CN106203484B - A kind of human motion state classification method based on classification layering - Google Patents
A kind of human motion state classification method based on classification layering Download PDFInfo
- Publication number
- CN106203484B CN106203484B CN201610509467.8A CN201610509467A CN106203484B CN 106203484 B CN106203484 B CN 106203484B CN 201610509467 A CN201610509467 A CN 201610509467A CN 106203484 B CN106203484 B CN 106203484B
- Authority
- CN
- China
- Prior art keywords
- sample
- classification
- action
- sample action
- movement
- 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
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Abstract
A kind of human motion state classification method based on classification layering, constructs with classification mark, sample action data prediction, sample action feature extraction, the feature selecting based on sample action data, classifier based on classification layering comprising different classes of movements design and sample action acquisition, the segmentation of different classes of sample action and classification method validation verification.The present invention is using classification layering as core, it is intended to reduce in assorting process due to classifier is influenced by other classification training samples and the phenomenon that judged by accident to test sample classification, achieve the purpose that improve human motion state recognition accuracy and recall rate.The present invention can be used as the core classification method of any human motion state identification, have stronger versatility and portability.
Description
Technical field
The present invention relates to motion state monitoring technical field, in particular to a kind of human motion based on classification layering
State classification method, it is intended to reduce in assorting process since classifier is influenced by other classification training samples and to test sample
The phenomenon that classification is judged by accident achievees the purpose that improve human motion state recognition accuracy and recall rate.The present invention, which can be used as, to be appointed
A kind of what core classification method of human motion state identification, has stronger versatility and portability.
Background technique
In today's society, the increasing of the accelerating rhythm of life and operating pressure is so that more and more people are in inferior health
State.People's health status for also just increasingly paying close attention to oneself in this way, take various measures to the health status for improving oneself, than
Such as start to adjust the work and rest rhythm of oneself, rational diet, moderately do various movements.In the various measures for improving health status
In, movement is a very important measure.Movement appropriate can enhance the metabolism of human body, mould perfect body
State helps people to exclude unhealthy emotion.In consideration of it, company's (such as Jawbone, millet, happy heart) releases independently one after another both at home and abroad
The motion monitoring equipment based on 3-axis acceleration and three-axis gyroscope sensor of research and development, it is intended to human motion state (such as row
Walk, run, sleep) real-time monitoring is carried out, and then record human body daily routines amount and provide a user feedback, help people's tune
Itself whole daily schedule achievees the purpose that improve physical condition.
In addition, can also be played to the clinical decision in medical field to the real-time monitoring of human body behavior pattern auxiliary well
Helping property acts on.For example, correlative study shows that physical activity level and depression have stronger association in depression Quantified therapy field
Property, human body will appear behavioral disorder phenomenon in the case where being influenced by depression, i.e. human body behavioural characteristic can be used as one kind
The effectively quantizating index of reflection patients with depression behavioral disorder.The mood reflection of decline, the body movement weakened and social functions
The behavioural characteristics such as obstacle have been concluded as depressed feature.In fact, above-mentioned behavior symptom can be viewed as by major depressive disorder
The caused blunt behavior presentation of self-discipline.In addition, psychomotor slow (being defined as slowing down for thinking, language and movement) is also suppression
A kind of one of strongly fragrant main symptom of disease, can be used for diagnosing depression, differentiation depression hypotype and progress curative effect evaluation, (drug is commented
Estimate).Specifically, it paces back and forth (touch turn), sitting (seat), lie and (lie) long, going in the interior for including in human body behavior pattern
The Clinical symptoms that all can serve as patients with depression for behaviors such as slow (walkings, run), in conjunction with other physiological phenomenons (such as palpitaition,
Perspire etc.), it can effectively reflect the clinical symptoms of patients with depression.In consideration of it, pace back and forth to daily routines amount, interior,
The behaviors such as sitting, behavior be slow carry out identification operation and are of great significance in terms of depression clinical symptoms are monitored with diagnosis.
However, the motion monitoring equipment sold currently on the market can only simple people several to running, walking and sleep etc.
Body motion state is effectively identified, the real-time monitoring demand of the human body behavior pattern comprising various motion classification is unable to satisfy.
In addition, the related human body behavior pattern recognition based on 3-axis acceleration and three-axis gyroscope sensor is the study found that certain is several dynamic
The data characteristics made classification (as sat and standing, from station and slave station is sat on to seat) has similitude.Simultaneously as different subjects
Between human body behavior pattern have differences, there is also data characteristicses for the sample actions of same action classifications that difference subject generates
Inconsistent situation.Above situation causes the decline of the human body behavior pattern recognition accuracy comprising plurality of classes.Therefore, a kind of
Human motion state classification method with good robustness is to the human body based on 3-axis acceleration and three-axis gyroscope sensor
Behavior pattern recognition research is most important.
Summary of the invention
It is an object of the invention to know for based on the human body behavior pattern of 3-axis acceleration and three-axis gyroscope sensor
Do not study, propose a kind of human motion state classification method based on classification layering, it is intended to reduce in assorting process due to
The phenomenon that classifier is influenced by other classification training samples and judged by accident to test sample classification, reaches raising human motion state
The purpose of recognition accuracy and recall rate.
To achieve the above object, the technical solution adopted by the present invention is that: a kind of human motion based on classification layering
State classification method includes different classes of movements design and sample action acquisition, the segmentation of different classes of sample action and classification mark
Note, sample action data prediction, sample action feature extraction, the feature selecting based on sample action data, based on classification point
The classifier building of layer mechanism and classification method validation verification step.
The different classes of movements design and sample action acquisition refer to human motion state is defined as lying, is sat, is stood,
Walk, hurry up, running, upstairs, downstairs, turn round, lie to sitting, sitting on station, standing to sitting, sitting on 13 kinds of different classes of movements of lying, and design
Experimental paradigm comprising the movement of above-mentioned classification, acquires the sample action of respective classes.The sample action of acquisition is by 3-axis acceleration
Data and three axis angular rate data composition, i.e., any data comprising 3-axis acceleration sensor and three-axis gyroscope sensor are adopted
Collection equipment is suitable for the sample action acquisition, and the data acquisition equipment is smart phone, movement sensing module.
The different classes of sample action segmentation marked with classification refer to acted in the experimental paradigm according to design it is successive
Sequentially, semi-automatic segmentation is carried out to different classes of sample action using sample action segmentation software, and will be dynamic after segmentation
Corresponding sports status categories are labeled as sample.Sample action segmentation software reads in a full experiment normal form every time and collects
Sample action data, according to 3-axis acceleration and three axis angular rates in the sequencing and sample action acted in experimental paradigm
The waveforms of data changes, by the pointer drag operation in software interface, define in experimental paradigm the starting points of different movements and
End point, finally inputs corresponding action classification title again in edit box, click data save button is by starting point and end point
Between sample action data, and then the data for completing different classes of sample action save and classification labeling operation.
The sample action data prediction refers to according to sequencing, successively using adding window overlapping, denoising and normalization
Three kinds of operating methods pre-process the sample action after segmentation, and it is special for subsequent sample to form final sample action
Levy extraction, feature selecting, classifier building and classification method verification operation.The adding window overlap operation refers to using regular length
Time window simultaneously carries out secondary splitting to the sample action after each segmentation by the Duplication of fixed percentage, dynamic after each segmentation
Make sample and is all finally divided into the fixed sample action of one or more length.Wherein, turn round, lie to sit back and wait movement due to occur
Time is shorter, and the sample action data length after segmentation is less than time window length, and the sample action after dividing in this case is adopted
With the strategy of front and back data augmentation, a length and time window sample action consistent in length are obtained.It walks, lie, seat movement hair
The raw time is longer, and the sample action length after segmentation is greater than time window length, in this case simultaneously using regular length time window
Secondary splitting is carried out to the sample action after each segmentation by the Duplication of fixed percentage and obtains multiple length and time window
Mouth sample action consistent in length.Sample action after segmentation is referred to as in the sample action obtained after adding window overlap operation
Sample action after secondary splitting.The denoising of sample action data prediction, which refers to, adopts the sample action after each secondary splitting
Data de-noising is carried out with median filtering denoising method.The normalization of sample action data prediction refers in order to avoid different subjects
Data variance between same action counts each subject according to the collected different classes of sample action of experimental paradigm
According to normalization operation.Specifically, each collected different classes of sample action of subject can obtain one after secondary splitting
A includes the other sample action collection of 13 types, and each sample action length is consistent in sample set.Normalization operation is exactly to each
It is tested collected sample action collection and operation is normalized.
The sample action feature extraction refer to according to time domain, frequency domain Feature Extraction Method respectively to each secondary splitting
The 3-axis acceleration data for including in sample action afterwards and three axis angular rate data carry out time domain and frequency domain character extracts.Into one
Step is said, for each axis motion state data in sliding window, calculates its mean value, standard deviation, the 25th percentile, the 5000th
Related coefficient, frequency spectrum energy and the preceding ten kinds of frequencies of quantile, the 75th percentile, 6 axis motion state datas between any two
Energy time domain and frequency domain data feature.Furtherly, the sample action after feature extraction operation, after each secondary splitting
It can be indicated by a high dimension vector, each sample action collection is made of the high dimensional feature matrix that a m row n is arranged.Wherein, m table
Show that sample action concentrates the sample action quantity for including, n indicates that the intrinsic dimensionality of high dimension vector, that is, the time domain extracted, frequency domain are special
Levy number.
The feature selecting of the sample action data, which refers to, drops high dimensional feature matrix using principal component analytical method
Dimension operation, to accumulate contribution degree greater than 75% for standard, eigenmatrix after obtaining dimensionality reduction and constructing as subsequent classifier is grasped
The data basis of work.The feature selection step that the present invention uses has alternative, that is, chooses whether using feature selection step
Dimensionality reduction operation is carried out to the high dimensional data feature that sample action feature extraction obtains.
Different human motion states is divided into quiet by the characteristics of classification layering is referred to according to human motion state
Only property movement and motility act two kinds.Wherein, motility movement can be subdivided into continuity movement and transformational movement two again
Kind.Usually, inactive movement includes three kinds of human motion states routinely of lying, sit and stand.Continuity movement include walk,
Hurry up, run, upstairs and downstairs five kinds of human motion states routinely, transformational movement include turn round, from lying to sitting, from sitting on
It lies, from station and slave station is sat on to sitting five kinds of human motion states routinely.It is acted according to inactive movement, motility, is continuous
Property movement and transformational act the divisions of four kinds of human motion states, classification layering is defined as follows:
It defines 1, will judge that test sample belongs to the classifying step of inactive movement or motility movement as first layer
Subseries；
It defines 2, will judge that test sample belongs to the classifying step lain, sit or stood as second acted towards inactive
Hierarchical classification；
It defines 3, will judge that test sample belongs to the classifying step of continuity movement or transformational movement as towards fortune
Second hierarchical classification of dynamic property movement；
It defines 4, will judge test sample and belong to walk, hurry up, run, upstairs or the classifying step gone downstairs is as towards continuous
Property movement third level classification；
5 are defined, will judge that test sample belongs to and turn round, lie and sit, sit on station, stand to seat or sit on the classifying step lain
As the third level classification acted towards transformational.
Classifier building based on classification layering refers to meet the classificating requirement of classification layering, all two
Sample action after secondary segmentation all has multiple action classification label.Wherein, the other sample action of three types of lying, sit and stand has
Classification itself and inactive act two categories label.It walks, hurry up, running, the other sample action of five types has upstairs and downstairs
Classification itself, continuity movement and motility act three kinds of class labels.It turns round, lie to sitting, sit on station, stand to sitting and sit on and lie
There is the other sample action of five types classification itself, transformational movement and motility to act three kinds of class labels.For all two
Sample action after secondary segmentation selects a sample action as test sample every time according to a verification method is stayed, remaining dynamic
Make sample as training sample, each class hierarchy is all made of K and carries out classification behaviour to test sample adjacent to classification method (KNN)
Make.The sort operation comprising the following steps:
The test sample of selection is input in the classifier based on classification layering of building by step 1；
Step 2 carries out the first hierarchical classification operation.It is motility that all categories label is taken out from remaining sample action
The sample action of movement and inactive movement calculates test sample and each training using KNN classification method as training sample
Euclidean distance between sample, point of the most classes of the smallest sample action of K Euclidean distance as the first hierarchical classification before taking
Class is as a result, judge that the test sample belongs to inactive movement or motility movement.When classification results are that inactive acts,
Classification process gos to step 3, otherwise, gos to step 4.
The second hierarchical classification that step 3, progress are acted towards inactive.All categories are taken out from remaining sample action
Label is that the sample action lain, sit and stood calculates test sample and each training using KNN classification method as training sample
Euclidean distance between sample, most classes of K Euclidean distance the smallest sample action towards inactive as acting before taking
The classification results of second hierarchical classification judge test sample classification to lie, sitting or stand.
The second hierarchical classification that step 4, progress are acted towards motility.All categories are taken out from remaining sample action
Label is that continuity movement and the sample action of transformational movement calculate test specimens using KNN classification method as training sample
This each training sample between Euclidean distance, before taking most classes of the smallest sample action of K Euclidean distance as towards
The classification results of second hierarchical classification of motility movement judge that test sample classification is dynamic for continuity movement or transformational
Make.When classification results are that continuity acts, classification process gos to step 5, otherwise, gos to step 6.
Step 5 carries out the third level acted towards continuity classification.All categories are taken out from remaining sample action
Label is to walk, hurry up, running, upstairs and sample action downstairs is as training sample, using KNN classification method, calculating test specimens
This each training sample between Euclidean distance, before taking most classes of the smallest sample action of K Euclidean distance as towards
Continuity movement third level classification classification results, that is, judge test sample classification be walk, hurry up, running, upstairs or under
Building.
Step 6 carries out the third level acted towards transformational classification.All categories are taken out from remaining sample action
Label is to turn round, lie and sit, sit on station, stand to seat and sit on the sample action lain as training sample, using the classification side KNN
Method calculates the Euclidean distance between test sample and each training sample, and the smallest sample action of K Euclidean distance is more before taking
It is several classes of as acted towards transformational third level classification classification results, that is, judge test sample classification for turn round, lie to
It sits, sit on station, stand to sitting or sit on and lie.
The classification method validation verification refers to according to the human motion state classification method based on classification layering
Obtained classification results, using accuracy Accuracy, error rate Error rate, sensitivity S ensitive, special efficacy degree
Five indexs of Specificity and precision Precision evaluate the classifying quality of classification method, and five indexs calculate public
Formula is described as follows:
(1) accuracy Accuracy, calculation formula are as follows:
(2) error rate Error rate, calculation formula are as follows:
(3) sensitivity S ensitive, calculation formula are as follows:
(4) special efficacy degree Specificity, calculation formula are as follows:
(5) precision Precision, calculation formula are as follows:
In formula, TP indicates that actual act classification is A and is classified the sample size that device is divided into A class, and FP indicates practical dynamic
Make classification to be non-A and be classified the sample size that device is divided into A class, FN indicates that actual act classification is A and is classified device division
For the sample size of non-A class, TN indicates that actual act classification is non-A and is classified the sample size that device is divided into non-A class.P table
Show actual act classification be A sample size, N indicate actual act classification be non-A sample size, A expression lie, sit, standing,
Walk, hurry up, running, upstairs, downstairs, turn round, lie to sitting, sit on station, stand to sitting and sit on one of the action classifications such as lie.
The present invention has the advantage that and effect is described as follows:
1, using classification layering, the action classification sample of similitude can is divided into other number data characteristics
According under tag class sample.Furtherly, when certain class action classification sample is as test sample, movement class similar with the category
It will not participate in the category classification of the sample, reduce in assorting process since classifier is by other not as training sample
The phenomenon that classification training sample is influenced and is judged by accident to test sample classification.
2, compared with other traditional human motion state classification methods based on machine learning and pattern-recognition, it is based on classification
The human motion state classification method of layering has higher human motion state recognition accuracy and recall rate.
3, the present invention can be used as the core classification method of any human motion state identification.Furtherly, each
It can be practised and algorithm for pattern recognition (such as support vector machines, Piao using the machine generallyd use at present during secondary hierarchical classification
Plain Bayes and neural network etc.), classification layering has stronger versatility and portability.
Detailed description of the invention
Fig. 1 is a kind of human motion state classification method flow chart based on classification layering
Fig. 2 is the classification method schematic diagram based on classification layering
Specific embodiment
Invention is described in further detail below with reference to examples and drawings, it is notable that the present invention
The embodiment of patent is without being limited thereto.
As shown in Figure 1, a kind of human motion state classification method specific implementation step explanation based on classification layering
It is as follows:
Step 1, design include the experimental paradigm of different classes of movement, using containing 3-axis acceleration sensor and three axis
The data acquisition equipment of gyro sensor acquires the sample action of respective classes according to experimental paradigm.The present invention is by human motion
State is defined as lying, sits, stands, walking, hurrying up, running, upstairs, downstairs, turn round, lie to sitting, sit on station, stand to sitting, sit on and lie 13 kinds
Different classes of movement.
Step 2, according to the sequencing acted in the experimental paradigm of design, using sample action segmentation software to inhomogeneity
Other sample action carries out semi-automatic segmentation, and the sample action after segmentation is labeled as corresponding sports status categories.
Step 3 successively uses adding window overlapping, denoising and three kinds of operating methods of normalization to carry out the sample action after segmentation
Pretreatment, and final sample action is formed for subsequent operation.Wherein, the data window length setting in adding window overlap operation
For the integral multiple of data acquisition equipment sample rate, windows overlay rate is set as 50%.Denoising operation using median filter method into
Row denoising.Sample action after each subject segmentation obtains the consistent sample action of multiple data lengths after pretreatment, often
The classification of a sample action belongs to one of 13 action classifications defined in step 1.
Step 4, to the 3-axis acceleration data for including in the sample action after each secondary splitting and three axis angular rate numbers
It is extracted according to progress time domain and frequency domain character.For each axis motion state data in sliding window, its mean value, standard are calculated
Difference, the 25th percentile, the 50th percentile, the 75th percentile, 6 axis motion state datas related coefficient between any two,
The time domains such as frequency spectrum energy and preceding ten kinds of frequency energies and frequency domain data feature.After feature extraction operation, each secondary point
Sample action after cutting can be by a high dimension vector expression, the high dimensional feature square that each sample action collection is arranged by a m row n
Battle array composition.
Step 5 chooses whether to carry out feature selecting operation.If necessary to carry out feature selecting operation, then go to step
6.Otherwise, 7 are gone to step.
Step 6 carries out dimensionality reduction operation to high dimensional feature matrix using principal component analytical method, is greater than with accumulating contribution degree
75% is standard, the eigenmatrix after obtaining dimensionality reduction.
Step 7 increases the sample action after all secondary splittings other action classification labels, and building classifier carries out
Sample action sort operation.In addition to class label itself, the other sample action of three types of lying, sit and stand increases inactive and acts class
Distinguishing label.In addition to class label itself, walks, hurries up, running, the other sample action of five types increases continuity movement upstairs and downstairs
Two categories label is acted with motility.In addition to class label itself, turns round, lies to sitting, sit on station, stand to sitting and sit on and lie five
The other sample action of type increases transformational movement and motility acts two categories label.For dynamic after all secondary splittings
Make sample, according to a verification method is stayed, selects a sample action as test sample every time, remaining sample action is as instruction
Practice sample, each class hierarchy is all made of K and carries out sort operation to test sample adjacent to classification method (KNN).Based on classification point
Layer mechanism classification method schematic diagram as shown in Fig. 2, comprising the following steps:
The test sample of selection is input in the classifier based on classification layering of building by step 71；
Step 72 carries out the first hierarchical classification operation.It is movement that all categories label is taken out from remaining sample action
Property movement and the sample action of inactive movement be used as training sample, using KNN classification method, calculate test sample and each instruction
Practice the Euclidean distance between sample, most classes of the smallest sample action of K Euclidean distance are as the first hierarchical classification before taking
Classification results judge that the test sample belongs to inactive movement or motility movement.When classification results are inactive movement
When, classification process gos to step 73, otherwise, gos to step 74.
The second hierarchical classification that step 73, progress are acted towards inactive.All classes are taken out from remaining sample action
Distinguishing label is that the sample action lain, sit and stood calculates test sample and each instruction using KNN classification method as training sample
Practice the Euclidean distance between sample, most classes of the smallest sample action of K Euclidean distance are used as before taking acts towards inactive
The second hierarchical classification classification results, that is, judge test sample classification to lie, sitting or stand.
The second hierarchical classification that step 74, progress are acted towards motility.All classes are taken out from remaining sample action
Distinguishing label is that continuity movement and the sample action of transformational movement calculate test using KNN classification method as training sample
Euclidean distance between sample and each training sample, most classes of the smallest sample action of K Euclidean distance are as face before taking
The classification results of the second hierarchical classification acted to motility judge test sample classification for continuity movement or transformational
Movement.When classification results are that continuity acts, classification process gos to step 75, otherwise, gos to step 76.
Step 75 carries out the third level acted towards continuity classification.All classes are taken out from remaining sample action
Distinguishing label is to walk, hurry up, running, upstairs with sample action downstairs as training sample, using KNN classification method, calculating test
Euclidean distance between sample and each training sample, most classes of the smallest sample action of K Euclidean distance are as face before taking
To continuity act third level classify classification results, that is, judge test sample classification be walk, hurry up, running, upstairs or
Downstairs.
Step 76 carries out the third level acted towards transformational classification.All classes are taken out from remaining sample action
Distinguishing label is to turn round, lie and sit, sit on station, stand to seat and sit on the sample action lain as training sample, using the classification side KNN
Method calculates the Euclidean distance between test sample and each training sample, and the smallest sample action of K Euclidean distance is more before taking
It is several classes of as acted towards transformational third level classification classification results, that is, judge test sample classification for turn round, lie to
It sits, sit on station, stand to sitting or sit on and lie.
The classification results that step 8, the human motion state classification method based on classification layering obtain calculate accuracy
Accuracy, error rate Error rate, sensitivity S ensitive, special efficacy degree Specificity and precision Precision five
A index simultaneously evaluates the classifying quality of classification method.
Presently preferred embodiments of the present invention is described above.It is to be appreciated that the present invention is the portion of detailed disclosure
Belong to techniques known, i.e., limitation of the invention and above-mentioned particular implementation, wherein be not described in detail
Equipment and structure are construed as being practiced using the common mode in this field；Anyone skilled in the art,
Without departing from the scope of the technical proposal of the invention, all using the methods and technical content of the disclosure above to the technology of the present invention
Scheme makes many possible changes and modifications or equivalent example modified to equivalent change, this has no effect on of the invention
Substantive content.Therefore, anything that does not depart from the technical scheme of the invention, technical spirit according to the present invention is to above embodiments
Any simple modifications, equivalents, and modifications, all of which are still within the scope of protection of the technical scheme of the invention.
Claims (3)
1. a kind of human motion state classification method based on classification layering, it is characterised in that: this method includes inhomogeneity
Other movements design and sample action acquisition, the segmentation of different classes of sample action and classification mark, are moved sample action data prediction
Make sample characteristics extraction, the feature selecting based on sample action data, classifier building and classification based on classification layering
Method validation verification step；
The different classes of movements design and sample action acquisition refer to be defined as human motion state lying, sit, standing, walking, fastly
Walk, run, upstairs, downstairs, turn round, lie to sitting, sit on station, stand to sitting, sit on 13 kinds of different classes of movements of lying, and design comprising upper
The experimental paradigm for stating classification movement, acquires the sample action of respective classes；The sample action of acquisition by 3-axis acceleration data and
Three axis angular rate data composition, i.e., any data acquisition equipment comprising 3-axis acceleration sensor and three-axis gyroscope sensor
It is suitable for the sample action acquisition, the data acquisition equipment is smart phone, movement sensing module；
The different classes of sample action segmentation refers to the sequencing acted in the experimental paradigm according to design with classification mark,
Semi-automatic segmentation is carried out to different classes of sample action using sample action segmentation software, and by the sample action after segmentation
It is labeled as corresponding sports status categories；Sample action segmentation software reads collected movement in a full experiment normal form every time
Sample data, according to 3-axis acceleration in the sequencing and sample action acted in experimental paradigm and three axis angular rate data
Waveform variation defines the starting point and end point of different movements in experimental paradigm by the pointer drag operation in software interface,
Corresponding action classification title is finally inputted in edit box again, click data save button will be dynamic between starting point and end point
The data made sample data, and then complete different classes of sample action save and classification labeling operation；
The sample action data prediction refers to according to sequencing, successively using three kinds of adding window overlapping, denoising and normalization
Operating method pre-processes the sample action after segmentation, and forms final sample action and mention for subsequent sample characteristics
It takes, feature selecting, classifier construct and classification method verification operation；The adding window overlap operation referred to using the regular length time
Window simultaneously carries out secondary splitting to the sample action after each segmentation by the Duplication of fixed percentage, the movement sample after each segmentation
This is all finally divided into the fixed sample action of one or more length；Wherein, turn round, lie to seat movement due to time of origin compared with
Short, the sample action data length after segmentation is less than time window length, and the sample action after dividing in this case uses front and back
The strategy of data augmentation obtains a length and time window sample action consistent in length；It walks, lie, seat movement time of origin
Longer, the sample action length after segmentation is greater than time window length, in this case using regular length time window and by fixation
The Duplication of percentage carries out secondary splitting to the sample action after each segmentation and obtains multiple length and time window length
Consistent sample action；Sample action after segmentation is referred to as secondary point in the sample action obtained after adding window overlap operation
Sample action after cutting；The denoising of sample action data prediction refers to the sample action after each secondary splitting using intermediate value
Filtering and noise reduction method carries out data de-noising；The normalization of sample action data prediction refers in order to avoid different subjects are identical dynamic
Data variance between work carries out data normalizing according to the collected different classes of sample action of experimental paradigm to each subject
Change operation；Specifically, the collected different classes of sample action of each subject can obtain one after secondary splitting and include
The other sample action collection of 13 types, each sample action length is consistent in sample set；Normalization operation is exactly to each by pilot production
Operation is normalized in the sample action collection collected；
The sample action feature extraction refer to according to time domain, frequency domain Feature Extraction Method respectively to each secondary splitting after
The 3-axis acceleration data for including in sample action and three axis angular rate data carry out time domain and frequency domain character extracts；Further
It says, for each axis motion state data in sliding window, calculates its mean value, standard deviation, the 25th percentile, the 50th percentage
Related coefficient, frequency spectrum energy and the preceding ten kinds of frequency energy of digit, the 75th percentile, 6 axis motion state datas between any two
Measure time domain and frequency domain data feature；Furtherly, after feature extraction operation, the sample action after each secondary splitting can
To be indicated by a high dimension vector, each sample action collection is made of the high dimensional feature matrix that a m row n is arranged；Wherein, m is indicated
Sample action concentrates the sample action quantity for including, and n indicates the intrinsic dimensionality of high dimension vector, that is, the time domain extracted, frequency domain character
Number；
The feature selecting of the sample action data, which refers to, carries out dimensionality reduction behaviour to high dimensional feature matrix using principal component analytical method
Make, to accumulate contribution degree greater than 75% for standard, eigenmatrix after obtaining dimensionality reduction and as subsequent classifier building operation
Data basis；The feature selection step of use has alternative, that is, chooses whether using feature selection step to sample action
The high dimensional data feature that feature extraction obtains carries out dimensionality reduction operation；
Different human motion states is divided into inactive by the characteristics of classification layering is referred to according to human motion state
Movement and motility act two kinds；Wherein, motility movement can be subdivided into continuity movement again and transformational acts two kinds；It is logical
For often, inactive movement includes three kinds of human motion states routinely of lying, sit and stand；Continuity movement include walk, hurry up,
Run, upstairs and downstairs five kinds of human motion states routinely, transformational movement include turn round, from lying to sitting, from sit on lie, from
Station and slave station are sat on to sitting five kinds of human motion states routinely；According to inactive movement, motility movement, continuity movement
The division of four kinds of human motion states is acted with transformational, classification layering is defined as follows:
It defines 1, will judge that test sample belongs to the classifying step of inactive movement or motility movement as the first level point
Class；
It defines 2, will judge that test sample belongs to the classifying step lain, sit or stood as the second level acted towards inactive
Classification；
It defines 3, will judge that test sample belongs to the classifying step of continuity movement or transformational movement as towards motility
Second hierarchical classification of movement；
It defines 4, will judge test sample and belong to walk, hurry up, run, upstairs or the classifying step gone downstairs is moved as towards continuity
The third level of work is classified；
Define 5, will judge test sample and belong to turn round, lie sit, sit on station, stand to seat or sit on the classifying step lain as
Third level classification towards transformational movement；
Classifier building based on classification layering refers to meet the classificating requirement of classification layering, all secondary points
Sample action after cutting all has multiple action classification label；Wherein, the other sample action of three types of lying, sit and stand has itself
Classification and inactive act two categories label；It walks, hurry up, running, the other sample action of five types has itself upstairs and downstairs
Classification, continuity movement and motility act three kinds of class labels；It turns round, lie to sitting, sit on station, stand to sitting and sit on and lie five kinds
There is the sample action of classification classification itself, transformational movement and motility to act three kinds of class labels；For all secondary points
Sample action after cutting selects a sample action as test sample, remaining movement sample every time according to a verification method is stayed
, as training sample, each class hierarchy is all made of K and carries out sort operation to test sample adjacent to classification method (KNN) for this.
2. a kind of human motion state classification method based on classification layering according to claim 1, feature exist
In: the sort operation comprising the following steps:
The test sample of selection is input in the classifier based on classification layering of building by step 1；
Step 2 carries out the first hierarchical classification operation；All categories label is taken out from remaining sample action as motility movement
Test sample and each training sample are calculated using KNN classification method as training sample with the sample action of inactive movement
Between Euclidean distance, classification knot of the most classes of K Euclidean distance the smallest sample action as the first hierarchical classification before taking
Fruit judges that the test sample belongs to inactive movement or motility movement；When classification results are that inactive acts, classification
Process gos to step 3, otherwise, gos to step 4；
The second hierarchical classification that step 3, progress are acted towards inactive；All categories label is taken out from remaining sample action
Sample action to lie, sitting and stand calculates test sample and each training sample using KNN classification method as training sample
Between Euclidean distance, most classes of K Euclidean distance the smallest sample action are as second acted towards inactive before taking
The classification results of hierarchical classification judge test sample classification to lie, sitting or stand；
The second hierarchical classification that step 4, progress are acted towards motility；All categories label is taken out from remaining sample action
Sample action for continuity movement and transformational movement is used as training sample, using KNN classification method, calculate test sample and
Euclidean distance between each training sample, most classes of the smallest sample action of K Euclidean distance are used as towards movement before taking
Property movement the second hierarchical classification classification results, i.e., judge test sample classification be continuity act or transformational act；
When classification results are that continuity acts, classification process gos to step 5, otherwise, gos to step 6；
Step 5 carries out the third level acted towards continuity classification；All categories label is taken out from remaining sample action
To walk, hurrying up, run, upstairs and sample action downstairs is used as training sample, using KNN classification method, calculate test sample and
Euclidean distance between each training sample, most classes of the smallest sample action of K Euclidean distance are used as towards continuous before taking
Property movement third level classification classification results, i.e., judge test sample classification be walk, hurry up, running, upstairs or go downstairs；
Step 6 carries out the third level acted towards transformational classification；All categories label is taken out from remaining sample action
It sits to turn round, lying, sit on station, stand to sitting and sitting on the sample action lain as training sample, using KNN classification method, count
The Euclidean distance between test sample and each training sample is calculated, most classes of the smallest sample action of K Euclidean distance before taking
As the classification results of the third level classification acted towards transformational, that is, test sample classification is judged to turn round, lying to seat, sit
It arrives at a station, stand to sitting or sit on and lie；
The classification method validation verification refers to be obtained according to the human motion state classification method based on classification layering
Classification results, using accuracy Accuracy, error rate Error rate, sensitivity S ensitive, special efficacy degree
Five indexs of Specificity and precision Precision evaluate the classifying quality of classification method, and five indexs calculate public
Formula is described as follows:
(1) accuracy Accuracy, calculation formula are as follows:
(2) error rate Error rate, calculation formula are as follows:
(3) sensitivity S ensitive, calculation formula are as follows:
(4) special efficacy degree Specificity, calculation formula are as follows:
(5) precision Precision, calculation formula are as follows:
In formula, TP indicates that actual act classification is A and is classified the sample size that device is divided into A class, and FP indicates actual act class
Non- A and it Wei not be classified the sample size that device is divided into A class, FN indicates that actual act classification is A and is classified device and is divided into non-A
The sample size of class, TN indicate that actual act classification is non-A and is classified the sample size that device is divided into non-A class；P indicates practical
Action classification be A sample size, N indicate actual act classification be non-A sample size, A expression lie, sit, standing, walking, hurrying up,
Run, upstairs, downstairs, turn round, lie to sitting, sit on station, stand to seat and sit on one of action classification of lying.
3. a kind of human motion state classification method based on classification layering according to claim 1, feature exist
In:
Step 1, design include the experimental paradigm of different classes of movement, using containing 3-axis acceleration sensor and three axis accelerometer
The data acquisition equipment of instrument sensor acquires the sample action of respective classes according to experimental paradigm；Human motion state is defined as
Lie, sit, standing, walking, hurrying up, running, upstairs, downstairs, turn round, lie to sitting, sit on station, stand to sitting, sit on 13 kinds of inhomogeneities of lying and do not move
Make；
Step 2, according to the sequencing acted in the experimental paradigm of design, using sample action segmentation software to different classes of
Sample action carries out semi-automatic segmentation, and the sample action after segmentation is labeled as corresponding sports status categories；
Step 3 successively uses three kinds of adding window overlapping, denoising and normalization operating methods to locate the sample action after segmentation in advance
Reason, and final sample action is formed for subsequent operation；Wherein, the data window length in adding window overlap operation is set as several
According to the integral multiple of acquisition equipment sample rate, windows overlay rate is set as 50%；Denoising operation is gone using median filter method
It makes an uproar；Sample action after each subject segmentation obtains the consistent sample action of multiple data lengths, Mei Gedong after pretreatment
The classification for making sample belongs to one of 13 action classifications defined in step 1；
Step 4, to the 3-axis acceleration data for including in the sample action after each secondary splitting and three axis angular rate data into
Row time domain and frequency domain character extract；For each axis motion state data in sliding window, its mean value, standard deviation, are calculated
25 percentiles, the 50th percentile, the 75th percentile, related coefficient, the frequency spectrum of 6 axis motion state datas between any two
Energy and preceding ten kinds of frequency energy time domains and frequency domain data feature；It is dynamic after each secondary splitting after feature extraction operation
Making sample can be indicated that each sample action collection is made of the high dimensional feature matrix that a m row n is arranged by a high dimension vector；
Step 5 chooses whether to carry out feature selecting operation；If necessary to carry out feature selecting operation, then 6 are gone to step；It is no
Then, 7 are gone to step；
Step 6 carries out dimensionality reduction operation to high dimensional feature matrix using principal component analytical method, is to accumulate contribution degree greater than 75%
Standard, the eigenmatrix after obtaining dimensionality reduction；
Step 7 increases the sample action after all secondary splittings other action classification labels, and building classifier is acted
Sample classification operation；In addition to class label itself, the other sample action of three types of lying, sit and stand increases inactive action classification mark
Label；In addition to class label itself, walks, hurries up, running, the other sample action of five types increases continuity movement and fortune upstairs and downstairs
Dynamic property acts two categories label；In addition to class label itself, turns round, lies to sitting, sit on station, stand to sitting and sit on five types of lying
Other sample action increases transformational movement and motility acts two categories label；For the movement sample after all secondary splittings
This selects a sample action as test sample, remaining sample action is as training sample every time according to a verification method is stayed
This, each class hierarchy is all made of K and carries out sort operation to test sample adjacent to classification method KNN.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610509467.8A CN106203484B (en) | 2016-06-29 | 2016-06-29 | A kind of human motion state classification method based on classification layering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610509467.8A CN106203484B (en) | 2016-06-29 | 2016-06-29 | A kind of human motion state classification method based on classification layering |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106203484A CN106203484A (en) | 2016-12-07 |
| CN106203484B true CN106203484B (en) | 2019-06-21 |
Family
ID=57464423
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610509467.8A Active CN106203484B (en) | 2016-06-29 | 2016-06-29 | A kind of human motion state classification method based on classification layering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106203484B (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106897740A (en) * | 2017-02-17 | 2017-06-27 | 重庆邮电大学 | EEMD DFA feature extracting methods under Human bodys' response system based on inertial sensor |
| 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 |
| CN107329563A (en) * | 2017-05-22 | 2017-11-07 | 北京红旗胜利科技发展有限责任公司 | A kind of recognition methods of type of action, device and equipment |
| CN107273857B (en) * | 2017-06-19 | 2021-03-02 | 深圳市酷浪云计算有限公司 | Motion action recognition method and device and electronic equipment |
| CN108008151A (en) * | 2017-11-09 | 2018-05-08 | 惠州市德赛工业研究院有限公司 | A kind of moving state identification method and system based on 3-axis acceleration sensor |
| CN107837087A (en) * | 2017-12-08 | 2018-03-27 | 兰州理工大学 | A kind of human motion state recognition methods based on smart mobile phone |
| CN109993037B (en) * | 2018-01-02 | 2021-08-06 | 中国移动通信有限公司研究院 | Action recognition method and device, wearable device and computer-readable storage medium |
| 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 |
| CN109508698B (en) * | 2018-12-19 | 2023-01-10 | 中山大学 | Human behavior recognition method based on binary tree |
| CN111127733B (en) * | 2019-10-29 | 2021-12-28 | 杭州智策略科技有限公司 | Mobile crowd sensing-based canteen queuing time detection system and method |
| CN113208576A (en) * | 2021-02-01 | 2021-08-06 | 安徽华米健康科技有限公司 | PAI value calculation method, device, equipment and storage medium |
| CN114239724B (en) * | 2021-12-17 | 2023-04-18 | 中南民族大学 | Cuball motion recognition and skill evaluation method based on inertial sensor |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7447334B1 (en) * | 2005-03-30 | 2008-11-04 | Hrl Laboratories, Llc | Motion recognition system |
| CN101866429A (en) * | 2010-06-01 | 2010-10-20 | 中国科学院计算技术研究所 | Training method of multi-moving object action identification and multi-moving object action identification method |
| CN103489000A (en) * | 2013-09-18 | 2014-01-01 | 柳州市博源环科科技有限公司 | Achieving method of human movement recognition training system |
| CN103761510A (en) * | 2014-01-02 | 2014-04-30 | 华南理工大学 | Method for motion recognition for simulating human visual cortex perception mechanism |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101640077B1 (en) * | 2009-06-05 | 2016-07-15 | 삼성전자주식회사 | Apparatus and method for video sensor-based human activity and facial expression modeling and recognition |
-
2016
- 2016-06-29 CN CN201610509467.8A patent/CN106203484B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7447334B1 (en) * | 2005-03-30 | 2008-11-04 | Hrl Laboratories, Llc | Motion recognition system |
| CN101866429A (en) * | 2010-06-01 | 2010-10-20 | 中国科学院计算技术研究所 | Training method of multi-moving object action identification and multi-moving object action identification method |
| CN103489000A (en) * | 2013-09-18 | 2014-01-01 | 柳州市博源环科科技有限公司 | Achieving method of human movement recognition training system |
| CN103761510A (en) * | 2014-01-02 | 2014-04-30 | 华南理工大学 | Method for motion recognition for simulating human visual cortex perception mechanism |
Non-Patent Citations (2)
| Title |
|---|
| An adaptive rule-based approach to classifying;Saif Okour etc.;《2015 International Conference on Healthcare Informatics》;20151231;第1-4页 |
| Robust Action Recognition Based on a Hierarchical Model;Xinbo Jiang etc.;《2013 International Conference on Cyberworlds》;20131231;第1-8页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106203484A (en) | 2016-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106203484B (en) | A kind of human motion state classification method based on classification layering | |
| Lin et al. | An explainable deep fusion network for affect recognition using physiological signals | |
| CN104127193B (en) | Assessment system and its appraisal procedure that a kind of depression degree quantifies | |
| Rehg et al. | Mobile health | |
| CN101766484B (en) | Method and equipment for identification and classification of electrocardiogram | |
| CN102663450B (en) | Method for classifying and identifying neonatal pain expression and non-pain expression based on sparse representation | |
| CN104484644B (en) | A kind of gesture identification method and device | |
| Motka et al. | Diabetes mellitus forecast using different data mining techniques | |
| CN101268938A (en) | Method and apparatus for electrocardiogram recognition and specification | |
| CN111009321A (en) | Application method of machine learning classification model in juvenile autism auxiliary diagnosis | |
| CN107887032A (en) | A kind of data processing method and device | |
| CN110097975A (en) | A kind of nosocomial infection intelligent diagnosing method and system based on multi-model fusion | |
| CN109065171A (en) | The construction method and system of Kawasaki disease risk evaluation model based on integrated study | |
| Xu et al. | Identification of autism spectrum disorder based on short-term spontaneous hemodynamic fluctuations using deep learning in a multi-layer neural network | |
| CN109273093A (en) | A kind of construction method and building system of Kawasaki disease risk evaluation model | |
| CN112133407A (en) | Rapid intelligent emotion assessment analysis method based on voice and expression | |
| CN108305680A (en) | Intelligent parkinsonism aided diagnosis method based on multi-element biologic feature and device | |
| CN114358194A (en) | Gesture tracking based detection method for abnormal limb behaviors of autism spectrum disorder | |
| CN110415818A (en) | A kind of intelligent pediatric disease interrogation system and method based on observable illness | |
| CN115040086A (en) | Data processing system and method based on digital biomarkers | |
| Feng et al. | Development and application of artificial intelligence in auxiliary TCM diagnosis | |
| CN106901689A (en) | Cognitive and motor function detecting system and data processing method based on interaction in kind | |
| Ninh et al. | Analysing the performance of stress detection models on consumer-grade wearable devices | |
| Wang et al. | The Recognition Method of Athlete Exercise Intensity Based on ECG and PCG. | |
| WO2023097780A1 (en) | Classification method and device for classifying patient‑ventilator asynchrony phenomenon in mechanical ventilation process |
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 |