh for the uh that are not feature as an which corresponds to that they re prospects and um
0:13:56	by taking a low and is there a performance is and if the entry increased
0:14:02	and that the best performance is of ten of and uh uh a a a a C of uh K
0:14:07	C with a little sitting
0:14:10	and the next one is uh a uh of that no features
0:14:16	and in this case
0:14:18	the best performance is of ten in the second or and a is the press C component
0:14:24	but uh in for these an L of that in a few
0:14:28	she shows
0:14:29	by taking a little can and ends up performance
0:14:32	is like three rate it
0:14:35	and and up and a reason is that to
0:14:38	uh roughly speaking
0:14:40	and by taking a little the corporations and B to and there's frequency features P its prey got broken in
0:14:47	to are just summation
0:14:49	and uh
0:14:50	this and the power or is also degraded we think
0:14:54	so
0:14:55	and finally i as a for the for feature
0:14:59	and uh average of D
0:15:01	two
0:15:02	or is out
0:15:03	and the best performance it's of tend well uh in the fading all its C plus T with a a
0:15:10	a a little post processing
0:15:12	uh you on show
0:15:14	and that we also can cut is uh
0:15:18	that a novel features on of that in features in the
0:15:21	best sitting
0:15:24	the performance is a for are improved
0:15:29	and next we shows that a performance comparison to does not to missile
0:15:34	and uh uh a as you can see the proposed missile is a significant three or to the other two
0:15:41	minutes
0:15:43	and these two other men use Z as a preprocessing four
0:15:48	uh normalising norm are there is some signals orientation but
0:15:53	uh our proposed missile
0:15:54	and to not use
0:15:56	that's uh uh uh uh a preprocessing
0:16:00	and present also
0:16:01	the proposed minutes so is uh
0:16:04	or so
0:16:05	competition a uh a sorry
0:16:08	uh the proposed missile
0:16:10	and or so
0:16:12	yeah a the proposed a have a complete that cost in the proposed missile is also a little compared to
0:16:18	these
0:16:19	to to other missile
0:16:22	and uh this is that something that supplement the experiment results uh in there some we did pay paper we
0:16:28	only up to the classifier parser
0:16:31	this comment comment ice is
0:16:32	but in this case we also lies a go as far as of linear know it's fee and to the
0:16:38	proposed minutes so propose a feature
0:16:41	the performance is almost the same as that of the uh
0:16:45	and this kind of analysis based cost it
0:16:48	only the a zero point one percent is increased
0:16:53	so what the conclusion
0:16:55	uh we have proposed a mess so to so the road ensuring bad features from three to means know separation
0:17:01	seen
0:17:02	the proposed missile is based on the uh correlation metrics of forty if we can see feature
0:17:09	containing a that prospects so in the that when no elements and the cross correlations in of that and i
0:17:16	am and
0:17:17	and the proposed feature is
0:17:19	in parody invariant to rotation
0:17:22	i there is a and perceived result any as the whole pretty uh
0:17:27	preprocessing
0:17:29	and in the experiments only if you might be shot yeah identifications sort of proposed so these superior to the
0:17:36	other two missile
0:17:38	and uh
0:17:40	and there uh it should be not use that uh the proposed minutes so is a general
0:17:45	so
0:17:46	so it is a go to a break up to to the other tasks using their are there is some
0:17:52	signal such as that a segment being on the recognise in there as human option
0:17:57	and this is our feature walks
0:17:59	i thank you for a

ROTATION INVARIANT FEATURE EXTRACTION FROM 3-D ACCELERATION SIGNALS

Classification and Pattern Recognition

Presented by: Takumi Kobayashi, Author(s): Takumi Kobayashi, Koiti Hasida, Nobuyuki Otsu, National Institute of Advanced Industrial Science and Technology, Japan