CN104598880A - Behavior identification method based on fuzzy support vector machine - Google Patents

Abstract

The invention discloses a behavior identification method based on a fuzzy support vector machine and adopts the fuzzy support vector machine to realize identification for various human behaviors (including normal behaviors such as standing, walking, running, going upstairs/downstairs and abnormal behaviors such as falling down); the behavior identification method is mainly applied to eliminating the influences of isolated points and noise points in the sample points to classification effects and improving behavior identification precision. The main contents for realization of the behavior identification are as follows: firstly, behavior data acquisition is realized by using a tri-axial accelerometer to obtain an X-axis acceleration, a Y-axis acceleration and a Z-axis acceleration; the mean value, the variance and the energy of the resultant acceleration as well as the correlation coefficient between any two dimensions of three-dimensional data are respectively extracted by means of resultant acceleration extracting characteristic values, and a six-dimensional characteristic vector is obtained; secondly, the degree of membership of each sampling point to the affiliated classification is calculated; thirdly, the construction of a classification model is realized by using the fuzzy support vector machine; and fourthly, the identification for human behaviors is realized at the online stage.

Description

A kind of Activity recognition method based on fuzzy support vector machine

Technical field

The present invention relates to Activity recognition technical field, relate in particular to a kind of Human bodys' response method based on fuzzy support vector machine.

Background technology

In recent years, health becomes the focus that people pay close attention to, and carries out identifying to human body behavior and monitors, can the activity of indirectly assessor, identification for abnormal behaviour (such as falling) can well process some unexpected situations, especially for old man.Also indirectly can be obtained some habits and customs of monitored person by the identification of the behavior to human body, the more physiological situation of human body can be obtained thus.

Traditional Human bodys' response is the mode based on image graphics, although based on the Human bodys' response technical development relatively morning of image, Theory comparison is ripe, but there is deficiency based on the Human bodys' response technology of image, be mainly manifested in the problems such as image processing algorithm is more complicated, hardware cost is higher, application scenarios is limited, accuracy rate is not high.

Sensor-based Human bodys' response, has fairly obvious advantage, is mainly manifested in, and is free to obtain human body daily behavior data, can be carried at measured easily and wait with it.

For sensor-based Human bodys' response technology, the main sorting algorithm of application has KNN algorithm, decision Tree algorithms, SVM, very fast learning machine algorithm.

KNN algorithm, exist calculate and storage overhead large; Classification speed is slow, there is higher-dimension disaster problem under big-sample data, and the precision of classification is low, cannot the impact of stress release treatment point and isolated point.

Decision Tree algorithms, although computing velocity is fast, operand is little, and when classification is too many, mistake increases than very fast.

Neural network algorithm, has comparatively faster pace of learning, but its Generalization Capability is not but very desirable.

Very fast learning machine algorithm is a kind of Novel learning Method and kit for, has that solving speed is fast, study precision high, but its learning ability is but based upon in the consumption of a large amount of hidden node.

SVM, support vector machine is structure based principle of minimization risk, ensure that good generalization ability.Therefore, support vector machine is widely used in fields such as pattern-recognition, regretional analysis, Function Estimation, time series forecastings.But from support vector machine theoretical analysis, support vector cassification lineoid is easily subject to the impact of noise spot and isolated point, thus cause the degree of accuracy reducing classification.

From upper surface analysis, all there is respective Pros and Cons in various sorting technique, investigation and analysis display, at present, for sensor-based Activity recognition, and the innovatory algorithm using SVM and SVM more.

Summary of the invention

For problem above-mentioned, the present invention adopts a kind of Activity recognition method based on fuzzy support vector machine, and fuzzy theory and support vector machine theory are merged, object is that stress release treatment point and isolated point are on the impact of classifying quality, improves the degree of accuracy of classification.

For achieving the above object, particular content of the present invention comprises:

S1, employing three axis accelerometer image data, extract eigenwert to resultant acceleration, eigenwert comprises: the correlation coefficient in average, variance, energy, three-dimensional data between any two-dimensional data, is expressed as follows S={s ₁, s ₂..., s _n, s={t ₁, t ₂, t ₃, t ₄, t ₅, t ₆, in formula, n represents total sample points, and can be normalized the eigenwert gathered, and eliminates the problem of ' large number eats decimal ';

S2, calculate each Lei Lei center, computing formula is as follows l represents the sample points of this class, when calculating sample point in each class to the Euclidean distance at class center, need get the degree of membership of each sample point to generic according to Euclidean distance calculation.Its Euclidean distance computing formula is as follows:

$D_{i(a,s_i)}={\left[\underset{j=1}{\overset{6}{\mathrm{\Σ}}}{|t_{\mathrm{ij}}^{\′}-t_j^{\′}|}^{\mathrm{\λ}}\right]}^{1-\mathrm{\λ}}$

When adopting the computing method based on the degree of membership of Affinity Among Samples, the calculation access method of the tight ness rating between sample is as follows:

$\mathrm{\φ}(R,a,\mathrm{\ξ})=R^2+C\left[\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i\right]$

D _i ²≤R ²+ξ _i,i＝1,…,n

S.t

ξ _i≥0,i＝1,…,n

In formula, a is class center, and C is penalty factor, and goes the size of R to represent tight ness rating between sample by calculating; Calculate degree of membership by tight ness rating, its degree of membership is calculated and is gone method as follows:

${\mathrm{\μ}}_{\left(x_i\right)}=\left\{\begin{array}{cc}0.6*\left[\frac{1-D_i/R}{1+D_i/R}\right]+0.4,& D_i\≤R\\ 0.4*\left[\frac{1}{1+D_i-R}\right],& D_i>R\end{array}\right.$

S3, employing fuzzy support vector machine realize classification and identify, its detailed process pretreated sample set is carried out classification based training to fuzzy support vector machine obtain a disaggregated model, in the ONLINE RECOGNITION stage, the sample point collected is input to disaggregated model and obtains classification results.

Particularly, described resultant acceleration has X-axis, Y-axis, Z axis acceleration calculation to get, and computing formula is as follows:

$\mathrm{AA}=\sqrt{a_x^2+a_y^2+a_z^2}$

Calculate resultant acceleration, the acceleration in space is synthesized a vector, the impact brought in direction can be ignored thus, contribute to reducing computing.

Particularly, eigenwert is extracted to resultant acceleration, comprising: the correlation coefficient in average, variance, energy, three-dimensional data between any two-dimensional data.Adopt overlapping moving window that continuous print sensor data stream is divided into the entity of regular length for data acquisition.

Mean value computation formula:

$A\stackrel{\‾}{A}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{AA}}_i$

Variance computing formula:

$V_{\mathrm{AA}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{({\mathrm{AA}}_i-A\stackrel{\‾}{A})}^2$

Energy balane formula:

${\mathrm{AA}}_{\mathrm{energy}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{AA}}_i^2$

Correlation coefficient computing formula:

$\mathrm{corr}(x,y)=\frac{\mathrm{cov}(x,y)}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Particularly, for feature set, be normalized, eliminate the impact of " large number eats decimal " problem, all eigenwerts are mapped to 0 to 1 interval, its normalized formula is as follows:

$t_i^{\′}=\frac{t_i-t_{\min }}{t_{\max }-t_{\min }}$

Particularly, for the training of fuzzy support vector machine, adopt multistratum classification model: first, by behavior be divided into motionless (stand still, the sit quietly) class of geo-stationary and action class, build one and fuzzy classification model of classifying; Secondly, normal behaviour in action class and abnormal behaviour are carried out to the structure of disaggregated model, abnormal behaviour is mainly fallen the identification of behavior; Many disaggregated models are finally adopted to be divided into by normal behaviour: to stand, walk, run, stair activity.Being configured with of such disaggregated model helps to stand and the identification of behavior of falling to sitting, for a long time, is conveniently connected mutually with post-service.

Beneficial effect

1, the present invention adopts multistratum classification Construction of A Model, highlights falling and the complexity first identifying, effectively reduce algorithm of the behavior such as sitting.

2, fuzzy factor is incorporated in support vector machine by the present invention, builds fuzzy support vector machine, contributes to stress release treatment point and isolated point to the impact of classifying quality, improves the accurate rate to Activity recognition.

Described on end, the present invention adopts three axis accelerometer image data, and data are synthesized, sextuple eigenwert is extracted to the acceleration of synthesis, the membership function of S type function is adopted to calculate the degree of membership of sample, degree of membership is introduced support vector machine, builds fuzzy support vector machine, adopt fuzzy support vector machine to realize the structure of multistratum classification model.Main object is to eliminate in sample point the noise spot and isolated point that exist to the impact of classifying quality, improves the accurate rate of Activity recognition.

Accompanying drawing explanation

Accompanying drawing 1 is the Activity recognition process flow diagram based on fuzzy support vector machine.

Accompanying drawing 2 is schematic diagram of multistratum classification model.

Embodiment

Below in conjunction with Figure of description, the invention will be further elaborated.

A kind of Activity recognition method based on fuzzy support vector machine of the present invention, mainly comprises:

S1, employing three axis accelerometer image data, extract eigenwert to resultant acceleration, eigenwert comprises: the correlation coefficient in average, variance, energy, three-dimensional data between any two-dimensional data, is expressed as follows S={s ₁, s ₂..., s _n, s={t ₁, t ₂, t ₃, t ₄, t ₅, t ₆, in formula, n represents total sample points, and is normalized the eigenwert gathered, and eliminate the problem of ' large number eats decimal ', its resultant acceleration is calculated as follows:

$\mathrm{AA}=\sqrt{a_x^2+a_y^2+a_z^2}$

A in formula _x, a _y, a _zthe acceleration of the X, Y, Z axis that three axis accelerometer gathers respectively.

Normalized formula is as follows:

$t_i^{\′}=\frac{t_i-t_{\min }}{t_{\max }-t_{\min }}$

S2, calculate each Lei Lei center, computing formula is as follows l represents the sample points of this class, and when calculating sample point in each class to the Euclidean distance at class center, need get the degree of membership of each sample point to generic according to Euclidean distance calculation, its Euclidean distance computing formula is as follows:

$D_{i(a,s_i)}={\left[\underset{j=1}{\overset{6}{\mathrm{\Σ}}}{|t_{\mathrm{ij}}^{\′}-t_j^{\′}|}^{\mathrm{\λ}}\right]}^{1-\mathrm{\λ}}$

S3, employing fuzzy support vector machine realize classification and identify, its detailed process pretreated sample set is carried out classification based training to fuzzy support vector machine obtain a disaggregated model, in the ONLINE RECOGNITION stage, the sample point collected is input to disaggregated model and obtains classification results.

For the training of fuzzy support vector machine, adopt multistratum classification model: first, by behavior be divided into motionless (stand still, the sit quietly) class of geo-stationary and action class, build one and fuzzy classification model of classifying; Secondly, normal behaviour in action class and abnormal behaviour are carried out to the structure of disaggregated model, abnormal behaviour is mainly fallen the identification of behavior; Many disaggregated models are finally adopted to be divided into by normal behaviour: to stand, walk, run, stair activity.

When adopting the computing method based on the degree of membership of Affinity Among Samples, the calculation access method of the tight ness rating between sample is as follows:

$\mathrm{\φ}(R,a,\mathrm{\ξ})=R^2+C\left[\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i\right]$

$S.t\left\{\begin{array}{c}{D_i}^2\≤R^2+{\mathrm{\ξ}}_i,i=1,...,n\\ {\mathrm{\ξ}}_i\≥0,i=1,...,n\end{array}\right.$

In formula, a is class center, and C is penalty factor.

Going the size of R to represent tight ness rating between sample by calculating, calculating degree of membership by tight ness rating, its degree of membership is calculated and is gone method as follows:

${\mathrm{\μ}}_{\left(x_i\right)}=\left\{\begin{array}{cc}0.6*\left[\frac{1-D_i/R}{1+D_i/R}\right]+0.4,& D_i\≤R\\ 0.4*\left[\frac{1}{1+D_i-R}\right],& D_i>R\end{array}\right.$ 。

Claims (5)

1. based on an Activity recognition method for fuzzy support vector machine, it is characterized in that, described method comprises:

S1, employing three axis accelerometer image data, extract eigenwert to resultant acceleration, eigenwert comprises: the correlation coefficient in average, variance, energy, three-dimensional data between any two-dimensional data, is expressed as follows S={s ₁, s ₂..., s _n, s={t ₁, t ₂, t ₃, t ₄, t ₅, t ₆, in formula, n represents total sample points, and is normalized sample characteristics collection;

S2, calculate each Lei Lei center, computing formula is as follows l represents the sample points of this class, when calculating sample point in each class to the Euclidean distance at class center, need get the degree of membership of each sample point to generic according to Euclidean distance calculation;

S3, employing fuzzy support vector machine realize classification and identify.

2. a kind of Activity recognition method based on fuzzy support vector machine according to claim 1, it is characterized in that, adopt three axis accelerometer image data, feature extraction is carried out to resultant acceleration, the eigenwert extracted includes: the correlation coefficient in average, variance, energy, three-dimensional data between any two-dimensional data, and the eigenwert obtained is normalized, the formula of normalized is as follows:

。

3. a kind of Activity recognition method based on fuzzy support vector machine according to claim 1, it is characterized in that, calculate each Lei Lei center, calculate the Euclidean distance at sample point and the class center of getting in each class respectively, calculate according to Euclidean distance the degree of membership that each sample point is under the jurisdiction of this class, its Euclidean distance computing formula is as follows:

In formula, λ gets 2, and when adopting the computing method based on the degree of membership of Affinity Among Samples, the calculation access method of the tight ness rating between sample is as follows:

In formula, a is class center, and C is penalty factor, and goes the size of R to represent tight ness rating between sample by calculating; Calculate degree of membership by tight ness rating, degree of membership is calculated and is gone method as follows:

。

4. a kind of Activity recognition method based on fuzzy support vector machine according to claim 1, it is characterized in that, pretreated sample set is carried out classification based training to fuzzy support vector machine and obtains a disaggregated model, in the ONLINE RECOGNITION stage, the sample point collected is input to disaggregated model and obtains classification results.

5. a kind of Activity recognition method based on fuzzy support vector machine according to claim 1, it is characterized in that, for the training of fuzzy support vector machine, adopt multistratum classification model: first, by behavior be divided into motionless (stand still, the sit quietly) class of geo-stationary and action class, build one and fuzzy classification model of classifying; Secondly, normal behaviour in action class and abnormal behaviour are carried out to the structure of disaggregated model, abnormal behaviour is mainly fallen the identification of behavior; Finally, many disaggregated models are adopted to be divided into by normal behaviour: to stand, walk, run, stair activity.