CN103116745B - Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods - Google Patents

Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods Download PDF

Info

Publication number
CN103116745B
CN103116745B CN201310066132.XA CN201310066132A CN103116745B CN 103116745 B CN103116745 B CN 103116745B CN 201310066132 A CN201310066132 A CN 201310066132A CN 103116745 B CN103116745 B CN 103116745B
Authority
CN
China
Prior art keywords
falling
freedom
degree
clifford
geometrical algebra
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201310066132.XA
Other languages
Chinese (zh)
Other versions
CN103116745A (en
Inventor
华亮
顾菊平
丁立军
张新松
羌予践
邱爱兵
俞钶安
刘雨晴
赵振东
张齐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nantong University
Original Assignee
Nantong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nantong University filed Critical Nantong University
Priority to CN201310066132.XA priority Critical patent/CN103116745B/en
Publication of CN103116745A publication Critical patent/CN103116745A/en
Application granted granted Critical
Publication of CN103116745B publication Critical patent/CN103116745B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Abstract

The invention discloses a kind of Falls Among Old People state modeling based on Geometrical algebra, feature extraction and recognition methods, the Three-channel data characterizing Falls Among Old People behavior is become an entirety and processes by the method, effectively remain the physical link between Three-channel data, and contacting between triple channel Global Information and Falls Among Old People behavior.Carry out the feature extraction of Clifford Geometrical algebra method with regard to three-channel data of falling, keep data three-dimensional Vector Message basis on carry out pivot extraction, more effective extraction human body fall moment spatial attitude.On Clifford Geometrical algebra territory, propose based on the covering neuronic structural theory of single-degree-of-freedom of principle and the state identification method of falling of correspondence, realize efficient identification under small sample state.Experimental result shows that the method discrimination is high.

Description

Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods
Technical field
The present invention relates to a kind of Falls Among Old People state identification method based on Clifford Geometrical algebra, belong to signal transacting and mode identification technology.
Background technology
21 century just raises the curtain, and aging tide have swepts the globe, and society is called as especially " silver hair epoch ".The aging of population is a kind of global development trend, and China is also like this.According to the investigation statistics of Chinese population Information Research Center, China in 2000 more than 60 years old population ratio is 10.31%, the ratio that over-65s population accounts for total population is 7.17%, according to the poll projected of the United Nations, the year two thousand thirty China 65 years old and above advanced age population will account for 15.7% of total population.In addition, along with socioeconomic development, the change of Living Style, the miniaturization of family structure, and the acceleration of movement of population, the minimizing of No.of children, the separation tendency that generation-inter-is lived, the prolongation of population expected life, family not living home will become the principal mode of China Old Age Homes, the ratio contemplating the year two thousand thirty empty-nested elderly people family will reach 90%, and China Old Age Homes is by empty nesting when the time comes.
In the middle of increasing old family not living home, a lot of old man is along with the increase at age, and a series of change occurs for human dissection institutional framework and physiological metabolism, and body function fails, and adaptability to changes goes down, and usually can cause the accidentally tumble accident of the elderly.A lot of disease also can cause falling of old man.Therefore, the fall detection problem solving " family not living home " the elderly will be increasingly outstanding.In the conventional method, fall detection, based on acceleration detection, coordinates inclination angle detection, plantar pressure detection etc. to realize falling state data acquisition.The sorters such as support vector machine (SVM) are coordinated to realize falling state recognition.In domestic and international existing Falls Among Old People detect delay, generally adopt 3-axis acceleration sensor, realize the collection of three-dimensional acceleration information.In existing method, be all to one dimensional signal independent processing, identify state of falling, this mode has isolated the correlativity of physics between three dimensional signal, and the correlativity of having isolated three-dimension integrally signal and having fallen between behavior.
Summary of the invention
The object of the present invention is to provide a kind of effectively identification Falls Among Old People state, differentiation is fallen and non-motion state of falling, and can posture of falling be identified, the intelligent distinguishing for the injury order of severity of falling provides the Falls Among Old People state modeling based on Geometrical algebra, feature extraction and the recognition methods on basis.
Technical solution of the present invention is:
Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and a recognition methods, it is characterized in that:
(1) the triple channel acceleration signal adopting 3-axis acceleration sensor to gather, triple channel acceleration signal pairwise orthogonal;
(2) Clifford Geometrical algebra is adopted to carry out modeling to triple channel Falls Among Old People data: to utilize the Geometrical algebra be defined on three dimensions theoretical, choose 2-vector subspace collection { e wherein 12, e 23, e 31; characterize triple channel respectively to fall data direction; X, Y, Z triple channel acceleration signal data same sampling instant collected are as hypercomplex three imaginary parts; calculation process is carried out as an arrangement; make triple channel acceleration information of falling become a vector on Clifford Geometrical algebra territory, realize falling data modeling; Adopt Clifford Geometrical algebra subspace discrete cosine transform behavioural characteristic of carrying out falling to extract, take out and drop to signal low-frequency component sequence and to fall as triple channel the proper vector of data;
(3) to old man fall and non-behavior state of falling is added up, obtain the behavior classification number needing research, repeatedly acceleration information collection experiment is carried out to each behavior classification, identical 3-axis acceleration sensor is all adopted in experiment, and 3-axis acceleration sensor is placed in old man's waist same position in each experiment, the multi-group data collected, adopt the modeling method in step (2), the modeling of Clifford Geometrical algebra and feature extraction are carried out to the multi-group data collected, the proper vector of taking out is as the sample set of each behavior classification, each behavior classification all builds oneself sample set,
(4) sample set of an optional behavior classification, calculate wherein any two data samples of falling in Clifford Geometrical algebra distance spatially, obtain two sample points that distance is the shortest, and construct first Clifford Geometrical algebra territory single-degree-of-freedom neuron;
(5) state sample of falling concentrated the feature samples of falling be enclosed in first single-degree-of-freedom neuron overlay area to remove, obtain new training set 1; New training set finds the point that distance first single-degree-of-freedom neuron is nearest, and constructs second single-degree-of-freedom neuron; On new training set 1 basis, the sample of falling be enclosed in second single-degree-of-freedom neuron overlay area is removed, obtains new training set 2; According to second neuronic building method of single-degree-of-freedom, looping construct single-degree-of-freedom neuron, until when the set of data samples of falling finally obtained is empty set, stops single-degree-of-freedom neuron to build;
(6) union is asked to all single-degree-of-freedom neurons in each obtained behavior classification; For M the behavior classification existed, after training, then obtain the neuronic union of M single-degree-of-freedom altogether; For behavior of falling a classification sample to be identified, first the Clifford Geometrical algebra territory distance of this sample to each behavior classification single-degree-of-freedom neuron union is calculated, then for M behavior classification, M Clifford Geometrical algebra territory bee-line can be obtained; M bee-line is compared, obtains minimum Clifford Geometrical algebra territory distance wherein; Behavior classification then corresponding to this minor increment is the most close with behavior classification to be identified.
The Three-channel data characterizing Falls Among Old People behavior, in clifford Geometrical algebra territory, is become an entirety and processes, effectively remain the physical link between Three-channel data by the present invention, and contacting between triple channel Global Information and old man's behavior.Propose Falls Among Old People identification new method on Clifford Geometrical algebra territory.The method that the present invention proposes can effectively identify Falls Among Old People state, distinguishes and falls and non-motion state of falling, and can identify posture of falling, for the intelligent distinguishing of the injury order of severity of falling provides basis.The method is intelligent by force, discrimination is high, fast operation.Make a general survey of domestic existingly to fall and other behavior state recognition methods and systems, the method that the present invention carries there is no unit and realizes.
Accompanying drawing explanation
Below in conjunction with drawings and Examples, the invention will be further described.
Fig. 1, Fig. 2, Fig. 3, Fig. 4 be the old man that gathers respectively right, trip, lie down, data under layback state.Data acquisition 3-axis acceleration sensor collection in figure, horizontal ordinate is the sampling number of acquisition system, and ordinate is 3-axis acceleration sensor output voltage amplitude.The triple channel of 3-axis acceleration sensor exports and adopts different colours to mark in the drawings respectively.3-axis acceleration sensor is placed in old man's loins.The axle vertical with human body is defined as Z axis, and the axle parallel with human body pitch orientation is defined as X-axis, falls that direction is parallel is defined as Y-axis with people side.
Fig. 5 is the Falls Among Old People state recognition process flow diagram based on Clifford Geometrical algebra.
Fig. 6 Euclidean space single-degree-of-freedom neuron areas schematic diagram.
The covering schematic diagram of Fig. 7 Euclidean space single-degree-of-freedom neuron pool.
Fig. 8 trips sample to the neuronic bee-line schematic diagram of 4 class.
Fig. 9 is that layback sample is to the neuronic bee-line schematic diagram of 4 class.
Figure 10 sits down sample to the neuronic bee-line schematic diagram of 4 class.
Embodiment
Three axles are fallen acceleration information collection
3 passage acceleration transducer human body are adopted to fall the signal (sensor is placed in old man's waist) of process, sample frequency is 240Hz, in order to analyze all kinds of state of falling, experimental process simulation " trips ", " layback ", " right side is fallen ", " left side is fallen ", " directly lying down ", " jump ", " turn-taking ", " sitting down rapidly ", " sit quietly down to peace ", 14 kinds of states such as " sliding along wall ", " trotting ", " normally walking ", " face down is fallen ", " slowly walking ".To typically fall attitude and recurrently trip attitude, attitude is fallen on the right side, attitude of swinging back and attitude data of lying down (triple channel, X, Y and Z-axis direction) as shown in Figure 1, Figure 2, Figure 3, Figure 4.
Clearly, from data analysis of falling, every class falls the sample data of attitude differing greatly of falling and characterize instantaneously, in addition falling in the time period, three orthogonal axes the information of data space vector relation each other of falling in fact characterize human body at the landing position of moment of falling, for judging that the height of fall risk is very important information, therefore the Clifford Geometrical algebra that the present invention proposes to describe each interchannel vector relations carries out modeling and signature analysis to data of falling, Three-channel data is processed (realizing bulk treatment by being configured to supercomplex) as a whole under Clifford Geometrical algebra territory, both the physical link between triple channel had been contained, contain again contacting between triple channel Global Information and Falls Among Old People state.
The signature analysis of Clifford Geometrical algebra and recognition methods
Clifford Geometrical algebra basic framework
Clifford Geometrical algebra develops on Grasssmann algebraic foundation, and at scientific algorithm, the aspects such as graphical analysis have to be applied comparatively widely.Be defined in real vector space R non Clifford Geometrical algebra, be denoted as G (R n), or G n.On Clifford Geometrical algebra, an important concept is exactly geometry long-pending (GeometricProduct), for given one group of orthogonal basis { e 1, e 2..., e n∈ R n, suppose existence two vectors i, β i∈ R), then the geometry of its correspondence is long-pending is defined as
ab=a·b+a∧b(1)
Wherein ab is the inner product of vector; A ∧ b is the apposition of vector.
The apposition B be made up of r linearly independent vector is called r-blade, and r value is called corresponding grade (routine 2-blade), and wherein B is
B = ^ i = 1 r a i , a i ∈ R n - - - ( 2 )
Further, several r-blade(example 2-blade) linear combination be called r-vector(example 2-vector), obviously at G nthe r-vector collection of upper correspondence has element, this set is G non n-dimensional subspace n, uses symbol represent.Therefore whole G ncan be expressed by the linear combination of (3) formula.
G n = Σ i = 0 n G n i - - - ( 3 )
Wherein for G nthe subspace of upper corresponding dimension, represent G nupper corresponding dimension is the subspace of i, j, from G nin arbitrary element M ∈ G of deriving nbe called multivector, namely
M = &Sigma; r = 0 n < M > r - - - ( 4 )
Wherein <M> rbe called G non r-vector part, G nspace altogether dimension.
In addition about Outer Product of Vectors character and the G (R of the three-dimensional real space 3) on multivector as follows.
With G 3space is example, if vectorial a=α 1e 1+ α 2e 2+ α 3e 3, b=β 1e 1+ β 2e 2+ β 3e 3, c=γ 1e 1+ γ 2e 2+ γ 3e 3; According to character e j &CenterDot; e j = 0 , i &NotEqual; j 1 , i = j , e i ^ e j = 1 , i &NotEqual; j 0 , i = j , And remember e i∧ e j=e ij=-e j∧ e i=-e ji, i, j=1,2,3.Then the geometry of vectorial a, b, c amasss abc can to express an accepted way of doing sth as follows:
abc=(ab)c=ε 0·1+ε 1e 12e 23e 3
4e 125e 236e 317e 123
Wherein ε i∈ R, i ∈ [0,7]; The form of above formula is G 3on a multivector, { 1, e 1, e 2, e 3, e 12, e 23, e 31, e 123be G 3on one group of orthogonal basis, in addition k-at different levels vector and number as shown in table 1 below, be 2 altogether 3individual.
Table 1G 3upper k-vector at different levels and number
For multivector A ∈ G n( ), it can be expressed as
A = &Sigma; k = 0 n < A > k
Have in addition
( 3 ) , < A > n = &lambda; 12 &CenterDot; &CenterDot; &CenterDot; n n e 12 &CenterDot; &CenterDot; &CenterDot; n
The data characteristics of falling of Clifford geometry-DCT is extracted
The present invention proposes analysis and falls low-frequency component as data characteristics, and adopts the subspace discrete cosine transform of Clifford Geometrical algebra.The low-frequency component fetched data as feature, the basic general picture of reflected signal.
Because attitude of falling has very important reference value for analysis risk height, the packet of falling of system acquisition contains the process data of X, Y and Z-direction in the whole process of falling of human body.To on any sampled point, three axial data of process of falling can be expressed as a three-dimensional vector, are therefore mapped in spatially for one group of three axle sample sequence, and the change of its sequence vector can be characterized by human body and to fall the attitudes vibration of process.In order to disclose the attitude of falling of data of falling, first carrying out Clifford Geometrical algebra and carrying out three axles and to fall data modeling, making triple channel signal become bulk treatment.
At space G 3upper subspace { 1 a, e 1, e 2, e 3, e 12, e 23, e 31, e 123, arbitrary multivector M can be expressed as
Wherein 2-vector subspace collection { e 12, e 23, e 31,
Here e is defined 12direction be the X axis data direction of data of falling, e 23and e 31be respectively the Y-axis data of data of falling and the direction of Z-axis direction data.Therefore, given one group of triple channel is fallen data
s x = { x i } i = 1 n , - - - ( 6 - 1 )
s y = { y i } i = 1 n , - - - ( 6 - 2 )
s z = { z i } i = 1 n - - - ( 6 - 3 )
Wherein x i, y i, z ibe respectively the data of falling of X-axis, Y-axis, Z axis i-th sampled point sampling, x i, y i, z i∈ R.S x, s yand s zfor corresponding X, Y and Z axis are fallen data, x i, y i, z i∈ R, N are the length of signal.Can be expressed as
s c={c η=x η·e 12+y ηe 23+z ηe 31|η∈1,2,…,n}(7)
The data of falling of one group of three axle can be expressed by (7) formula entirety thus, the c in each sampling instant ibe a G (R 3) on multivector.Or (7) express by formula (8) equivalence, in (8), adopt hypercomplex form, i, j, k are hypercomplex three imaginary parts respectively, that is:
s c={c η=x η·i+y η·j+z η·k|η∈1,2,…,n}(8)
Wherein e 12=i, e 23=j, e 31=k; And have i 2=j 2=k 2=-1, ij=-ji=k, jk=-kj=i, ki=-ik=j.For the subspace discrete cosine transform of Clifford Geometrical algebra, point left and right conversion, wherein left conversion F l1, ω 2), right conversion F r1, ω 2) be respectively:
F L ( &omega; 1 , &omega; 2 ) = &delta; ( &omega; 1 ) &delta; ( &omega; 2 ) &Sigma; x = 0 m - 1 &Sigma; y = 0 n - 1 uf ( x , y ) cos [ &pi; ( 2 x + 1 ) &omega; 1 2 m ] cos [ &pi; ( 2 x + 1 ) &omega; 2 2 n ] - - - ( 9 )
F R ( &omega; 1 , &omega; 2 ) = &delta; ( &omega; 1 ) &delta; ( &omega; 2 ) &Sigma; x = 0 m - 1 &Sigma; y = 0 n - 1 f ( x , y ) cos [ &pi; ( 2 x + 1 ) &omega; 1 2 m ] cos [ &pi; ( 2 x + 1 ) &omega; 2 2 n ] u - - - ( 10 )
Wherein u ∈ G (R 3), u 2=-1; F (x, y) is figure signal, when getting x or y and being 1, is then one-dimensional signal, &delta; ( &omega; 1 ) = 1 m &omega; 1 = 0 1 n &omega; 2 &NotEqual; 0 , &delta; ( &omega; 2 ) = 1 m &omega; 2 = 0 1 n &omega; 2 &NotEqual; 0
Especially, be a kind of special circumstances of 2D signal for one-dimensional signal, its computing method are similar.
For the result after conversion, such as formula (11)
F d=[h 1,h 2,…,h n],d∈{R,L},h i∈G(R 0,3)(11)
Get medium and low frequency components series [h 1, h 2..., h t] (t>1) fall as triple channel the proper vector of data.
The training of Clifford Geometrical algebra single-degree-of-freedom neuron and recognition methods
The recognizer covering thought based on single-degree-of-freedom has original advantage in small sample state, and its most basic basic point is that it meets process and the principle that people is familiar with things.What it may be noted that here desalinates the concept of neuronic weights in the present invention, only emphasizes its geometrical property, and the present invention gets single-degree-of-freedom neuron and carries out training and identify.On the single-degree-of-freedom recognizer basis of the present invention in existing theorem in Euclid space, propose the single-degree-of-freedom neuron building method on Clifford Geometrical algebra territory and recognition methods of falling.
Single-degree-of-freedom neuron on Euclidean space
Below the single-degree-of-freedom neuron on Euclidean space is simply introduced.If tie up on Euclidean space, by 2 p at n 1, p 2∈ R nthe neuronic area of space of the single-degree-of-freedom determined can be expressed by (12) formula, namely
Ω n={f|ρ(f,s n)≤σ,σ∈R}(12)
Wherein s n=η p 1+ (1-η) p 2, and 0≤η≤1, ρ (f, s n) be line segment outer some f is to the shortest space length of this line segment, and σ is distance threshold, clearly, and the distance of which definition
&rho; 2 ( f , s n ) = | | f - f p | | 2 , | | f p - p 1 | | + | | f p - p 2 | | = | | p 1 - p 2 | | | | f - p 1 | | 2 , | | f p - p 1 | | + | | p 1 - p 2 | | = | | f p - p 2 | | | | f - p 2 | | 2 , otherwise - - - ( 13 )
Wherein
f p = Pr ( f , l p 1 p 2 ) - - - ( 14 )
Here f pnamely f is to straight line projection vector point.
Clifford Geometrical algebra single-degree-of-freedom neuron spatially
Due to the Falls Among Old People sample characteristics that the present invention extracts its characteristic component therefore need spatially set up single-degree-of-freedom neuron models and correspondence recognizer (for being different from point on Euclidean space or vector, if the present invention subsequent to point or vector without all acquiescences illustrated in space on).
Choose any two vectors of falling distance g (y 1, y 2) measure definitions is
g 2 ( y 1 , y 2 ) = &Sigma; i = 0,2 | | < y 1 > i - < y 2 > i | | 2 - - - ( 15 )
Wherein vector for vectorial y j(j=1,2) are corresponding to space on part; Norm and have g 2(y 1, y 2)=g 2(y 2, y 1).Therefore in space on, by 2 points the single-degree-of-freedom neuron Ω determined cthe overlay area of definition is
Ω C={f|ρ c(f,s c)≤σ,σ∈R}(16)
Wherein s c=η p 1+ (1-η) p 2represent by point the line segment determined; σ is distance threshold.Be similar to (12) formula, spatially, the some f outside line segment cto line segment bee-line be ρ c(f, s c).Or be expressed as
Ω C={f|ω c(f,p 1,p 2)≤σ,σ∈R}(17)
Suppose to there is M class 3-axis acceleration data sample, each class correspond to different modes of falling (fall as leaned forward, swing back and fall), or lie down, the non-mode of falling such as to sit down.Every class has N number of sample (the same mode of falling has carried out N secondary data collection experiment, obtains N number of sample), each sample then every class forms a sample set A 0={ y 1, y 2..., y nany two then the neuronic algorithm of structure single-degree-of-freedom is as follows:
Step 1: calculate every class and fall or the non-behavior sample collection A that falls 0the data sample y that falls on upper any two Clifford Geometrical algebra territories j, y k? distance spatially g ( y j , y k ) = &Sigma; i = 0,2 | | < y j > i - < y k > i | | 2 ; Put (r 1, r 2)=mp ({ g jk| j, k ∈ N, j ≠ k}),
(r 1,r 2∈N,r 1≠r 2)(18)
Subscript (the r of the shortest 2 of distance in matrix A that what wherein function mp returned is 1, r 2), namely and construct first single-degree-of-freedom neuron Ω c1, namely
Ω C1={f|ω c(f,y r1,y r2)≤σ,σ∈R}(19)
Wherein threshold value σ meaning is the same, and state sample of falling is concentrated and has been enclosed in single-degree-of-freedom neuron Ω c1sample of falling in overlay area is removed, and obtains new training set A 1=A 0c1, to point again v is designated as 1, v 2.
Step 2: at A 1on find distance Ω c1nearest point, is denoted as v again 3∈ A 1, namely
g 2 ( v 2 , v 3 ) = min y j &Element; A 1 { &Sigma; i = 0,2 | | < y j > i - < v 2 > i | | 2 } , - - - ( 20 )
And construct second single-degree-of-freedom neuron Ω c2, it is in space upper covering domain of sending out is
Ω C2={f|ω c(f,v 2,v 3)≤σ,σ∈R}(21)
Get back another new point set A 2=A 1c2.
Step 3: by the method for step 2, at point set A q=A q-1qon find distance Ω qnearest point, is denoted as v q+2∈ A q, then construct the neuron Ω of q single-degree-of-freedom c (q+1), its space covering domain is
Ω C(q+1)={f|ω c(f,v q,v q+1)≤σ,σ∈R}(22)
Obtain new point set A again q+1=A qc (q+1).
Step 4: continue step 3, the data point set of falling to the last obtained when being empty set, algorithm terminates.
After having trained, the neuron Ω of W single-degree-of-freedom altogether cr(r=1,2,3 ... W), they are in space on union N sdbe expressed as
N sd = &cup; r = 1 W &Omega; Cr - - - ( 23 )
Subsequently to discuss again based on the neuronic recognizer of single-degree-of-freedom.For the M class old man behavior sample (falling or other non-behaviors of falling) existed, after training, then obtain the also neuronic of M class single-degree-of-freedom altogether (i=1,2 ..., M).For a Falls Among Old People behavior sample x to be identified 0, first calculate x 0to every class (i=1,2 ..., M) distance and be defined as
d ( x 0 , N sd i ) = min j = 1 W i { g ( x 0 , &Omega; C ( i , j ) ) } - - - ( 24 )
Wherein belong to a jth single-degree-of-freedom neuron of the i-th class sample, W iit is the i-th class sample single-degree-of-freedom neuron number.Known by (22), for M class (i=1,2 ..., M), calculate M bee-line altogether (i=1,2 ..., M).
Specify herein: distance old man behavior sample x 0the generic of that nearest class sample is x 0generic, namely
k = arg min i = 1 M { d ( x 0 , N sd i ) } , ( k = 1,2 , &CenterDot; &CenterDot; &CenterDot; , M ) - - - ( 25 )
Argmin expression obtains the function of minimum value.
Experimental result and analysis
For verifying the validity of recognizer, herein to typically, and there is the attitude and carried out Feature extraction and recognition research with some common physical activity attitudes of falling of high risk, comprising " normally walking ", " trip ", " sitting down ", " layback is fallen ", the 5 class attitudes such as " to lie down ", wherein " normally walk " and belong to special attitude of falling with " lying down ".Every class attitude of falling gets 10 samples as training sample, then gets other 10 samples as independent test sample, be then divided into 50 training samples and 50 independent test samples.Carry out feature extraction as stated above, because the signal value of image data is comparatively large, in order to avoid making troubles to computing, characteristic signal carries out regular, jointly reduces 1000 times, and gets t=60, and namely proper vector length is 60.
Experimental result shows, when getting threshold value σ=200, have 3 samples to be known by mistake, wherein 1 " tripping " attitude is known for being sat down by mistake, 1 " layback " attitude is for knowing for " sitting down " by mistake, the attitude of 1 " sitting down " is for knowing for " tripping " by mistake, and " normally walking " and " lying down " two attitudes are all correctly validated, and overall correct recognition rata is 94.0%.
Further analysis, according to recognizer in this paper and above-mentioned selected threshold value, divided by three samples known by mistake and be clipped to wherein " tripping ", " layback ", " sitting down ", the neuronic bee-line of " lying down " 4 class is as shown in Fig. 8, Fig. 9, Figure 10.

Claims (1)

1., based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and a recognition methods, it is characterized in that:
(1) the triple channel acceleration signal adopting 3-axis acceleration sensor to gather, triple channel acceleration signal pairwise orthogonal;
(2) Clifford Geometrical algebra is adopted to carry out modeling to triple channel Falls Among Old People data: to utilize the Geometrical algebra be defined on three dimensions theoretical, choose 2-vector subspace collection { e wherein 12, e 23, e 31; characterize triple channel respectively to fall data direction; X, Y, Z triple channel acceleration signal data same sampling instant collected are as hypercomplex three imaginary parts; integrally carry out calculation process; make triple channel acceleration signal data become a vector on Clifford Geometrical algebra territory, realize falling data modeling; Adopt Clifford Geometrical algebra subspace discrete cosine transform behavioural characteristic of carrying out falling to extract, take out signal low-frequency component sequence of falling and to fall as triple channel the proper vector of data;
(3) to old man fall and non-behavior state of falling is added up, obtain the behavior classification number needing research, repeatedly acceleration information collection experiment is carried out to each behavior classification, identical 3-axis acceleration sensor is all adopted in experiment, and 3-axis acceleration sensor is placed in old man's waist same position in each experiment, the multi-group data collected, adopt the modeling method in step (2), the modeling of Clifford Geometrical algebra and feature extraction are carried out to the multi-group data collected, the proper vector of taking out is as the sample set of each behavior classification, each behavior classification all builds oneself sample set,
(4) sample set of an optional behavior classification, calculate wherein any two data samples of falling in Clifford Geometrical algebra distance spatially, obtain two sample points that distance is the shortest, and construct first Clifford Geometrical algebra territory single-degree-of-freedom neuron;
(5) state sample of falling concentrated the feature samples of falling be enclosed in first single-degree-of-freedom neuron overlay area to remove, obtain new training set 1; New training set finds the point that distance first single-degree-of-freedom neuron is nearest, and constructs second single-degree-of-freedom neuron; On new training set 1 basis, the feature samples of falling be enclosed in second single-degree-of-freedom neuron overlay area is removed, obtains new training set 2; According to second neuronic building method of single-degree-of-freedom, looping construct single-degree-of-freedom neuron, until when the state sample of falling finally obtained integrates as empty set, stops single-degree-of-freedom neuron to build;
(6) union is asked to all single-degree-of-freedom neurons in each obtained behavior classification; For M the behavior classification existed, after training, then obtain the neuronic union of M single-degree-of-freedom altogether; For behavior of falling a classification sample to be identified, first the Clifford Geometrical algebra territory distance of this sample to each behavior classification single-degree-of-freedom neuron union is calculated, then for M behavior classification, M Clifford Geometrical algebra territory bee-line can be obtained; M bee-line is compared, obtains minimum Clifford Geometrical algebra territory distance wherein; Behavior classification then corresponding to this minor increment is the most close with behavior classification to be identified.
CN201310066132.XA 2013-03-01 2013-03-01 Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods Active CN103116745B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310066132.XA CN103116745B (en) 2013-03-01 2013-03-01 Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310066132.XA CN103116745B (en) 2013-03-01 2013-03-01 Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods

Publications (2)

Publication Number Publication Date
CN103116745A CN103116745A (en) 2013-05-22
CN103116745B true CN103116745B (en) 2016-01-20

Family

ID=48415117

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310066132.XA Active CN103116745B (en) 2013-03-01 2013-03-01 Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods

Country Status (1)

Country Link
CN (1) CN103116745B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104899594B (en) * 2014-03-06 2018-01-02 中国科学院沈阳自动化研究所 Hand motion recognition method based on Decomposition Surface EMG
CN108830299A (en) * 2018-05-21 2018-11-16 千寻位置网络有限公司 The recognition methods and system, smart machine of smart machine wearing regime based on SVM
WO2020150990A1 (en) * 2019-01-25 2020-07-30 深圳大学 Human body enactive representation processing method and system, and electronic apparatus and storage medium
CN109974645B (en) * 2019-04-10 2021-09-10 成都安尔法智控科技有限公司 Sensor abnormal displacement monitoring method and system
CN110781808A (en) * 2019-10-23 2020-02-11 泰康保险集团股份有限公司 Fall detection method, device, equipment and storage medium
CN111024820B (en) * 2019-12-19 2023-01-31 南通大学 Health monitoring system for offshore wind power blade and data processing method thereof
CN112468956B (en) * 2020-11-12 2022-10-11 西安邮电大学 Human activity monitoring method for indoor positioning and motion state
CN114849162B (en) * 2022-04-13 2023-03-24 江苏医药职业学院 Monitoring system of treadmill for body building and rehabilitation of old people and treadmill

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1462990A2 (en) * 2003-03-28 2004-09-29 Matsushita Electric Industrial Co., Ltd. Apparatus and method for processing an image
CN101711401A (en) * 2007-04-19 2010-05-19 皇家飞利浦电子股份有限公司 Fall detection system
CN201853320U (en) * 2010-11-23 2011-06-01 南通大学 Monitoring and alarm system for the old

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1462990A2 (en) * 2003-03-28 2004-09-29 Matsushita Electric Industrial Co., Ltd. Apparatus and method for processing an image
CN101711401A (en) * 2007-04-19 2010-05-19 皇家飞利浦电子股份有限公司 Fall detection system
CN201853320U (en) * 2010-11-23 2011-06-01 南通大学 Monitoring and alarm system for the old

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
双权值神经网络结构性逼近理论的研究及其应用;王洪昌;《中国优秀博硕士学位论文全文数据库》;20061215;全文 *
基于三轴加速度传感器的人体跌倒检测系统设计与实现;王荣等;《计算机应用》;20120501;第32卷(第5期);全文 *
基于几何代数的矢量时空数据表达与建模方法;俞肇元等;《地球信息科学学报》;20120215;第14卷(第1期);全文 *

Also Published As

Publication number Publication date
CN103116745A (en) 2013-05-22

Similar Documents

Publication Publication Date Title
CN103116745B (en) Based on the Falls Among Old People state modeling of Geometrical algebra, feature extraction and recognition methods
US9679195B2 (en) Systems and methods for facial property identification
JP5848551B2 (en) Learning device, learning device control method, detection device, detection device control method, and program
US10231651B2 (en) Gait authentication system and method thereof
CN103605952B (en) Based on the Human bodys&#39; response method that Laplce&#39;s canonical group is sparse
CN105023022A (en) Tumble detection method and system
JP2010207488A (en) Behavior analyzing device and program
CN105512620A (en) Convolutional neural network training method and apparatus for face recognition
CN103593672A (en) Adaboost classifier on-line learning method and Adaboost classifier on-line learning system
CN102542295A (en) Method for detecting landslip from remotely sensed image by adopting image classification technology
CN110084286A (en) A kind of human motion recognition method of sensor-based ECOC technology
CN103500342A (en) Human behavior recognition method based on accelerometer
CN105354555A (en) Probabilistic graphical model-based three-dimensional face recognition method
CN102521574A (en) Human action identification method based on pyroelectric infrared information
CN111259956B (en) Inertial sensor-based method for rapidly identifying unconventional behaviors of people
Ning et al. Fall detection algorithm based on gradient boosting decision tree
CN109086667A (en) Similar active recognition methods based on intelligent terminal
CN102831445B (en) Target detection method based on semantic Hough transformation and partial least squares
Ashwini et al. Skeletal data based activity recognition system
CN203149599U (en) Geometric algebra-based identification system for falling states of old people
Wongpatikaseree et al. Range-based algorithm for posture classification and fall-down detection in smart homecare system
Jarchi et al. Transition detection and activity classification from wearable sensors using singular spectrum analysis
CN106355142A (en) A Method and Device for Recognizing Human Falling State
CN107122767A (en) A kind of physical activity end-point detecting method based on comentropy
CN115393956A (en) CNN-BilSTM fall detection method for improving attention mechanism

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant