CN105868779B - A kind of Activity recognition method based on feature enhancing and Decision fusion - Google Patents
A kind of Activity recognition method based on feature enhancing and Decision fusion Download PDFInfo
- Publication number
- CN105868779B CN105868779B CN201610182598.XA CN201610182598A CN105868779B CN 105868779 B CN105868779 B CN 105868779B CN 201610182598 A CN201610182598 A CN 201610182598A CN 105868779 B CN105868779 B CN 105868779B
- Authority
- CN
- China
- Prior art keywords
- axis
- sensor data
- acceleration sensor
- axis acceleration
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
A kind of Activity recognition method based on feature enhancing and Decision fusion, pre-processes acceleration transducer data；Calculate the multidimensional characteristic vector of data after pre-processing；Multidimensional characteristic vector is selected using forward sequence selection algorithm to obtain best features vector；Feature enhancing processing is carried out using Relief-F algorithms selection suitable characteristics value；One fundamental classifier of training and three Weak Classifiers；Using the recognition result of fundamental classifier and three Weak Classifiers as four parameters, while using the recognition result of last movement as another parameter, the weight of this five parameters is determined by training；Human body behavior is identified using two layers of classification, first layer classification is carried out using fundamental classifier, and second layer classification obtains final recognition result using Nearest Neighbor with Weighted Voting Decision fusion.The present invention can distinguish away effectively, upstairs, downstairs these similar movements being difficult to differentiate between, improve Human bodys' response rate.
Description
Technical field
The present invention relates to the fields such as 3-axis acceleration sensor, pattern-recognition, more particularly to are based on 3-axis acceleration data
Human bodys' response field.
Background technique
Carrying out Human bodys' response using sensor is always sensing data processing, the research of area of pattern recognition heat
Point.The daily behavior that people is detected by 3-axis acceleration sensor has the characteristics that convenient, discrimination is high.But acceleration
The case where data are often difficult to differentiate between there is various motion data is especially away, upstairs, downstairs these three act very phases
Seemingly, very small even if the difference under high precision collecting frequency between data.Once data set becomes larger, between these three movements
The existing data that interfere with each other are with regard to more, and since data acquisition platform might not be fixed on human body, so acquisition
Data inevitably have because of the error that shake generates, and in extreme circumstances, there may be undistinguishables for the data of these three movements
The case where, this is correctly identified that these three human actions bring very big difficulty.
Summary of the invention
In order to overcome existing action identification method due to 3-axis acceleration data similitude caused by action recognition just
True rate is low, the big deficiency of identification error, and the present invention proposes a kind of Activity recognition method based on feature enhancing and Decision fusion, can
Effectively distinguish, upstairs, downstairs these similar movements for being difficult to differentiate between, improve Human bodys' response rate.
The technical solution adopted by the present invention to solve the technical problems is:
A kind of Activity recognition method based on feature enhancing and Decision fusion, comprising the following steps:
Step 1, collected acceleration transducer data are pre-processed, including noise processed and data segmentation；
Step 2, sensing data characteristic value after pre-processing is calculated, the multidimensional characteristic arrow for characterizing the sensing data is obtained
Amount；
Step 3, the characteristic vector that step 2 obtains is selected to feature selection approach with before sequence, obtains characterization and passes
The best features vector of sensor data；
Step 4, feature enhancing is carried out using Relief-F algorithms selection characteristic value to the best features vector that step 3 obtains
Processing, process are as follows:
The weight of each characteristic value, given threshold k are calculated using Relief-F algorithm, weight is higher than the characteristic value of the threshold value
Feature enhancing is carried out, four characteristic values is finally selected, is respectively: the difference of two squares of the 3-axis acceleration sensor data in z-axis, three axis
Related coefficient between acceleration transducer data x-axis and y-axis is related between 3-axis acceleration sensor data x-axis and z-axis
Coefficient, the related coefficient between 3-axis acceleration sensor data y-axis and z-axis carry out feature enhancing processing using formula (1),
Wherein piIt is characterized value, m is characterized enhancing coefficient；
Step 5, data four support vector machine classifiers of training obtained using step 4, including a fundamental classifier
With three according to AdaBoost algorithm idea obtain be directed to away, upstairs, downstairs three movement Weak Classifiers；
Step 6, four classifiers obtained using step 5 to currently the movement identified being needed to identify, are obtained respectively
Four recognition results as four parameters, the recognition result that upper one is acted is as the 5th parameter, by this five parameters
It is denoted as label respectivelyi, i=1,2,3,4,5, the weight w of this five parameters is determined by trainingi, i=1,2,3,4,5；
Step 7, human body behavior act is identified using two layers of classification, the basis point that first layer is obtained using step 5
Class device identification, if recognition result be walk, upstairs, downstairs just carry out second layer classification, the second layer classification to described in step 6
Five parameters are weighted ballot Decision fusion, obtain final recognition result.
Further, in the step 2, the process of multidimensional characteristic vector is calculated are as follows: calculate 3-axis acceleration sensor data
In the average value of x-axis, 3-axis acceleration sensor data are in the average value of y-axis, and 3-axis acceleration sensor data are in the flat of z-axis
Mean value, the difference of two squares of the 3-axis acceleration sensor data in x-axis, the difference of two squares of the 3-axis acceleration sensor data in y-axis, three axis
The difference of two squares of the acceleration transducer data in z-axis, the mould of 3-axis acceleration sensor statistical average, 3-axis acceleration sensor
The mould of the data difference of two squares, 3-axis acceleration sensor data are in the degree of bias of x-axis, and 3-axis acceleration sensor data are in the inclined of y-axis
Degree, the degree of bias of the 3-axis acceleration sensor data in z-axis, root mean square of the 3-axis acceleration sensor data in x-axis, the acceleration of three axis
Spend root mean square of the sensing data in y-axis, root mean square of the 3-axis acceleration sensor data in z-axis, 3-axis acceleration sensor
Energy of the data in x-axis, energy of the 3-axis acceleration sensor data in y-axis, energy of the 3-axis acceleration sensor data in z-axis
Amount, entropy of the 3-axis acceleration sensor data in x-axis, entropy of the 3-axis acceleration sensor data in y-axis, 3-axis acceleration sensing
Entropy of the device data in z-axis, related coefficient between 3-axis acceleration sensor data x-axis and y-axis, 3-axis acceleration sensor number
According to the related coefficient between x-axis and z-axis, related coefficient between 3-axis acceleration sensor data y-axis and z-axis, signal amplitude
Region.
Further, in the step 6, the process of weight training are as follows: setup parameter alpha=1/num, num are first
Walk, upstairs, downstairs three movement training samples sum, set the initial value of weight: wi=1, i=1,2,3,4,5, weight instruction
Fundamental classifier recognition result label is only updated during practicing1Weight w1With last action recognition result label5Weight
w5, the weight setting of other three Weak Classifier recognition results is 1 constant, i.e. w2=1, w3=1, w4=1 is constant；Using step 7
In method training sample is identified, if sample it is correct movement be activity, one sample of every identification just uses public affairs
Formula (2) is updated weight, the n=1 when specimen discerning result is consistent with sample concrete class, otherwise n=8, and weight updates
After to w5It is formula (3) processing, PtA.tAIndicate walk, upstairs, downstairs three movement be not transfer to this three movement except its
The probability that it is acted；
w5=w5*(1/PtA.tA) (3)。
Further, in the step 7, Nearest Neighbor with Weighted Voting Decision fusion process are as follows: to five parameter labeli, i=1,2,
3,4,5 using formula (4) calculate leave, upstairs, downstairs three movement scores, wherein activity_n indicate current action,
Score_n indicates the representative score acted of activity_n, PA.AIndicate the probability that a upper movement remains unchanged, PlastA.ATable
Show label5Affiliated movement is transferred to the probability of current action, and the highest action classification of final score is recognition result；
Technical concept of the invention are as follows: the present invention improves the discrimination of similar movement by two methods: first is that by
Characteristic vector that sensing data is calculated carries out feature enhancing, enhancing be conducive to identification walk, upstairs, these three movements downstairs
Characteristic value.Second is that first layer classification distinguishes the movement of easy differentiation using two layers of classification and Decision fusion, will be not easy
The movement of differentiation carries out second layer classification.Using the recognition result of first layer as a parameter in second layer classification, by basis
The recognition result that three Weak Classifiers that AdaBoost algorithm idea trains obtain is as three parameters, by the knowledge of last movement
Other result is determined the weight of this five parameters by training, is determined to this five parameters using Nearest Neighbor with Weighted Voting as another parameter
Plan merges to obtain final recognition result.The Activity recognition method based on feature enhancing and Decision fusion can effectively improve similar
The correct recognition rata of movement.
Beneficial effects of the present invention are mainly manifested in: the characteristic value beneficial to identification similar movement carries out feature enhancing, leads to
Cross Nearest Neighbor with Weighted Voting Decision fusion and obtain final recognition result, can effective district divide similar movement, greatly improve identification similar movement
Correct recognition rata.
Detailed description of the invention
Fig. 1 is the flow chart of the Activity recognition method the present invention is based on feature enhancing and Decision fusion.
Fig. 2 is human body behavior act transition diagram.
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and examples.
Fig. 1 is the main-process stream of the Activity recognition method based on feature enhancing and Decision fusion.
The present embodiment is implemented on Android mobile phone platform, does training and test using the sensing data independently acquired.Altogether
Acquire the walking, run of 10 people, upstairs, downstairs, five action datas of standing, each movement acquisition time is 10 minutes, thus
Share 500 minutes data, sample frequency 100Hz.
Collected 3-axis acceleration data need first to carry out both sides processing before pre-processing: being on the one hand due to three
The data of axle acceleration sensor acquisition include gravity, in order to reduce influence of the gravity to acceleration information, need to subtract
Weight component on sensing data.On the other hand be because the coordinate system where mobile phone relative to mobile phone be it is fixed, relatively
It can but change with the variation in mobile phone orientation in earth coordinates, in order to which behavior identifying system can be identified, mobile phone is in various
Behavior act when orientation needs that mobile phone coordinate system is converted to earth coordinates in real time.
Data prediction includes noise processed and data segmentation, since the data of acquisition are one section of continuous data flows, is
Feature extraction and identification model training, pretreatment after convenience need for data to be divided into 6 seconds window sizes, 50% data
The data slot of overlapping.
After data prediction, need to calculate the characteristic value of data, main includes the feature of time domain and frequency domain, and the present invention has altogether
24 characteristic values are calculated from the acceleration information window of 6 seconds 50% data overlaps, comprising: 3-axis acceleration sensor data
In the average value of x-axis, 3-axis acceleration sensor data are in the average value of y-axis, and 3-axis acceleration sensor data are in the flat of z-axis
Mean value, the difference of two squares of the 3-axis acceleration sensor data in x-axis, the difference of two squares of the 3-axis acceleration sensor data in y-axis, three axis
The difference of two squares of the acceleration transducer data in z-axis, the mould of 3-axis acceleration sensor statistical average, 3-axis acceleration sensor
The mould of the data difference of two squares, 3-axis acceleration sensor data are in the degree of bias of x-axis, and 3-axis acceleration sensor data are in the inclined of y-axis
Degree, the degree of bias of the 3-axis acceleration sensor data in z-axis, root mean square of the 3-axis acceleration sensor data in x-axis, the acceleration of three axis
Spend root mean square of the sensing data in y-axis, root mean square of the 3-axis acceleration sensor data in z-axis, 3-axis acceleration sensor
Energy of the data in x-axis, energy of the 3-axis acceleration sensor data in y-axis, energy of the 3-axis acceleration sensor data in z-axis
Amount, entropy of the 3-axis acceleration sensor data in x-axis, entropy of the 3-axis acceleration sensor data in y-axis, 3-axis acceleration sensing
Entropy of the device data in z-axis, related coefficient between 3-axis acceleration sensor data x-axis and y-axis, 3-axis acceleration sensor number
According to the related coefficient between x-axis and z-axis, related coefficient between 3-axis acceleration sensor data y-axis and z-axis, signal amplitude
Region.
Some characteristic values may include redundancy or unrelated information, these information will affect identification accuracy, therefore need
Feature selecting is carried out, not only can effectively improve identification accuracy in this way, and computation complexity can be reduced and simplify instruction
Practice model.Present invention employs the selection algorithm before sequence to feature selecting (SFS) algorithm as character subset, final choice goes out
For 11 characteristic values as optimal feature subset, this 11 characteristic values include: 3-axis acceleration sensor data being averaged in x-axis
Value, 3-axis acceleration sensor data are in the average value of y-axis, and in the difference of two squares of z-axis, three axis add 3-axis acceleration sensor data
The mould of velocity sensor statistical average, the degree of bias of the 3-axis acceleration sensor data in y-axis, 3-axis acceleration sensor data
In the degree of bias of z-axis, 3-axis acceleration sensor data are in the root mean square of x-axis, and 3-axis acceleration sensor data are in the square of y-axis
Root, the related coefficient between 3-axis acceleration sensor data x-axis and y-axis, 3-axis acceleration sensor data x-axis and z-axis it
Between related coefficient, the related coefficient between 3-axis acceleration sensor data y-axis and z-axis.
After selected character subset, because each characteristic value is different to the contribution of action recognition in selected feature, institute
Differentiation is walked with the present invention, upstairs, downstairs three biggish characteristic values of movements contribution carry out feature enhancing processing.It uses
Relief-F algorithm calculates the weight of each characteristic value, given threshold k=0.05, and the characteristic value that weight is higher than the threshold value carries out special
Sign enhancing, finally selectes four characteristic values, is respectively: the difference of two squares of the 3-axis acceleration sensor data in z-axis, 3-axis acceleration
Related coefficient between sensing data x-axis and y-axis, the related coefficient between 3-axis acceleration sensor data x-axis and z-axis,
Related coefficient between 3-axis acceleration sensor data y-axis and z-axis carries out feature enhancing processing using formula (1), wherein pi
It is characterized value, it is 1.6 that m, which is characterized enhancing coefficient value,；
Classifier used in the present invention is support vector machines, needs four classifiers of training, including a basis altogether
Classifier and three Weak Classifiers obtained according to AdaBoost algorithm idea.Fundamental classifier complete to walk, run, upstairs, under
Building, five identifications acted of standing.Three Weak Classifiers are specific to away, upstairs, downstairs three action trainings, it is each weak
The effect of classifier is that guarantee is very high to some action recognition rate, without guaranteeing the discrimination to other movements.AdaBoost
Algorithm itself is to obtain Weak Classifier one by one by changing data distribution, to be directed to the weak typing of " walking " this action training
For device, need to do in Weak Classifier training process is exactly constantly to find out to act identification " walking " from original training data
Advantageous sample is deleted to the noisy sample of " walking " action recognition, is finally obtained one and is advised exclusively for " walking " this movement
The training sample marked.
Fig. 2 is human body behavior act transition diagram, here by except walk, upstairs, downstairs in addition to movement and be other dynamic
Make, as can be seen from Figure 2 the behavior act of people can remain unchanged, can also be converted to other movements, and convert only with it is upper
A behavior act is unrelated in relation to and with behavior act before, therefore human body behavior act can be converted and regard a Ma Er as
Can husband's process, can be helped to identify current action with the recognition result of last movement using relationship existing between behavior act.
The present invention set the probability matrix converted between human body behavior act asWherein row table
The state successively indicated is: walk, upstairs, downstairs, it is other, the state that list successively indicates is: walk, upstairs, downstairs, it is other, in P
Each element PijExpression state i is transferred to the probability of state j.According to Markov forecast techniques method, the present invention classifies four
Device is to the recognition result of current action as four parameters, and the recognition result that upper one is acted is as the 5th parameter, therefore
There are five parameters altogether, are denoted as label respectivelyi, i=1,2,3,4,5, the weight w of this five parameters is determined by trainingi, i=
1,2,3,4,5。
Weight training method is as follows: setup parameter alpha=1/num first, num walk, upstairs, three movements downstairs
Training sample sum, sets the initial value of weight: wi=1, i=1,2,3,4,5, weight training only updates base categories in the process
Device recognition result label1Weight w1With last action recognition result label5Weight w5, other three Weak Classifiers identifications
As a result weight setting is 1 constant, i.e. w2=1, w3=1, w4=1 is constant；Human body behavior act is known using two layers of classification
Not；If the correct movement of sample is activity, one sample of every identification just uses formula (2) to be updated weight, works as sample
N=1 when this recognition result is consistent with sample concrete class, otherwise n=8, weight update after to w5Formula (3) processing is done,
PtA.tAIndicate walk, upstairs, downstairs three movement be not transfer to this three movement except other movements probability；
w5=w5*(1/PtA.tA) (3)。
Finally using two layers of classification, first layer to walk, run, upstairs, downstairs, five movements of standing identify, if knowledge
Other result is that run or stand be final recognition result, if it is be difficult to differentiate between walk, upstairs, three movements just progress the downstairs
Two layers of classification, the second layer use Nearest Neighbor with Weighted Voting Decision fusion.Detailed process is as follows for Nearest Neighbor with Weighted Voting Decision fusion: to five parameters
labeli, i=1,2,3,4,5 calculated using formula (4) leave, upstairs, the scores of three movements downstairs, wherein activity_n
Indicate that current action, score_n indicate the representative score acted of activity_n, PA.AIndicate what a upper movement remained unchanged
Probability, PlastA.AIndicate label5Affiliated movement is transferred to the probability of current action, and the highest action classification of final score is to know
Other result；
Table 1 is the correct recognition ratas that feature enhancing and Decision fusion method and the method for the present invention five movements of identification is not used
Compare, it can be seen that the method for the present invention can effectively identify similar movement, such as walk, upstairs, downstairs, hence it is evident that improve human body behavior
Discrimination.
Table 1
It is clear that the present invention under being described herein can under the premise of without departing from true spirit and scope of the present invention
With there are many variations.Therefore, all obvious changes some to those skilled in the art, are intended to be included in this power
Within the scope of sharp claim is covered.Scope of the present invention is only defined by described claims.
Claims (4)
1. a kind of Activity recognition method based on feature enhancing and Decision fusion, it is characterised in that: the Activity recognition method packet
Include following steps:
Step 1, collected acceleration transducer data are pre-processed, including noise processed and data segmentation；
Step 2, sensing data characteristic value after pre-processing is calculated, the multidimensional characteristic vector for characterizing the sensing data is obtained；
Step 3, the characteristic vector that step 2 obtains is selected to feature selection approach with before sequence, obtains characterization sensor
The best features vector of data；
Step 4, the best features vector that step 3 obtains is carried out at feature enhancing using Relief-F algorithms selection characteristic value
Reason, process are as follows:
The weight of each characteristic value, given threshold k are calculated using Relief-F algorithm, the characteristic value that weight is higher than the threshold value carries out
Feature enhancing, finally selectes four characteristic values, is respectively: the difference of two squares of the 3-axis acceleration sensor data in z-axis, the acceleration of three axis
Spend the related coefficient between sensing data x-axis and y-axis, the phase relation between 3-axis acceleration sensor data x-axis and z-axis
Number, the related coefficient between 3-axis acceleration sensor data y-axis and z-axis carry out feature enhancing processing using formula (1),
Middle pi is characterized value, and m is characterized enhancing coefficient；
Step 5, data four support vector machine classifiers of training obtained using step 4, including a fundamental classifier and three
It is a according to AdaBoost algorithm idea obtain be directed to away, upstairs, downstairs three movement Weak Classifiers；
Step 6, four classifiers obtained using step 5 are respectively to currently needing the movement identified to identify, four obtained
A recognition result is as four parameters, and the recognition result that upper one is acted is as the 5th parameter, by this five parameter difference
It is denoted as labelj, j=1,2,3,4,5, the weight w of this five parameters is determined by trainingj, j=1,2,3,4,5；
Step 7, human body behavior act is identified using two layers of classification, the fundamental classifier that first layer is obtained using step 5
Identification, if recognition result be walk, upstairs, downstairs just carry out second layer classification, the second layer classification to five described in step 6
Parameter is weighted ballot Decision fusion, obtains final recognition result.
2. a kind of Activity recognition method based on feature enhancing and Decision fusion as described in claim 1, it is characterised in that: institute
It states in step 2, calculates the process of multidimensional characteristic vector are as follows: calculate average value of the 3-axis acceleration sensor data in x-axis, three axis
Average value of the acceleration transducer data in y-axis, average value of the 3-axis acceleration sensor data in z-axis, 3-axis acceleration biography
The difference of two squares of the sensor data in x-axis, the difference of two squares of the 3-axis acceleration sensor data in y-axis, 3-axis acceleration sensor data
In the difference of two squares of z-axis, the mould of 3-axis acceleration sensor statistical average, the mould of the 3-axis acceleration sensor data difference of two squares,
The degree of bias of the 3-axis acceleration sensor data in x-axis, the degree of bias of the 3-axis acceleration sensor data in y-axis, 3-axis acceleration biography
Sensor data are in the degree of bias of z-axis, and 3-axis acceleration sensor data are in the root mean square of x-axis, and 3-axis acceleration sensor data are in y
The root mean square of axis, 3-axis acceleration sensor data z-axis root mean square, 3-axis acceleration sensor data x-axis energy,
Energy of the 3-axis acceleration sensor data in y-axis, energy of the 3-axis acceleration sensor data in z-axis, 3-axis acceleration biography
Sensor data are in the entropy of x-axis, and 3-axis acceleration sensor data are in the entropy of y-axis, and 3-axis acceleration sensor data are in z-axis
Entropy, the related coefficient between 3-axis acceleration sensor data x-axis and y-axis, 3-axis acceleration sensor data x-axis and z-axis it
Between related coefficient, the related coefficient between 3-axis acceleration sensor data y-axis and z-axis, signal amplitude region.
3. a kind of Activity recognition method based on feature enhancing and Decision fusion as claimed in claim 1 or 2, feature exist
In: in the step 6, the process of weight training are as follows: setup parameter alpha=1/num first, num be walk, upstairs, downstairs three
The training sample sum of a movement, sets the initial value of weight: wj=1, j=1,2,3,4,5, weight training only updates in the process
Fundamental classifier recognition result label1Weight w1With last action recognition result label5Weight w5, other three weak point
The weight setting of class device recognition result is 1 constant, i.e. w2=1, w3=1, w4=1 is constant；Using the method in step 7 to training
Sample is identified that, if the correct movement of sample is activity, one sample of every identification just uses formula (2) to carry out weight
It updates, the n=1 when specimen discerning result is consistent with sample concrete class, otherwise n=8, to w after weight update5Do formula
(3) it handles, PtA.tAIndicate walk, upstairs, downstairs three movement be not transfer to this three movement except other movements probability；
w5=w5*(1/PtA.tA) (3)。
4. a kind of Activity recognition method based on feature enhancing and Decision fusion as claimed in claim 1 or 2, feature exist
In: in the step 7, Nearest Neighbor with Weighted Voting Decision fusion process are as follows: to five parameter labelj, j=1,2,3,4,5 use formula
(4) calculate leave, upstairs, downstairs three movement scores, wherein activity_n indicate current action, score_n indicate
The representative score acted of activity_n, PA.AIndicate the probability that a upper movement remains unchanged, PlastA.AIndicate label5It is affiliated
Movement be transferred to the probability of current action, the highest action classification of final score is recognition result；
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610182598.XA CN105868779B (en) | 2016-03-28 | 2016-03-28 | A kind of Activity recognition method based on feature enhancing and Decision fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610182598.XA CN105868779B (en) | 2016-03-28 | 2016-03-28 | A kind of Activity recognition method based on feature enhancing and Decision fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105868779A CN105868779A (en) | 2016-08-17 |
| CN105868779B true CN105868779B (en) | 2018-12-18 |
Family
ID=56624914
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610182598.XA Active CN105868779B (en) | 2016-03-28 | 2016-03-28 | A kind of Activity recognition method based on feature enhancing and Decision fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105868779B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106446876B (en) * | 2016-11-17 | 2019-10-18 | 南方科技大学 | A kind of sensing Activity recognition method and apparatus |
| CN108152059B (en) * | 2017-12-20 | 2021-03-16 | 西南交通大学 | High-speed train bogie fault detection method based on multi-sensor data fusion |
| CN108363490A (en) * | 2018-03-01 | 2018-08-03 | 深圳大图科创技术开发有限公司 | A kind of good intelligent robot system of interaction effect |
| CN108875597B (en) * | 2018-05-30 | 2021-03-30 | 浙江大学城市学院 | Large-scale data set-oriented two-layer activity cluster identification method |
| CN109086698B (en) * | 2018-07-20 | 2021-06-25 | 大连理工大学 | Human body action recognition method based on multi-sensor data fusion |
| CN109726195B (en) * | 2018-11-26 | 2020-09-11 | 北京邮电大学 | Data enhancement method and device |
| CN110569898A (en) * | 2019-09-02 | 2019-12-13 | 河海大学 | Human behavior recognition method |
| CN112084852B (en) * | 2020-08-04 | 2022-08-26 | 河海大学 | Human body track similar behavior identification method based on data fusion |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984921A (en) * | 2014-04-29 | 2014-08-13 | 华南理工大学 | Three-axis feature fusion method used for human movement recognition |
| CN104268577A (en) * | 2014-06-27 | 2015-01-07 | 大连理工大学 | Human body behavior identification method based on inertial sensor |
| CN104899564A (en) * | 2015-05-29 | 2015-09-09 | 中国科学院上海高等研究院 | Human behavior real-time recognition method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI356357B (en) * | 2007-12-24 | 2012-01-11 | Univ Nat Chiao Tung | A method for estimating a body pose |
-
2016
- 2016-03-28 CN CN201610182598.XA patent/CN105868779B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984921A (en) * | 2014-04-29 | 2014-08-13 | 华南理工大学 | Three-axis feature fusion method used for human movement recognition |
| CN104268577A (en) * | 2014-06-27 | 2015-01-07 | 大连理工大学 | Human body behavior identification method based on inertial sensor |
| CN104899564A (en) * | 2015-05-29 | 2015-09-09 | 中国科学院上海高等研究院 | Human behavior real-time recognition method |
Non-Patent Citations (4)
| Title |
|---|
| A framework for human activity recognition based on accelerometer data;Itishree Mandal et al;《2014 5th International Conference- Confluence The Next Generation Information Technology Summit》;20141231;第600-603页 * |
| Activity Recognition with Smartphone Sensors;Xing Su et al;《TSINGHUA SCIENCE AND TECHNOLOGY》;20140630;第19卷(第3期);第235-249页 * |
| 基于三轴加速度传感器的人体动作识别研究;罗初发 等;《工业控制计算机》;20151231;第28卷(第11期);第31-32页 * |
| 基于手机加速度传感器的人体行为识别;衡霞 等;《西安邮电大学学报》;20141130;第19卷(第6期);第76-79页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105868779A (en) | 2016-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105868779B (en) | A kind of Activity recognition method based on feature enhancing and Decision fusion | |
| CN105930767B (en) | A kind of action identification method based on human skeleton | |
| Bujari et al. | Movement pattern recognition through smartphone's accelerometer | |
| WO2018070414A1 (en) | Motion recognition device, motion recognition program, and motion recognition method | |
| CN106096662B (en) | Human motion state identification based on acceleration transducer | |
| CN104123545B (en) | A kind of real-time human facial feature extraction and expression recognition method | |
| CN107908288A (en) | A kind of quick human motion recognition method towards human-computer interaction | |
| CN110245718A (en) | A kind of Human bodys' response method based on joint time-domain and frequency-domain feature | |
| CN109597485B (en) | Gesture interaction system based on double-fingered-area features and working method thereof | |
| CN105912142B (en) | A kind of note step and Activity recognition method based on acceleration sensor | |
| CN105224104B (en) | Pedestrian movement's state identification method based on smart mobile phone grip mode | |
| CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
| CN105930770B (en) | A kind of human motion recognition method based on Gaussian process latent variable model | |
| JP2016062610A (en) | Feature model creation method and feature model creation device | |
| CN110490080A (en) | A kind of human body tumble method of discrimination based on image | |
| CN105373810B (en) | Method and system for establishing motion recognition model | |
| CN103345626A (en) | Intelligent wheelchair static gesture identification method | |
| CN106210269A (en) | A kind of human action identification system and method based on smart mobile phone | |
| CN107392939A (en) | Indoor sport observation device, method and storage medium based on body-sensing technology | |
| CN110796101A (en) | Face recognition method and system of embedded platform | |
| CN106599785A (en) | Method and device for building human body 3D feature identity information database | |
| CN113065505A (en) | Body action rapid identification method and system | |
| CN102831408A (en) | Human face recognition method | |
| CN110388926A (en) | A kind of indoor orientation method based on mobile phone earth magnetism and scene image | |
| CN113663312A (en) | Micro-inertia-based non-apparatus body-building action quality evaluation method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |