CN106446778A - Method for identifying human motions based on accelerometer - Google Patents

Method for identifying human motions based on accelerometer Download PDF

Info

Publication number
CN106446778A
CN106446778A CN201610751821.8A CN201610751821A CN106446778A CN 106446778 A CN106446778 A CN 106446778A CN 201610751821 A CN201610751821 A CN 201610751821A CN 106446778 A CN106446778 A CN 106446778A
Authority
CN
China
Prior art keywords
feature
graph
dimension
learning
graphs
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.)
Pending
Application number
CN201610751821.8A
Other languages
Chinese (zh)
Inventor
褚晶辉
罗薇
吕卫
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.)
Tianjin University
Original Assignee
Tianjin 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 Tianjin University filed Critical Tianjin University
Priority to CN201610751821.8A priority Critical patent/CN106446778A/en
Publication of CN106446778A publication Critical patent/CN106446778A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24143—Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Signal Processing (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a method for identifying human motions based on an accelerometer. The method includes the following steps: conducting operable window processing on an original acceleration signal, extracting the frequency domain characteristics of the acceleration signal; based on the relevance of characteristics in each dimension in a characteristics space, generating an incidence matrix of the entire characteristic space; conducting spectral clustering on the incidence matrix, generating a plurality of subsets of which each is represented in the form of a graph, until which completing the part of multi-graph represented learning; analyzing each acquired graph in multi-dimensions, removing redundant spectrum information of each graph so as to acquire most relevant information in graphs and achieving local dimension reduction; applying multi-set canonical correlation analysis to the plurality of graphs that are subject to locality dimension reduction, obtaining the final representation form of the characteristics by computing the relevance of main information among graphs, and completing multi-graph encapsulated learning; conducting category learning through a nearest neighbor classifier. According to the invention, the method can reduce dimension of characteristics and increase accuracy of motion identification.

Description

Human motion identification method based on acceleration transducer
Technical field
The invention belongs to area of pattern recognition, particularly to a kind of human motion identification method.
Background technology
Human motion pattern recognition is an interesting and challenging problem, it be widely used in health care, The aspect such as old man's nurse and context-aware computing.Basic step is related to collecting sensor signal, and signal processing and pattern are divided Class.According to the difference of sensing electronics and algorithm, the method for human motion pattern recognition can be divided into two classes：1) it is based on and calculate The learning method of machine vision；2) method based on acceleration transducer.
Mans motion simulation is had been widely used for based on the technology of computer vision, this technology needs in monitoring place peace Fill single or multiple video cameras, realize the identification of human motion by object detection, Object Segmentation, feature extraction and classification.But It is that technology based on computer vision requires strictly to illumination condition, and data volume is easily caused greatly dimension disaster.It is based on and add The human motion identification method of velocity sensor, by being put on acceleration transducer, can carry out not under field conditions (factors) The noticeable motion detection with Noninvasive.This technology is mainly by carrying out pretreatment, feature extraction to acceleration signal Realize the identification of human motion with selection, feature analysiss and classification.Although compared with the technology based on computer vision, based on plus The data volume of the human motion identification method of velocity sensor is less, but is as the increase of the feature of extraction, still may make Become dimension disaster, the accuracy of impact identification.
The numerous studies of the human motion identification based on acceleration transducer are all many according to placing in body different parts Individual sensor is carrying out Motion Recognition, but multiple sensor collection continuous data is not only unrealistic but also tester can be made to feel not Comfortable, so, the research of the human motion identification based on single acceleration transducer receives bigger concern.Based on acceleration Sensor distinguish human body motor pattern a kind of be analyzed by the time domain of acceleration signal, frequency domain and wavelet character and Judge.The temporal signatures of wherein acceleration signal mainly include average, variance, the correlation coefficient of two between centers and energy etc.；Frequency domain Feature is first to carry out fast Fourier transform to acceleration signal, then extracts the feature such as Fourier coefficient or frequency domain entropy；Small echo Feature, by acceleration signal is carried out with wavelet transformation, extracts the features such as wavelet energy, wavelet energy distribution.Another kind is to pass through Directly study obtains the character representation of data-driven, is the effective ways of mining data collection potential structure based on the algorithm of figure.
Content of the invention
The present invention provides a kind of human motion recognizer based on acceleration transducer, can sufficiently utilize in feature Potential useful information, removes the redundancy in feature, had both reduced the dimension of feature, improves the accuracy rate of Motion Recognition again. Technical scheme is as follows：
A kind of human motion identification method based on acceleration transducer, comprises the following steps：
(1) determine suitable sliding window size, active window process is carried out to original acceleration signal, accelerate after extraction process The frequency domain character of degree signal；
(2) Hilbert-Schmidt independent criteria is used to this feature, according to the pass between dimensional feature every in feature space Connection property, generates the incidence matrix of whole feature space；
(3) method using spectral clustering to this incidence matrix, generates multiple subsets, each subset is expressed as figure, until this In complete many charts dendrography habit part；
(4) to each the figure application Multidimensional Scaling method obtaining, remove the unnecessary spectrum information of each figure, to obtain figure Interior maximally related information, realizes local dimensionality reduction；
(5) with many, canonical correlation analysis are collected to the many figures after the dimensionality reduction of local, by calculating the phase of main information between figure Guan Xing, obtains the final representation of feature, completes the study scheming embedded expression more；
(6) classification learning is carried out by nearest neighbor classifier.
The present invention proposes a kind of new method on the basis of obtaining acceleration signal, and the core of the method is many charts Dendrography is practised and the embedded study of many figures.The first step is to realize many charts dendrography using the method for analysis of spectrum to practise；Second step passes through local Dimensionality reduction and the overall situation merge the embedded study realizing many figures.Wherein, local dimensionality reduction refers to remove the unnecessary spectral information of each figure, with Obtain maximally related information in figure；The overall situation merges the dependency referring to main information between calculating figure, final to obtain feature Representation.The method can effectively reduce the dimension of feature, and improves the knowledge of the human motion based on acceleration transducer Not rate.
Brief description
The broad flow diagram of the human motion identification method based on acceleration transducer that Fig. 1 provides for the present invention.
Specific embodiment
In order to more clearly describe the present invention, below embodiment of the present invention is described in further detail.
Classified based on the human motion of acceleration transducer to easily facilitate, improved based on acceleration transducer The accuracy rate of human motion identification, the invention provides the human motion identification method based on acceleration transducer, method and step As follows：
(1) sliding window is adopted to process on tri- axles of x, y and z of current original acceleration signal respectively, each window comprises 512 sample points, Duplication is 50%；Then respectively FFT is carried out to each sliding window, because first coefficient represents directly Flow component, then take front 64 coefficients and remove first for FFT feature, that is, FFT is characterized as 63 dimensions.By each action in each window Three axle FFT coefficients in mouthful couple together, then in each window, FFT is characterized as 63 × 3-dimensional, obtain every group of final FFT of action special Levy and tie up for 945, be designated as X；
(2) determine the relation that in current feature space X, x dimension is tieed up with y, by Hilbert-Schmidt independent criteria Original feature space X is expressed as associated diagram by (Hilbert-Schmidt Independence Criterion, HSIC).For X dimension and y dimension, define Q={ (x1,y1),…,(xm,ym), and Q has m sample.By regenerating in Hilbert space Two kernel function k (x meeting integrabilityi,xj)=<φ(xi),φ(xj)>With l (yi,yj)=<ψ(yi),ψ(yj)>, Ke Yixiang That answers calculates matrix K and L, and in δijCan obtain the matrix H of incidence matrix K and L under=1e-2, wherein calculating matrix K, The formula of L and H is：
Kij=k (xi,xj) (1)
Lij=l (yi,yj) (2)
Hij=δij-m-1(3)
(3) represent to simplify nuclear matrix with contraction.OrderThen calculate x dimension and associating that y ties up The l of propertycFormula lcAnd its approximate evaluation is：
lc(x, y)=HSIC (Q)=(m-1)2tr(KHLH) (4)
(4) l is calculated to each pair feature x in the initial characteristic data collection X having m sample and yc, obtain initial association Figure A (Aij=lc(xi,xj))；
(5) with Spectral Clustering, original feature space is divided into n separate character subset on AFor the n separate character subset generating, each subset has m sample, can be expressed as m The figure on individual summit, so can generate altogether n figure.
(6) for each figure, adjacency matrix D ∈ Rm×m, for the s on i-th and j-th summit representing in figureiAnd sj, σ2 It is zooming parameter, and in the case of σ=0.01, can be as follows with adjacency matrix D：
(7), after obtaining n figure, to each the figure application Multidimensional Scaling method obtaining, remove each figure unnecessary Spectrum information, reaches the purpose of local dimensionality reduction.By Multidimensional Scaling is used on adjacency matrix D to each figure, to each figure Carry out dimensionality reduction, obtain the principal coordinate of each figure simultaneously.Detailed process is as follows：
The first step, according to the adjacency matrix D obtaining, by calculating the average similarity of the i-th row and jthWithAnd It is the average similarity of all row and columnsCarry out structural matrix T.The public affairs being averaging similarity and matrix T adopting in this step Formula is as follows：
Second step, carries out characteristic vector analysis to matrix T, obtains r characteristic vector e1,e2,…,erWith r non-zero characteristics Value, this r eigenvalue is λ according to descending1≥λ2≥…≥λr＞ 0.Matrix E is made up of as row characteristic vector, as [e1…er]；Diagonal matrix Λ is made up of the r eigenvalue corresponding to matrix E, then the characteristic vector of t-th sample is expressed asThe principal coordinate obtaining i-th figure is defined as：
(8), after obtaining the principal coordinate of each figure, the Cross-covariance between principal coordinate and projection vector ω can be obtained, As principal coordinate GiWith GjBetween covariance matrix and principal coordinate GiAuto-covariance matrix be：
Cij=E [Gi(Gj)T] (9)
Cii=E [Gi(Gi)T] (10)
(9) with many, canonical correlation analysis are collected to the many figures after the dimensionality reduction of local, by calculating the phase of main information between figure Guan Xing, obtains the final representation G of feature, completes the study scheming embedded expression more.Many collection allusion quotation employed in this step Type formula is as follows：
(10) pass through to solve above formula, then k (k≤min (r of i-th figure1..., rn)) canonical variable of collection more than rank isSo final characteristic of division is expressed asSo pass through many collection allusion quotations N figure is associated measuring and the final character representation of embedded formation by type correlation analysiss.
(11) finally by nearest neighbor classifier, classification learning is carried out to the final representation G of feature, show that motion is known Other result.
Below the effectiveness of method provided by the present invention is verified with a specific experiment, described below：
Experiment uses SCUT-NAA data base, and this data base is the mankind disclosed in first based on three-dimensional acceleration Behavior database.By placing 3-axis acceleration sensor ADXL 330 in belt, coat pocket and three fixations of trouser pocket , there are 44 data pickers position, and sample frequency is 100Hz, acquires 1278 samples altogether.This 44 data pickers are by 34 Name male and 10 women compositions, mean age and variance are 21.2 years old and 0.7 years old respectively.This data base acquires 10 classes and moves Make, the light intensity action such as sit quietly, remain where one is and wait middle intensity kinesis, and jump and the high intensity such as race motion, so SCUT- NAA data base is suitable for the research of this paper.In the present invention, select to be collected in the acceleration information of waist, arbitrarily choose 44 samples , as training set, remaining sample is as test set for this.Wherein, because only that 30 pickers provide number by bike According to, so experimental data by bike uses existing all data, and arbitrarily take 30 samples as training set, Remaining sample is as test set.Under default situations, the quantity of setting subset is 2 (n=2), and arranges how collection canonical correlations divide The exponent number of analysis is 19 (k=19).
Experimental result is as follows：Principal component analysiss (Principal Component Analysis, PCA), linear discriminent Analysis (Linear Discriminant Analysis, LDA), local retaining projection (Locality Preserving Projection, LPP) and have local retaining projection (the Supervised Locality Preserving of supervision Projection, SLPP) it is conventional dimension reduction method, framework can be embedded with linear graph and they are united, and can use The structure of figure and figure embedded being explained.Local sensitivity discriminant analysiss (Locality Sensitive Discriminant Analysis, LSDA) it is the effective supervision algorithm that can find data set local geometric characteristic.Interval Fisher analyzes (Marginal Fisher ' s Analysis, MFA)[25]It is to be calculated based on the new manifold learning of the embedded framework of figure Method, this algorithm describes the diversity between the similarity in class and class by two figures.Scheme proposed by the invention is obvious Better than other dimension reduction methods, dimension can be reduced, improve discrimination again.

Claims (1)

1. a kind of human motion identification method based on acceleration transducer, comprises the following steps：
(1) determine suitable sliding window size, original acceleration signal is carried out with active window process, extraction process post-acceleration is believed Number frequency domain character；
(2) Hilbert-Schmidt independent criteria is used to this feature, according to the association between dimensional feature every in feature space Property, generate the incidence matrix of whole feature space；
(3) method using spectral clustering to this incidence matrix, generates multiple subsets, each subset is expressed as figure, complete until here Become many charts dendrography habit part；
(4) to each the figure application Multidimensional Scaling method obtaining, remove the unnecessary spectrum information of each figure, to obtain in figure Related information, realizes local dimensionality reduction；
(5) with many, canonical correlation analysis are collected to the many figures after the dimensionality reduction of local, by calculating the dependency of main information between figure, Obtain the final representation of feature, complete the study scheming embedded expression more；
(6) classification learning is carried out by nearest neighbor classifier.
CN201610751821.8A 2016-08-27 2016-08-27 Method for identifying human motions based on accelerometer Pending CN106446778A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610751821.8A CN106446778A (en) 2016-08-27 2016-08-27 Method for identifying human motions based on accelerometer

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610751821.8A CN106446778A (en) 2016-08-27 2016-08-27 Method for identifying human motions based on accelerometer

Publications (1)

Publication Number Publication Date
CN106446778A true CN106446778A (en) 2017-02-22

Family

ID=58182726

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610751821.8A Pending CN106446778A (en) 2016-08-27 2016-08-27 Method for identifying human motions based on accelerometer

Country Status (1)

Country Link
CN (1) CN106446778A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111199216A (en) * 2020-01-07 2020-05-26 上海交通大学 Motion prediction method and system for human skeleton
CN115015390A (en) * 2022-06-08 2022-09-06 华侨大学 MWTLMDS-based curtain wall working modal parameter identification method and system

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101034441A (en) * 2007-03-29 2007-09-12 浙江大学 Human motion date recognizing method based on integrated Hidden Markov model leaning method
CN102521843A (en) * 2011-11-28 2012-06-27 大连大学 Three-dimensional human body motion analysis and synthesis method based on manifold learning
CN103886323A (en) * 2013-09-24 2014-06-25 清华大学 Behavior identification method based on mobile terminal and mobile terminal
CN104063511A (en) * 2014-07-09 2014-09-24 哈尔滨工业大学 Complex system supervision graph embedding structural data visualized monitoring method based on relevant measurement
CN104166981A (en) * 2014-06-17 2014-11-26 南京信息工程大学 Human body movement learning method based on multigraph expression
CN104899604A (en) * 2015-06-08 2015-09-09 哈尔滨工程大学 Feature-level fusion method based on data set merging
CN105786189A (en) * 2016-04-28 2016-07-20 深圳大学 Finger independent action recognition method and system based on MMG signal
CN105893947A (en) * 2016-03-29 2016-08-24 江南大学 Bi-visual-angle face identification method based on multi-local correlation characteristic learning

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101034441A (en) * 2007-03-29 2007-09-12 浙江大学 Human motion date recognizing method based on integrated Hidden Markov model leaning method
CN102521843A (en) * 2011-11-28 2012-06-27 大连大学 Three-dimensional human body motion analysis and synthesis method based on manifold learning
CN103886323A (en) * 2013-09-24 2014-06-25 清华大学 Behavior identification method based on mobile terminal and mobile terminal
CN104166981A (en) * 2014-06-17 2014-11-26 南京信息工程大学 Human body movement learning method based on multigraph expression
CN104063511A (en) * 2014-07-09 2014-09-24 哈尔滨工业大学 Complex system supervision graph embedding structural data visualized monitoring method based on relevant measurement
CN104899604A (en) * 2015-06-08 2015-09-09 哈尔滨工程大学 Feature-level fusion method based on data set merging
CN105893947A (en) * 2016-03-29 2016-08-24 江南大学 Bi-visual-angle face identification method based on multi-local correlation characteristic learning
CN105786189A (en) * 2016-04-28 2016-07-20 深圳大学 Finger independent action recognition method and system based on MMG signal

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
BORG I等: ""Modern Multidimensional Scaling:Theory and Applications"", 《JOURNAL OF EDUCATIONAL MEASUREMENT》 *
NIELSEN A A等: ""Multiset canonical correlations analysis and multispectral,truly multitemporal remote sensing data"", 《IEEE TRANSACTIONS ON IMAGE PROCESSING A PUBLICATION OF THE IEEE SIGNAL PROCESSING SOCIETY》 *
王磊等: ""基于多集典型相关分析的雷达辐射源指纹识别"", 《西安电子科技大学学报（自然科学版）》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111199216A (en) * 2020-01-07 2020-05-26 上海交通大学 Motion prediction method and system for human skeleton
CN111199216B (en) * 2020-01-07 2022-10-28 上海交通大学 Motion prediction method and system for human skeleton
CN115015390A (en) * 2022-06-08 2022-09-06 华侨大学 MWTLMDS-based curtain wall working modal parameter identification method and system

Similar Documents

Publication Publication Date Title
Lu et al. Deep learning for fall detection: Three-dimensional CNN combined with LSTM on video kinematic data
Singh et al. Transforming sensor data to the image domain for deep learning—An application to footstep detection
Tran et al. A multi-modal multi-view dataset for human fall analysis and preliminary investigation on modality
Kwolek et al. Calibrated and synchronized multi-view video and motion capture dataset for evaluation of gait recognition
Cuzzolin et al. Metric learning for Parkinsonian identification from IMU gait measurements
Le et al. Robust hand detection in vehicles
CN103886341A (en) Gait behavior recognition method based on feature combination
Zaki et al. Using automated walking gait analysis for the identification of pedestrian attributes
CN110084211B (en) Action recognition method
CN111009321A (en) Application method of machine learning classification model in juvenile autism auxiliary diagnosis
Kobayashi et al. Three-way auto-correlation approach to motion recognition
Zhang et al. A deep learning-based approach for gait analysis in Huntington disease
Zheng et al. Evaluation framework on translation-invariant representation for cumulative foot pressure image
Bu Human motion gesture recognition algorithm in video based on convolutional neural features of training images
CN106529377A (en) Age estimating method, age estimating device and age estimating system based on image
Mastorakis et al. Fall detection without people: A simulation approach tackling video data scarcity
Ortega-Anderez et al. A multi-level refinement approach towards the classification of quotidian activities using accelerometer data
Minija et al. Food image classification using sphere shaped—Support vector machine
Lu et al. Visual guided deep learning scheme for fall detection
Parameswari et al. Human activity recognition using SVM and deep learning
Ge et al. Local global relational network for facial action units recognition
Li et al. Dynamic long short-term memory network for skeleton-based gait recognition
CN107045624A (en) A kind of EEG signals pretreatment rolled into a ball based on maximum weighted and sorting technique
CN106446778A (en) Method for identifying human motions based on accelerometer
Javeed et al. Deep activity recognition based on patterns discovery for healthcare monitoring

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20170222