CN103400123A - Gait type identification method based on three-axis acceleration sensor and neural network - Google Patents
Gait type identification method based on three-axis acceleration sensor and neural network Download PDFInfo
- Publication number
- CN103400123A CN103400123A CN2013103673511A CN201310367351A CN103400123A CN 103400123 A CN103400123 A CN 103400123A CN 2013103673511 A CN2013103673511 A CN 2013103673511A CN 201310367351 A CN201310367351 A CN 201310367351A CN 103400123 A CN103400123 A CN 103400123A
- Authority
- CN
- China
- Prior art keywords
- gait
- neural network
- acceleration
- sample
- data
- 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
Links
Images
Landscapes
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a gait type identification method based on a three-axis acceleration sensor and a neural network. The method specifically comprises steps as follows: step 1), establishing a database of gait acceleration signals; step 2), performing the segmentation stage of the signals in the corresponding period; step 3), removing a gravity factor; step 4), performing the gait feature extraction stage; step 5), performing gait presorting stage; step 6), performing dimensionality reduction operation of a gait feature set; and step 7), performing specific gait identification stage. According to the method, gait features are screened with a staged MIV (mean impact value) method, the gait type identification work is performed in combination of a BP (back propagation) neural network, the extracted features are taken as input independent variables of the neural network, six gait types of sitting, standing, walking in a low speed, walking in a high speed, going upstairs and going downstairs are effectively identified sequentially through the gait presorting stage and the specific gait identification stage, and the method can have higher accuracy and reliability through enlarging of a gait data capacity range and the optimized design of the neural network.
Description
Technical field
The invention belongs to the processing of biomedical signals technical field, relate to a kind of gait type discrimination method, be specifically related to a kind of type of gait based on 3-axis acceleration sensor and neural network discrimination method.
Background technology
Attitude when gait refers to people's walking is a kind of biological characteristic of complexity.It has the following advantages: less demanding to systemic resolution, be adapted to remote identification, be not subject to invade, be difficult to hide etc.Therefore the research of gait become to one of study hotspot of domestic and international research institution and university.
The research of abroad the gait type being differentiated and pay close attention to starting early, methods are mainly based on video image and two kinds of patterns of sensor.
Based on the research method of video image, its experiment is data from camera monitoring individual movement situation.This research method generally comprises: to processes such as the cutting apart of sequence image, image binaryzation, image space feature extraction, recognizer application.Its shortcoming be exactly fund input greatly, easily affected by environment, individual privacy can not be guaranteed and also the experimental data amount large, the view data of two dimension deals with comparatively trouble.
Another kind of research method is to detect and gather certain physiological signal data by individual wearable sensors, then is transferred to the analyzing and processing that host computer carries out data.The application that sensor type has breathing, heartbeat, energy, acceleration etc. to differentiate as the gait type.Acceleration signal have easily obtain, data accurately, the advantage such as the environmental interference factor is few.Acceleration transducer is widely applied to this problem.
Utilize the step state acceleration signal to carry out the type discriminating and will carry out analyzing and processing to the acceleration signal of the human body natural that obtains motion exactly, comprise usually that motor message obtains, the signal period is cut apart, signal characteristic abstraction, 4 processes of identification algorithm.
Wherein this process of signal characteristic abstraction directly affects final discriminating performance.Recent many researchers concentrate the feature that Sample Entropy, wavelet energy value, multi-scale wavelet entropy, Fourier descriptor are differentiated as the gait type analysis, and then will detect the feature of data and sample characteristics and apply all kinds of recognizers and complete final discriminating task.The shortcoming of this disposal route is to lack high accuracy and reliability, and one or both time-frequency characteristics of gait signal, when distinguishing two kinds of similar gaits, there will be the failed situation of differentiating.The method that also has the researcher to take to mix the time and frequency domain characteristics coupling is carried out the gait discriminating, but easily ignores the problem that may there be correlativity in different characteristic, the ill effect of getting half the result with twice the effort occurs.
In sum, the limitation that exists by traditional acceleration method has:
(1) aspect the sensor wearing position, wearing position all must be determined unified, and is attached on clothing, and this experimenter that can make is discomfort when experiment, makes gait lack naturality;
(2) splitting signal in chronological order, can make abrupt change signal sometime be separated separately, the impact erroneous judgement;
(3) there is very large correlativity in some feature of extracting between them, can replace all features by the feature that dimensionality reduction is chosen tool representative fully;
(4) some feature dimension reduction method is to screening gait signal characteristic poor effect, and resolution is not high.
Summary of the invention
The objective of the invention is for overcoming above-mentioned the deficiencies in the prior art, a kind of type of gait based on 3-axis acceleration sensor and neural network discrimination method is provided, by 3-axis acceleration sensor, obtain the body gait acceleration signal, the present invention is based on and obtain the body gait acceleration signal, adopt the method for differentiating stage by stage, 6 kinds of body gait types (sit, stand, be careful, hurry up,, downstairs) be can differentiate upstairs, 2 kinds of static gaits and 4 kinds of dynamic gaits comprised.
For achieving the above object, the present invention adopts following technical proposals:
Based on the gait type discrimination method of 3-axis acceleration sensor and neural network, concrete steps are as follows:
Step 1): set up the step state acceleration Signals Data Base, what obtain here is called original gait data;
Step 2): original gait data is removed to gravity factor by filtering, and the signal data that obtains is called the acceleration of motion data sample;
Step 3): signal is cut apart the stage same period: by original gait data and the step 2 in step 1)) in the acceleration of motion sample cut and obtain original gait data sample block and acceleration of motion blocks of data samples, all from interval the two class blocks of data samples of cutting apart, extract and obtain detecting sample and training sample respectively, detect sample this moment and comprise that original gait detects sample and acceleration of motion detects sample, training sample comprises original gait training sample and acceleration of motion training sample;
Step 4): Method of Gait Feature Extraction stage: from extracting eigenwert acceleration of motion training sample and acceleration of motion detection sample, detect sample and original gait training sample and extract eigenwert from original gait;
Step 5): gait is presorted the stage: utilize the eigenwert training first nerves network of the acceleration of motion training sample that step 4) obtains, utilize acceleration of motion that the first nerves network that trains obtains step 4) to detect sample and presort;
Step 6): gait feature collection dimensionality reduction operation: carry out respectively the dimensionality reduction of static gait feature collection and the dimensionality reduction of dynamic gait feature collection;
Step 7): gait is specifically differentiated the stage:
Training nervus opticus network, and with the nervus opticus network that trains, the original gait detection sample of presorting as static types is carried out to the discriminating of type;
Training third nerve network, and with the third nerve network that trains, the original gait detection sample of presorting as regime type is carried out to the discriminating of type.
The concrete grammar of described step 1) is:
11) the 3-axis acceleration sensor position of wearing all is unified in front, and is fixed with powerful adhesive plaster, and the data that gather XYZ tri-axles that come have represented the acceleration information on experimenter's vertical direction, left and right directions, fore-and-aft direction; 3-axis acceleration sensor MMA7260Q is used in the collection of gait signal, and the acceleration test range is ± 4g that g is free-fall velocity, 1g=9.8m/s
2, the systematic sampling frequency is 200Hz;
12) experimenter's number is some names, wear walking shoes with natural leg speed sit, stand, be careful, hurry up, upstairs, the experiment of 6 kinds of different gaits downstairs, every experimenter carried out every kind of gait Therapy lasted some minutes, thereby set up the step state acceleration Signals Data Base, the data obtained is called original gait signal.
Described step 2) for by acceleration of motion from original gait signal, separating, by original gait signal, through cutoff frequency, be 0.5Hz, ripple is the three rank Butterworth LPF of 0.01db, and gravity factor is removed, and obtains the acceleration of motion data.
The concrete grammar of described step 3) is: sample frequency is reduced to 100Hz from 200Hz, the Duplication that each data was used 5 seconds is that 50% rectangle time window is divided into 48 segment data samples, when the save data sample, for each data sample adds label 1～6, respectively representative sits, stand, be careful, hurry up, upstairs, the 6 kinds of different gaits of going downstairs, the data of everyone every kind gait 25% are chosen as detecting sample in interval, remaining 75% as training sample, and it is to choose first blocks of data samples in every four data sample block that described interval is extracted.
The step of described step 4) is as follows:
Step 41): from acceleration of motion training sample and the acceleration of motion of step 3), detect extraction signal amplitude area SMA and two eigenwerts of the average energy value AE sample respectively;
Step 42): from the original gait of step 3), detect the eigenwert of extracting X, Y, tri-axis data of Z sample and original gait training sample, the eigenwert of each axle comprises average, axle related coefficient, energy value, interquartile range, mean absolute difference, square root, standard deviation and 8 eigenwerts of variance, owing to there being three axles to amount to 24 eigenwerts.
The step of described step 5) is as follows:
Set up a BP neural network, using step 41) two eigenwerts of the acceleration of motion training sample that obtains feature set of merging into it inputs as neural network, training can be differentiated a BP neural network of dynamic gait and static gait, its error rate is met the demands: error rate, less than 0.02, then detects acceleration of motion the feature set of sample and presorts with a BP neural network that trains as input.
The step of described step 6) is as follows:
Step 61) static gait feature collection dimensionality reduction: 24 eigenwerts that label in step 3) belonged to static original gait training sample average influence value MIV method dimensionality reduction, the MIV value of gait being differentiated to the eigenwert that impact is larger is larger, chooses the maximum eigenwert of 8 MIV values and merges as static gait neural network input feature vector collection;
Step 62) dynamic gait feature collection dimensionality reduction: 24 eigenwerts that label in step 3) belonged to dynamic original gait training sample average influence value MIV method dimensionality reduction equally, choose the maximum eigenwert of 8 MIV values and merge as dynamic gait neural network input feature vector collection.
Described step 7) step is as follows:
Step 71) the concrete discriminating of static gait: set up the 2nd BP neural network for static gait is carried out to the particular type discriminating, utilize step 61) static gait neural network input feature vector collection as input training the 2nd BP neural network, after training completes, the original gait for static types of presorting in step 5) detects sample and enters in the 2nd BP neural network and specifically differentiate, differentiates that types results has two kinds: stand, sit;
Step 72) the dynamically concrete discriminating of gait: set up the 3rd BP neural network for dynamic gait is carried out to the particular type discriminating, utilize step 62) dynamic gait neural network input feature vector collection as input training the 3rd BP neural network, the original gait for regime type of presorting in step 5) after training completes detects sample and enters in the 3rd BP neural network and specifically differentiate, differentiates that types results has four kinds: hurry up, be careful, upstairs, downstairs.
The extraction of described step 4) gait feature is to use the feature calculation formula from gait data, calculating:
Step 41) computing formula be used to the stage of presorting is respectively:
Wherein w, N are as time window length, x
iy
iz
iRepresent the acceleration of motion data on xyz tri-axles of a window, A
kIt is the coefficient that the acceleration of motion data on xyz tri-axles of a window are carried out discrete FFT conversion;
Step 42) for concrete, differentiate that the part computing formula in type stage is:
Average
In formula, a
iBe the i sampled value of acceleration constantly, w is length of window; The axle related coefficient
In formula, a
ib
iRepresent respectively two out-of-alignment acceleration sampled signals, this formula is for calculating xy, the related coefficient between xz and yz tri-axles; Energy value
Variance
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value; Standard deviation
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value.
Described step 5) is presorted the stage in gait, and a BP neural network structure is: 2-5-2, and namely input layer has 2 nodes, and hidden layer has 5 nodes, and output layer has 2 nodes; Here the neural network activation function of hidden layer and output layer is all selected S type activation function
The anticipation error of neural network is set as 0.02, and the neural network algorithm that uses is the BP algorithm of Levnberg_Marquardt; During with acceleration of motion data sample neural network training, signal amplitude area SMA and the average energy value AE distinguish the static or dynamic input feature vector collection of gait as the neural network of presorting, Output rusults be 1,2 will be converted into [10]
TBe used for representing static gait, Output rusults is 3,4,5,6 will be converted into as [01]
TBe used for representing dynamic gait; 3 layers of BP be neural, and to flutter what open up that structural drawing expresses be the feature set column vector X of n * 1 dimension
n=(x
1, x
2..., x
n)
T(X
n∈ X) to m * 1 dimension identification result column vector Y
m=(Y
1, Y
2..., Y
m)
T(Y
m∈ Y) nonlinear function mapping relations, the BP neural network, by information forward-propagating and 2 training process of error back propagation, is regulated neural network weight and threshold value repeatedly, makes the error of neural network output valve and expectation value reach requirement.
Described step 6) static state/dynamically the dimension of the input feature vector collection of gait classification device is 24.Dimension to the input feature vector collection reduces, and removes the feature of redundancy, selects the feature set of the Nonlinear Mapping relation that can reflect neural network.Mean Impact Value (MIV) method is mainly after with the training of all gait signal characteristics, practicing a correct neural network, then each input feature vector collection is carried out to the difference computing is that each input variable ± 10% generates new samples p respectively
1, p
2(p
1p
2=former eigenwert * (1 ± 10%)), then as the neural network input of simulation sample, with the network of building up, carry out emulation respectively, obtain network Output simulation value A
1, A
2.Get again the arithmetic square root of its difference
It is Mean Impact Value.After Mean Impact Value calculates, choose the input independent variable of the feature of front 8 maximums of Mean Impact Value as the dynamic gait neural network classifier of optimizing.With identical Mean Impact Value method, choose the input independent variable of these 8 features of Mean Impact Value maximum as the static gait neural network classifier of optimizing.
Described step 7) is specifically differentiated the stage in the gait type, in its gait type stage of concrete discriminating, average, axle related coefficient, energy value, four minutes differences, mean absolute difference, root, standard deviation, these 8 features of variance are as the static/dynamic input feature vector collection of gait classification device, due to 3 axles being arranged, so the dimension of the input feature vector collection of static state/dynamic action sorter is 24.
Need to carry out the gait tag extraction and be converted into the operation of two-dimentional Output rusults: when dynamically gait was specifically classified, label was 3,4,5,6 Output rusults to be decided to be respectively to [1000]
T, [0100]
T, [0010]
T, [0001]
TWhen static gait classification, label is 1 Output rusults is decided to be [10]
T, label is 2 Output rusults to be decided to be to [01]
T.Dynamically the gait neural network structure is: 8-30-4.Static gait neural network structure is: 8-10-2.Here the activation function of hidden layer and output layer is all selected S type activation function:
The anticipation error of network is set as 0.02.The neural network algorithm that uses is the BP algorithm of Levnberg_Marquardt.The brief description of the BP algorithm of Levnberg_Marquardt: the network weight vector of establishing the k time iteration of W (k) expression, dimension is M, and new weight vector W (k+1) can try to achieve according to following rule: W (k+1)=W (k)+Δ W (k) (Δ W is the weights increment); In the algorithm of Levnberg_Marquardt, its Δ W form is: Δ W=-[J
T(W) J (W)+uI]
-1J (W) e (W), in formula, u is scale-up factor, is positive constant; I is unit matrix; J (W) is the Jacobian matrix; E (W) is desired output and the actual error vector of exporting.The each iteration efficiency of the BP algorithm of Levnberg_Marquardt is very high, can greatly improve the overall performance of neural network.
Beneficial effect of the present invention:
1, the present invention's data used are immediately obtained and process by wearing with oneself 3-axis acceleration sensor, and cost is lower, are easier to be integrated in other portable medical monitoring equipment.Wear acceleration transducer and be selected in front, avoid the experimenter to wear discomfort, guaranteed the naturality of gait.Every kind of gait data is good with the feature continuity of the data sample extraction that rectangular window is partitioned into 50% overlapping time of 5 seconds, can well keep general character each other.
2, the present invention utilizes neural network that the ability of stronger learning ability and autonomous learning complex mappings is arranged on distinguishing nonlinear can be classified, by information forward-propagating and 2 training process of error back propagation, regulating networks weights and threshold value, make the error of neural network forecast value and expectation value reach requirement repeatedly.
3, the present invention utilizes MIV method screening gait feature stage by stage, in conjunction with the BP neural network, carry out gait type discriminating work, the feature of extraction is inputted to independent variable as neural network, successively through gait presort, gait specifically differentiated for two stages, realized seat, stood, be careful, hurried up, upstairs, effective discriminating of 6 kinds of gait types downstairs, and, by from now on to the increase of gait data range of capacity and the optimal design of neural network, will have higher accuracy and reliability.
The accompanying drawing explanation
Fig. 1 is for cutting apart experimenter's gait signal of being careful with 50% overlapping rectangle time window;
Fig. 2 is that neural the flutterring of BP opened up structural drawing;
Fig. 3 is the comparison diagrams of experimenter's static state/dynamic two class gaits on the SMA/AE feature;
Fig. 4 (a) for four kinds of dynamic gaits (be careful, hurry up) of experimenter at X-axis standard deviation and X-axis energy value Characteristic Contrast figure;
Fig. 4 (b) is that four kinds of dynamic gaits (upstairs, downstairs) of experimenter are at X-axis standard deviation and X-axis energy value Characteristic Contrast figure;
Fig. 5 is experimenter's seat and stands two kinds of static gaits at X-axis root mean square and X-axis energy value Characteristic Contrast figure;
Fig. 6 is that the gait type is differentiated the simulation algorithm process flow diagram.
Embodiment
The invention will be further described below in conjunction with accompanying drawing and embodiment.
As shown in Figure 6, embodiment 1:
The present invention, according to the uniform requirement of portable cardiac monitor system, wears by human body, obtains on human body natural's step state acceleration basis of signals, carries out the discriminating of gait type.
1) collection of experimenter's gait data
11) the 3-axis acceleration sensor MMA7260Q of Freescale company is used in the collection of gait signal, and the acceleration test range is ± 4g that the systematic sampling frequency is 200Hz.The position that sensor is worn all is unified in front, and is fixed with powerful adhesive plaster, and the data that gather XYZ tri-axles that come have represented the acceleration information on experimenter's vertical direction, left and right directions, fore-and-aft direction.
12) experimenter's number is 10, (age is between 22～28 years old for 6 women of 4 male sex, body weight is between 45～75kg), wear walking shoes with natural leg speed sit, stand, be careful, hurry up, upstairs, the experiment of 6 kinds of different gaits downstairs, every experimenter carried out every kind of gait Therapy lasted 2 minutes.So just set up a capacity and be the step state acceleration Signals Data Base of 60 groups, the data of gained are called original gait data.
2) data cuts apart and pre-service
21) as can be known by other results of study, the frequency of the suffered gravity of human body is below 0.5Hz, for by acceleration of motion from the step state acceleration signal, separating, the design cutoff frequency is 0.5Hz, 0.01db three rank Butterworth LPF of ripple, gait data sample device after filtering, just gravity factor can be removed, what obtain is that the acceleration of motion data just can be used for the signal extraction of presorting.Dynamically/during the static classification of motion, use be original gait data.
22) adopt the method for two interval values of taking out (two data extract first in order), sample frequency is reduced to 100Hz from 200Hz, eliminate partial noise.After frequency reducing, everyone every kind gait data amount is 12000, and data mode corresponding to two classes arranged: acceleration of motion data and original gait data.Two kinds of data modes of every kind of gait all use 5 seconds 50% Duplication rectangle time windows to be divided into 48 segment data sample block, after cutting apart when the save data sample block, for every kind of gait data adds label, label is numeral 1～6, be used for representative stand, sit, be careful, hurry up, upstairs, the six kinds of different gaits of going downstairs, label position is placed on first of gait signal data.All experimenters' gait data sample summation is 2880 groups (48 * 6 * 10), and what Fig. 1 represented is how the gait signal of being careful to be carried out to time window to cut apart.Every kind of gait data sample of each class data has 480 groups, the data of everyone every kind gait 25% are chosen as detecting sample in interval, remaining 75% as training sample, annotate: detect sample and comprise that original gait detects sample and acceleration of motion detects sample, training sample comprises original gait training sample and acceleration of motion training sample.
3) extract the gait feature for differentiating
The extraction of gait feature is to use the feature calculation formula from gait data, calculating.In gait, presort the stage, signal amplitude area (SMA) and the average energy value (AE) are distinguished the input feature vector collection of static/dynamic gait as pre-classifier, and computing formula is respectively:
Wherein w, N are as time window length, x
iy
iz
iRepresent the acceleration of motion data on xyz tri-axles of a window, A
kIt is the coefficient that the acceleration of motion data on xyz tri-axles of a window are carried out discrete FFT conversion.As seen from Figure 3, from these two features that the acceleration of motion data are extracted, can easily distinguish static state/dynamic gait, because two eigenwerts of this of static gait are definitely less than dynamically.And the gait type is specifically differentiated the stage, average, axle related coefficient, energy value, four minutes differences, mean absolute difference, root, standard deviation, these 8 features of variance are as the static/dynamic input feature vector collection of gait classification device, and these 8 eigenwert part computing formula are: average
In formula, a
iBe the i sampled value of acceleration constantly, w is length of window; The axle related coefficient
In formula, a
ib
iRepresent respectively two out-of-alignment acceleration sampled signals, this formula is for calculating xy, the related coefficient between xz and yz tri-axles; Energy value
A in formula
kBe the coefficient that the acceleration of motion data on xyz tri-axles of a window are carried out discrete FFT conversion, N is for being time window length; Variance
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value; Standard deviation
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value; Due to 3 axles being arranged, so the dimension of the input feature vector collection of static state/dynamic action sorter is 24.
4) gait is presorted the stage
In gait, presort the stage, the tag extraction of at first gait being moved is converted into Output rusults operation and is: label is 1,2 Output rusults is decided to be [10]
TBe used for representing the static types gait; Label is 3,4,5,6 Output rusults to be decided to be to [01]
TBe used for representing the regime type gait.
That the 3 layers of BP neural network of using this paper realize is the feature set column vector X of (n * 1) dimension
n=(x
1, x
2..., x
n)
T(X
n∈ X) to (m * 1) dimension identification result column vector Y
m=(Y
1, Y
2..., Y
m)
T(Y
m∈ Y) nonlinear function mapping.As shown in Figure 2, the gait of the structure neural network structure of presorting is the structural drawing of BP neural network: 2-5-2, and namely input layer has 2 nodes, and hidden layer has 5 nodes, and output layer has 2 nodes, and (feature as input has two, therefore the input layer number is 2; The gait type of output has two kinds, therefore the output layer nodes is 2).Here the activation function of hidden layer and output layer is all selected S type activation function:
The anticipation error of network is set as 0.02.The neural network algorithm that uses is the BP algorithm of Levnberg_Marquardt.
First trained sample characteristics collection is transferred to hidden layer forward as input, through activation function f (u) effect, has Output rusults
Produce; The result of hidden layer back kick again is passed to output layer, has Output rusults
Produce, wherein w in above-mentioned two formulas
Ijw
JkFor the weights of each layer of BP neural network junction, θ
Hjθ
OkFor the threshold value of hidden layer and output layer, f is S type activation function, H
Jp, Y
KpBe respectively hidden layer Output rusults, output layer Output rusults.If final Output rusults Y
KpWith expectation value, there is undesirable error, then regulating networks weight w repeatedly forward
Ijw
JkWith threshold value θ
Hjθ
Ok, make the error of neural network forecast value and expectation value reach requirement.
5) original gait feature collection MIV method dimensionality reduction
Static/dynamically the dimension of the input feature vector collection of gait classification device is 24.In order to prevent the over-fitting of neural network, improve modeling accuracy, need to the dimension of input feature vector collection be reduced, remove the feature of redundancy, select the feature set of the Nonlinear Mapping relation that can reflect neural network.Mean Impact Value (MIV) method is mainly after with the training of all gait signal characteristics, practicing a correct neural network, then each input feature vector collection is carried out to the difference computing is that each input variable ± 10% generates new samples p respectively
1p
2, p
1p
2As simulation sample, with the network of building up, carry out emulation respectively again, obtain simulation value A
1A
2, then get the arithmetic square root of its difference, i.e. the MIV value.After MIV method dimensionality reduction, choose the feature of front 8 maximums of MIV value: X-axis standard deviation, X-axis energy value, X-axis root mean square, XZ axle related coefficient, YZ axle related coefficient, Y-axis energy value, Y-axis root mean square, X-axis variance are as the input independent variable of the dynamic gait neural network classifier of optimizing.With identical MIV method, choose the input independent variable of these 8 features of Y-axis root mean square, Y-axis energy value, X-axis root mean square, Z axis root mean square, YZ axle related coefficient, Z axis energy value, X-axis energy value, four minutes poor MIV value maximums of Z axis as the static gait neural network classifier of optimizing.
From Fig. 4 (a) and Fig. 4 (b), can find out that four kinds of dynamic gaits have good discrimination at the maximum that is filtered out by MIV front 2 eigenwert X-axis standard deviations, X-axis energy values, simultaneously two kinds of static gaits otherness on X-axis root mean square and 2 features of X-axis energy value is apparent in view as seen from Figure 5, and this shows that the MIV method can select the feature set of the Nonlinear Mapping relation that can reflect neural network.
6) the gait type is specifically differentiated the stage
In the gait type, specifically differentiate the stage, need to carry out the gait tag extraction and be converted into the operation of two-dimentional Output rusults: when dynamically gait was specifically classified, label was 3,4,5,6 Output rusults to be decided to be respectively to [1000]
T, [0100]
T, [0010]
T, [0001]
TWhen static gait classification, label is 1 Output rusults is decided to be [10]
T, label is 2 Output rusults to be decided to be to [01]
T.Dynamically the gait neural network structure is: 8-30-4, because choose 8 feature sets the most representative, the dynamic action that differentiate has 4 kinds.Static gait neural network structure is: 8-10-2, because choose 8 feature sets the most representative, the action of the static state that differentiate has 2 kinds.Here the activation function of hidden layer and output layer is all selected the S type:
The anticipation error of network is set as 0.02.The neural network algorithm that uses is the BP algorithm of Levnberg_Marquardt.
Although above-mentionedly by reference to the accompanying drawings the specific embodiment of the present invention is described; but be not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various modifications that creative work can make or distortion still in protection scope of the present invention.
Claims (10)
1. based on the gait type discrimination method of 3-axis acceleration sensor and neural network, it is characterized in that, concrete steps are as follows:
Step 1): set up the step state acceleration Signals Data Base, what obtain here is called original gait data;
Step 2): original gait data is removed to gravity factor by filtering, and the signal data that obtains is called the acceleration of motion data sample;
Step 3): signal is cut apart the stage same period: by original gait data and the step 2 in step 1)) in the acceleration of motion sample cut and obtain original gait data sample block and acceleration of motion blocks of data samples, all from interval the two class blocks of data samples of cutting apart, extract and obtain detecting sample and training sample respectively, detect sample this moment and comprise that original gait detects sample and acceleration of motion detects sample, training sample comprises original gait training sample and acceleration of motion training sample;
Step 4): Method of Gait Feature Extraction stage: from extracting eigenwert acceleration of motion training sample and acceleration of motion detection sample, detect sample and original gait training sample and extract eigenwert from original gait;
Step 5): gait is presorted the stage: utilize the eigenwert training first nerves network of the acceleration of motion training sample that step 4) obtains, utilize acceleration of motion that the first nerves network that trains obtains step 4) to detect sample and presort;
Step 6): gait feature collection dimensionality reduction operation: carry out respectively the dimensionality reduction of static gait feature collection and the dimensionality reduction of dynamic gait feature collection;
Step 7): gait is specifically differentiated the stage:
Training nervus opticus network, and with the nervus opticus network that trains, the original gait detection sample of presorting as static types is carried out to the discriminating of type;
Training third nerve network, and with the third nerve network that trains, the original gait detection sample of presorting as regime type is carried out to the discriminating of type.
2. discrimination method as claimed in claim 1, is characterized in that,
The concrete grammar of described step 1) is:
11) front is all unified to be fixed in the 3-axis acceleration sensor position of wearing, and the data that gather XYZ tri-axles that come have represented the acceleration information on experimenter's vertical direction, left and right directions, fore-and-aft direction; 3-axis acceleration sensor MMA7260Q is used in the collection of gait signal, and the acceleration test range is ± 4g that g is free-fall velocity, 1g=9.8m/s
2, the systematic sampling frequency is 200Hz;
12) experimenter's number is some names, wear walking shoes with natural leg speed sit, stand, be careful, hurry up, upstairs, the experiment of 6 kinds of different gaits downstairs, every experimenter carried out every kind of gait Therapy lasted some minutes, thereby set up the step state acceleration Signals Data Base, the data obtained is called original gait signal.
3. discrimination method as claimed in claim 1, is characterized in that,
Described step 2) for by acceleration of motion from original gait signal, separating, by original gait signal, through cutoff frequency, be 0.5Hz, ripple is the low-pass filter of 0.01db, and gravity factor is removed, and obtains the acceleration of motion data.
4. discrimination method as claimed in claim 1, is characterized in that,
The concrete grammar of described step 3) is: sample frequency is reduced to 100Hz from 200Hz, the Duplication that each data was used 5 seconds is that 50% rectangle time window is divided into 48 segment data samples, when the save data sample, for each data sample adds label 1～6, respectively representative sits, stand, be careful, hurry up, upstairs, the 6 kinds of different gaits of going downstairs, the data of everyone every kind gait 25% are chosen as detecting sample in interval, remaining 75% as training sample, and it is to choose first blocks of data samples in every four data sample block that described interval is extracted.
5. discrimination method as claimed in claim 1, is characterized in that,
The step of described step 4) is as follows:
Step 41): from acceleration of motion training sample and the acceleration of motion of step 3), detect extraction signal amplitude area SMA and two eigenwerts of the average energy value AE sample respectively;
Step 42): from the original gait of step 3), detect the eigenwert of extracting X, Y, tri-axis data of Z sample and original gait training sample, the eigenwert of each axle comprises average, axle related coefficient, energy value, interquartile range, mean absolute difference, square root, standard deviation and 8 eigenwerts of variance, owing to there being three axles to amount to 24 eigenwerts.
6. discrimination method as claimed in claim 1, is characterized in that, the step of described step 5) is as follows:
Set up a BP neural network, using step 41) two eigenwerts of the acceleration of motion training sample that obtains feature set of merging into it inputs as neural network, training can be differentiated a BP neural network of dynamic gait and static gait, its error rate is met the demands: error rate, less than 0.02, then detects acceleration of motion the feature set of sample and presorts with a BP neural network that trains as input.
7. discrimination method as claimed in claim 1, is characterized in that, the step of described step 6) is as follows:
Step 61) static gait feature collection dimensionality reduction: 24 eigenwerts that label in step 3) belonged to static original gait training sample average influence value MIV method dimensionality reduction, the MIV value of gait being differentiated to the eigenwert that impact is larger is larger, chooses the maximum eigenwert of 8 MIV values and merges as static gait neural network input feature vector collection;
Step 62) dynamic gait feature collection dimensionality reduction: 24 eigenwerts that label in step 3) belonged to dynamic original gait training sample average influence value MIV method dimensionality reduction equally, choose the maximum eigenwert of 8 MIV values and merge as dynamic gait neural network input feature vector collection.
8. discrimination method as claimed in claim 1, is characterized in that, described step 7) step as follows:
Step 71) the concrete discriminating of static gait: set up the 2nd BP neural network for static gait is carried out to the particular type discriminating, utilize step 61) static gait neural network input feature vector collection as input training the 2nd BP neural network, after training completes, the original gait for static types of presorting in step 5) detects sample and enters in the 2nd BP neural network and specifically differentiate, differentiates that types results has two kinds: stand, sit;
Step 72) the dynamically concrete discriminating of gait: set up the 3rd BP neural network for dynamic gait is carried out to the particular type discriminating, utilize step 62) dynamic gait neural network input feature vector collection as input training the 3rd BP neural network, the original gait for regime type of presorting in step 5) after training completes detects sample and enters in the 3rd BP neural network and specifically differentiate, differentiates that types results has four kinds: hurry up, be careful, upstairs, downstairs.
9. discrimination method as described as claim 1 or 5, is characterized in that, the extraction of described step 4) gait feature is to use the feature calculation formula from gait data, calculating:
Step 41) computing formula be used to the stage of presorting is respectively:
Wherein w, N are as time window length, x
iy
iz
iRepresent the acceleration of motion data on xyz tri-axles of a window, A
kIt is the coefficient that the acceleration of motion data on xyz tri-axles of a window are carried out discrete FFT conversion;
Step 42) for concrete, differentiate that the part computing formula in type stage is:
Average
In formula, a
iBe the i sampled value of acceleration constantly, w is length of window; The axle related coefficient
In formula, a
ib
iRepresent respectively two out-of-alignment acceleration sampled signals, this formula is for calculating xy, the related coefficient between xz and yz tri-axles; Energy value
Variance
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value; Standard deviation
A in formula
iBe the i sampled value of acceleration constantly, u is a
iMean value.
10. discrimination method as described as claim 1 or 6, is characterized in that, described step 5) is presorted the stage in gait, and a BP neural network structure is: 2-5-2, and namely input layer has 2 nodes, and hidden layer has 5 nodes, and output layer has 2 nodes; Here the neural network activation function of hidden layer and output layer is all selected S type activation function
The anticipation error of neural network is set as 0.02, and the neural network algorithm that uses is the BP algorithm of Levnberg_Marquardt; During with acceleration of motion data sample neural network training, signal amplitude area SMA and the average energy value AE distinguish the static or dynamic input feature vector collection of gait as the neural network of presorting, Output rusults be 1,2 will be converted into [10]
TBe used for representing static gait, Output rusults is 3,4,5,6 will be converted into as [01]
TBe used for representing dynamic gait; 3 layers of BP be neural, and to flutter what open up that structural drawing expresses be the feature set column vector X of n * 1 dimension
n=(x
1, x
2..., x
n)
T(X
n∈ X) to m * 1 dimension identification result column vector Y
m=(Y
1, Y
2..., Y
m)
T(Y
m∈ Y) nonlinear function mapping relations, the BP neural network, by information forward-propagating and 2 training process of error back propagation, is regulated neural network weight and threshold value repeatedly, makes the error of neural network output valve and expectation value reach requirement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013103673511A CN103400123A (en) | 2013-08-21 | 2013-08-21 | Gait type identification method based on three-axis acceleration sensor and neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013103673511A CN103400123A (en) | 2013-08-21 | 2013-08-21 | Gait type identification method based on three-axis acceleration sensor and neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103400123A true CN103400123A (en) | 2013-11-20 |
Family
ID=49563741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013103673511A Pending CN103400123A (en) | 2013-08-21 | 2013-08-21 | Gait type identification method based on three-axis acceleration sensor and neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103400123A (en) |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104269025A (en) * | 2014-09-29 | 2015-01-07 | 南京信息工程大学 | Wearable type single-node feature and position selecting method for monitoring outdoor tumble |
| WO2015120824A1 (en) * | 2014-02-17 | 2015-08-20 | Hong Kong Baptist University | Gait measurement with 3-axes accelerometer/gyro in mobile devices |
| CN105125220A (en) * | 2015-10-20 | 2015-12-09 | 重庆软汇科技股份有限公司 | Falling-down detection method |
| CN105760819A (en) * | 2016-01-28 | 2016-07-13 | 西南大学 | Daily activity recognition method based on acceleration signal |
| CN106101060A (en) * | 2016-05-24 | 2016-11-09 | 杭州华三通信技术有限公司 | A kind of information detecting method and device |
| CN106295680A (en) * | 2016-07-29 | 2017-01-04 | 深圳云智优泊科技有限公司 | Low-power consumption is driven and ambulatory status identification system |
| CN106344031A (en) * | 2016-08-29 | 2017-01-25 | 常州市钱璟康复股份有限公司 | Sound feedback-based gait training and estimating system |
| CN106563645A (en) * | 2016-11-01 | 2017-04-19 | 上海师范大学 | Intelligent piezoelectric film sensor sorting method based on tensor decomposition |
| CN107016411A (en) * | 2017-03-28 | 2017-08-04 | 北京犀牛数字互动科技有限公司 | Data processing method and device |
| CN107016346A (en) * | 2017-03-09 | 2017-08-04 | 中国科学院计算技术研究所 | gait identification method and system |
| CN107277222A (en) * | 2016-12-20 | 2017-10-20 | 浙江从泰网络科技有限公司 | User behavior state judging method based on mobile phone built-in sensors |
| CN107316052A (en) * | 2017-05-24 | 2017-11-03 | 中国科学院计算技术研究所 | A kind of robust Activity recognition method and system based on inexpensive sensor |
| CN107403154A (en) * | 2017-07-20 | 2017-11-28 | 四川大学 | A kind of gait recognition method based on dynamic visual sensor |
| CN107958221A (en) * | 2017-12-08 | 2018-04-24 | 北京理工大学 | A kind of human motion Approach for Gait Classification based on convolutional neural networks |
| CN108073154A (en) * | 2016-11-11 | 2018-05-25 | 横河电机株式会社 | Information processing unit, information processing method and recording medium |
| CN108563939A (en) * | 2018-04-25 | 2018-09-21 | 常州大学 | Human body identification based on gait geometric locus feature |
| CN108814585A (en) * | 2018-05-03 | 2018-11-16 | 深圳竹信科技有限公司 | ECG's data compression method, apparatus and computer readable storage medium |
| CN109009143A (en) * | 2018-07-12 | 2018-12-18 | 杭州电子科技大学 | A method of ecg information is predicted by body gait |
| CN109325428A (en) * | 2018-09-05 | 2019-02-12 | 周军 | Mankind's activity gesture recognition method based on multi-level end-to-end neural network |
| CN109512643A (en) * | 2017-09-20 | 2019-03-26 | 三星电子株式会社 | Device of walking aid and the method for controlling the device |
| TWI657800B (en) * | 2017-11-03 | 2019-05-01 | 國立成功大學 | Method and system for analyzing gait |
| CN109753172A (en) * | 2017-11-03 | 2019-05-14 | 矽统科技股份有限公司 | The classification method and system and touch panel product of touch panel percussion event |
| CN110705584A (en) * | 2019-08-21 | 2020-01-17 | 深圳壹账通智能科技有限公司 | Emotion recognition method, emotion recognition device, computer device and storage medium |
| CN111512178A (en) * | 2017-12-08 | 2020-08-07 | 认知系统公司 | Machine learning motion detection based on wireless signal attributes |
| CN111870248A (en) * | 2020-06-05 | 2020-11-03 | 安徽建筑大学 | Motion state feature extraction and identification method based on 3D acceleration signal |
| CN112932469A (en) * | 2021-01-26 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | CNN + Transformer-based triaxial acceleration activity identification method |
| CN112946318A (en) * | 2021-03-24 | 2021-06-11 | 苏州康旺聚贤智能科技有限公司 | Calibration algorithm of acceleration sensor |
| CN111512178B (en) * | 2017-12-08 | 2024-06-04 | 认知系统公司 | Machine-learned motion detection based on wireless signal properties |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807245A (en) * | 2010-03-02 | 2010-08-18 | 天津大学 | Artificial neural network-based multi-source gait feature extraction and identification method |
| WO2011026001A2 (en) * | 2009-08-28 | 2011-03-03 | Allen Joseph Selner | Characterizing a physical capability by motion analysis |
| CN201812295U (en) * | 2010-10-09 | 2011-04-27 | 天津职业技术师范大学 | Zigbee and accelerometer based human body gait data extraction device |
-
2013
- 2013-08-21 CN CN2013103673511A patent/CN103400123A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011026001A2 (en) * | 2009-08-28 | 2011-03-03 | Allen Joseph Selner | Characterizing a physical capability by motion analysis |
| CN101807245A (en) * | 2010-03-02 | 2010-08-18 | 天津大学 | Artificial neural network-based multi-source gait feature extraction and identification method |
| CN201812295U (en) * | 2010-10-09 | 2011-04-27 | 天津职业技术师范大学 | Zigbee and accelerometer based human body gait data extraction device |
Non-Patent Citations (5)
| Title |
|---|
| 刘蓉: ""人体运动信息获取及物理活动识别研究"", 《中国博士学位论文全文数据库 社会科学II辑》 * |
| 戴永，等: ""基于RBF神经网络的手绘电气草图分类研究"", 《湘谭大学自然科学学报》 * |
| 王科俊等: ""步态识别中的步态检测与序列预处理"", 《模式识别与仿真》 * |
| 王美，等: ""改进的BP神经网络在糖尿病危险因素分析中的应用"", 《软件导刊》 * |
| 顾姚媛: ""基于MIV特征筛选和BP神经网络的三维人体参数转换"", 《上海工程技术大学学报》 * |
Cited By (37)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015120824A1 (en) * | 2014-02-17 | 2015-08-20 | Hong Kong Baptist University | Gait measurement with 3-axes accelerometer/gyro in mobile devices |
| CN104269025A (en) * | 2014-09-29 | 2015-01-07 | 南京信息工程大学 | Wearable type single-node feature and position selecting method for monitoring outdoor tumble |
| CN104269025B (en) * | 2014-09-29 | 2016-09-28 | 南京信息工程大学 | Wearable single node feature and the position choosing method of monitoring is fallen down towards open air |
| CN105125220A (en) * | 2015-10-20 | 2015-12-09 | 重庆软汇科技股份有限公司 | Falling-down detection method |
| CN105760819A (en) * | 2016-01-28 | 2016-07-13 | 西南大学 | Daily activity recognition method based on acceleration signal |
| CN105760819B (en) * | 2016-01-28 | 2019-03-26 | 西南大学 | A kind of daily routines recognition methods based on acceleration signal |
| CN106101060A (en) * | 2016-05-24 | 2016-11-09 | 杭州华三通信技术有限公司 | A kind of information detecting method and device |
| CN106295680A (en) * | 2016-07-29 | 2017-01-04 | 深圳云智优泊科技有限公司 | Low-power consumption is driven and ambulatory status identification system |
| CN106344031A (en) * | 2016-08-29 | 2017-01-25 | 常州市钱璟康复股份有限公司 | Sound feedback-based gait training and estimating system |
| CN106563645A (en) * | 2016-11-01 | 2017-04-19 | 上海师范大学 | Intelligent piezoelectric film sensor sorting method based on tensor decomposition |
| CN108073154B (en) * | 2016-11-11 | 2021-05-18 | 横河电机株式会社 | Information processing apparatus, information processing method, and recording medium |
| CN108073154A (en) * | 2016-11-11 | 2018-05-25 | 横河电机株式会社 | Information processing unit, information processing method and recording medium |
| CN107277222A (en) * | 2016-12-20 | 2017-10-20 | 浙江从泰网络科技有限公司 | User behavior state judging method based on mobile phone built-in sensors |
| CN107277222B (en) * | 2016-12-20 | 2020-12-15 | 浙江斑智科技有限公司 | User behavior state judgment method based on built-in sensor of mobile phone |
| CN107016346A (en) * | 2017-03-09 | 2017-08-04 | 中国科学院计算技术研究所 | gait identification method and system |
| CN107016411B (en) * | 2017-03-28 | 2020-09-29 | 北京犀牛数字互动科技有限公司 | Data processing method and device |
| CN107016411A (en) * | 2017-03-28 | 2017-08-04 | 北京犀牛数字互动科技有限公司 | Data processing method and device |
| CN107316052A (en) * | 2017-05-24 | 2017-11-03 | 中国科学院计算技术研究所 | A kind of robust Activity recognition method and system based on inexpensive sensor |
| CN107403154A (en) * | 2017-07-20 | 2017-11-28 | 四川大学 | A kind of gait recognition method based on dynamic visual sensor |
| CN109512643A (en) * | 2017-09-20 | 2019-03-26 | 三星电子株式会社 | Device of walking aid and the method for controlling the device |
| TWI657800B (en) * | 2017-11-03 | 2019-05-01 | 國立成功大學 | Method and system for analyzing gait |
| CN109753172A (en) * | 2017-11-03 | 2019-05-14 | 矽统科技股份有限公司 | The classification method and system and touch panel product of touch panel percussion event |
| CN111512178B (en) * | 2017-12-08 | 2024-06-04 | 认知系统公司 | Machine-learned motion detection based on wireless signal properties |
| CN111512178A (en) * | 2017-12-08 | 2020-08-07 | 认知系统公司 | Machine learning motion detection based on wireless signal attributes |
| CN107958221A (en) * | 2017-12-08 | 2018-04-24 | 北京理工大学 | A kind of human motion Approach for Gait Classification based on convolutional neural networks |
| CN108563939B (en) * | 2018-04-25 | 2022-05-20 | 常州大学 | Human body identity recognition based on gait track curve characteristics |
| CN108563939A (en) * | 2018-04-25 | 2018-09-21 | 常州大学 | Human body identification based on gait geometric locus feature |
| CN108814585A (en) * | 2018-05-03 | 2018-11-16 | 深圳竹信科技有限公司 | ECG's data compression method, apparatus and computer readable storage medium |
| CN108814585B (en) * | 2018-05-03 | 2021-05-28 | 深圳竹信科技有限公司 | Electrocardiosignal processing method and device and computer readable storage medium |
| CN109009143A (en) * | 2018-07-12 | 2018-12-18 | 杭州电子科技大学 | A method of ecg information is predicted by body gait |
| CN109009143B (en) * | 2018-07-12 | 2021-01-29 | 杭州电子科技大学 | Method for predicting electrocardio information through human gait |
| CN109325428B (en) * | 2018-09-05 | 2020-11-27 | 周军 | Human activity posture recognition method based on multilayer end-to-end neural network |
| CN109325428A (en) * | 2018-09-05 | 2019-02-12 | 周军 | Mankind's activity gesture recognition method based on multi-level end-to-end neural network |
| CN110705584A (en) * | 2019-08-21 | 2020-01-17 | 深圳壹账通智能科技有限公司 | Emotion recognition method, emotion recognition device, computer device and storage medium |
| CN111870248A (en) * | 2020-06-05 | 2020-11-03 | 安徽建筑大学 | Motion state feature extraction and identification method based on 3D acceleration signal |
| CN112932469A (en) * | 2021-01-26 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | CNN + Transformer-based triaxial acceleration activity identification method |
| CN112946318A (en) * | 2021-03-24 | 2021-06-11 | 苏州康旺聚贤智能科技有限公司 | Calibration algorithm of acceleration sensor |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103400123A (en) | Gait type identification method based on three-axis acceleration sensor and neural network | |
| CN105678222B (en) | A kind of mobile device-based Human bodys' response method | |
| Javeed et al. | Wearable sensors based exertion recognition using statistical features and random forest for physical healthcare monitoring | |
| Tran et al. | Human activities recognition in android smartphone using support vector machine | |
| Kwon et al. | Unsupervised learning for human activity recognition using smartphone sensors | |
| Li et al. | An automatic user-adapted physical activity classification method using smartphones | |
| CN104598880A (en) | Behavior identification method based on fuzzy support vector machine | |
| Uddin et al. | A guided random forest based feature selection approach for activity recognition | |
| Tian et al. | MEMS-based human activity recognition using smartphone | |
| Uddin et al. | Random forests based recognition of human activities and postural transitions on smartphone | |
| Jia et al. | Two‐Dimensional Extreme Learning Machine | |
| Hnoohom et al. | Human activity recognition using triaxial acceleration data from smartphone and ensemble learning | |
| Radhika et al. | Smartphone-based human activities recognition system using random forest algorithm | |
| CN103345626A (en) | Intelligent wheelchair static gesture identification method | |
| CN112294295A (en) | Human body knee motion posture identification method based on extreme learning machine | |
| Jane et al. | Sign language interpreter: Classification of forearm emg and imu signals for signing exact english | |
| WO2023035093A1 (en) | Inertial sensor-based human body behaviour recognition method | |
| Ahamed et al. | Fuzzy inference system-based recognition of slow, medium and fast running conditions using a triaxial accelerometer | |
| CN114358194A (en) | Gesture tracking based detection method for abnormal limb behaviors of autism spectrum disorder | |
| Sheng et al. | An adaptive time window method for human activity recognition | |
| Shi et al. | Fall detection system based on inertial mems sensors: Analysis design and realization | |
| Yuan et al. | Smartphone-based activity recognition using hybrid classifier | |
| Nazari et al. | Human activity recognition from knee angle using machine learning techniques | |
| Bayat et al. | Human gait recognition using bag of words feature representation method | |
| CN109271889A (en) | A kind of action identification method based on the double-deck LSTM neural network |
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 |
Application publication date: 20131120 |
|
| RJ01 | Rejection of invention patent application after publication |