CN105678222B - A kind of mobile device-based Human bodys' response method - Google Patents
A kind of mobile device-based Human bodys' response method Download PDFInfo
- Publication number
- CN105678222B CN105678222B CN201511009411.8A CN201511009411A CN105678222B CN 105678222 B CN105678222 B CN 105678222B CN 201511009411 A CN201511009411 A CN 201511009411A CN 105678222 B CN105678222 B CN 105678222B
- Authority
- CN
- China
- Prior art keywords
- data
- feature
- disaggregated model
- behavior
- vector
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/245—Classification techniques relating to the decision surface
- G06F18/2451—Classification techniques relating to the decision surface linear, e.g. hyperplane
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
Abstract
A kind of mobile device-based Human bodys' response method, utilizes the multiple sensors real-time data collection built in mobile device；The data that sensor obtains are corrected, are filtered, calculate a series of data preprocessing operation such as generation data and data segmentation；Feature extraction is carried out to pretreated data segment, by the character pair vector input equipment position disaggregated model of extraction, obtains device location classification；Select corresponding behavior disaggregated model that the character pair vector input behavior disaggregated model of extraction is obtained final Activity recognition result according to the device location classification of acquisition.The present invention provides that a kind of versatility is good, the higher mobile device-based Human bodys' response method of accuracy.
Description
Technical field
The present invention relates to Activity recognition technical fields, relate in particular to a kind of mobile device-based Human bodys' response
Method.
Background technique
Human bodys' response is a kind of by obtaining and analyzing human body behavior related data, judges the skill of human body behavior state
Art.By knowing human body basis behavioral activity, which can be motion tracking, health monitoring, fall detection, the elderly's prison
Shield, patient resumes training, complex behavior identifies, support industry manufactures, human-computer interaction, augmented reality, indoor positioning and navigation, a
The research and application of the various fields such as the identification of people's feature, urbanization calculating provide human body relevant information, therefore have important answer
With value and research significance.
Traditional Human bodys' response technology is based primarily upon image information progress.This method captures human body by video equipment
The image information of movement analyzes the relevant image sequence of behavior, judges behavior classification.Behavior based on image information is known
Qi Bu not be early, theoretical relative maturity, recognition accuracy is higher, but since method is more complex, influences vulnerable to background light, and image
Acquire that equipment volume is big, power consumption is high, using being restricted.
Due to MEMS (Micro-Electro-Mechanical System, MEMS) manufacturing technology it is rapid into
Step, various types of sensors continue to develop, such as accelerometer, gyroscope, magnetometer and barometer etc..These sensor energy
The relevant information for enough collecting human motion, while having had both the advantages such as good portability, low in energy consumption, and to user without bothering, resist
Environmental disturbances ability is strong, becomes the preference data source for realizing duration Human bodys' response instead of video equipment.
In addition, with the rapid progress of mobile device the relevant technologies, especially using iOS and Android platform as the intelligence of representative
The equipment such as energy mobile phone, Intelligent flat have become the important carrier of all kinds of MEMS sensors.It is compared to wearing of specially designing
Formula sensor device is worn, this kind of mobile device has more powerful processing, storage and network-connectivity, while built-in a variety of biographies
Sensor carries or wears again extras without using person, and more convenient for user, friendly, development platform also has
There is the feature of universality.Therefore mobile device-based Activity recognition becomes the focus of researcher.
Existing sensor-based Human bodys' response method is directed to the design of dedicated wearable device mostly, to being based on
The Human bodys' response of common mobile devices is simultaneously not suitable for.Caused by due to the characteristics of wearable device, sensor is placed on use
On the fixation position of family body, generally also there is fixed direction.And all kinds of mobile devices, such as smart phone, due in day
That often lives carries and uses, and according to personal habits difference has different placement locations and different placement directions.It is related
MEMS sensor is placed on the data characteristics generated when the different parts of human body and has bigger difference by studies have shown that.Therefore will
Equipment is in the data that different location is collected into and mutually obscures the decline that will lead to whole recognition accuracy；And due to placement direction
Indefinite, the method based on certain specific direction data also can not be used directly；In addition, the operational capability of smart phone is flat from desktop
There are also larger gaps for platform, in order to realize real-time Activity recognition, need to choose and the suitable feature extracting method of design complexities
And disaggregated model.Therefore actual conditions and operational capability for smart phone is needed to design suitable method.
Summary of the invention
Versatility in order to overcome the shortcomings of existing Human bodys' response method is poor, accuracy is poor, and the present invention provides
A kind of versatility is good, the higher mobile device-based Human bodys' response method of accuracy.
The technical solution adopted by the present invention to solve the technical problems is:
A kind of mobile device-based Human bodys' response method, comprising the following steps:
Step (1), training device location disaggregated model cpWith the behavior disaggregated model ca based on distinct device positioni, cai
∈ C, C are the set being made of behavior disaggregated model, C={ ca1, ca2..., caI, behavior disaggregated model caiWith device location
piWith one-to-one relationship, device location pi∈ P, P are the set being made of device location predetermined, P={ p1,
p2..., pI, I is device location class number predetermined；
Step (2) acquires sensor raw data by mobile device built-in sensors in real time；
Step (3) carries out data prediction to the data that mobile device built-in sensors acquire in real time, obtains current time
The corresponding data segment of window；
Step (4), according to position identification feature collection Fp, feature extraction is carried out to the data segment of acquisition, obtains the data segment
Corresponding feature vectorNpIt is characterized collection FpIn number of features；By feature vector Vp
Input equipment position disaggregated model cp, obtain device location classification pi；
Step (5), according to Activity recognition feature set Fa, feature extraction is carried out to the data segment of acquisition, obtains the data segment
Corresponding feature vectorNaIt is characterized collection FaIn number of features；According to step (4)
The position classification p of acquisitioniSelect corresponding behavior disaggregated model cai, by feature vector VaInput behavior disaggregated model cai, output
Behavior classification aj, aj∈ A, A are the set being made of behavior predetermined, A={ a1, a2..., aJ, J is predetermined
Behavior class number, ajAs final recognition result.
Further, training device location disaggregated model c in the step (1)pWith row of the training based on distinct device position
For disaggregated model caiThe step of it is as follows:
Step (1-1) acquires distinct device position and different human body behavior feelings by mobile device built-in sensors offline
Sensor raw data under condition；
Step (1-2) carries out data prediction to the sensor raw data of acquisition, obtains corresponding different time window
Data segment；
Step (1-3) carries out feature extraction to each data segment, obtains the corresponding feature of the data segment according to feature set F
Vector Vk=[v1, v2..., vN]∈RN, N is characterized the number of features in collection F, k ∈ { 1,2 ..., K }, and all data segments extract
Feature vector collectively form sample set S, S=[V1；V2；...；VK], feature vector number K is data segment number；
Step (1-4), to each feature vector V in sample set Sk, flag data acquire when device location classification pi,
pi∈ P is obtained position training vector k ∈ { 1,2 ..., K }, and all training vectors after label constitute position
Set training set Tp,
Step (1-5), utilizes training set TpTraining device location disaggregated model, obtains device location disaggregated model cp；
Step (1-6), according to disaggregated model cp, the Partial Feature composition position in feature set F for the model is selected to know
Other feature set Fp；
Step (1-7) carries out feature selecting to feature set F, and selection is wherein suitable for the Partial Feature of Activity recognition model
Form Activity recognition feature set Fa；
Step (1-8), Screening Samples, which integrate, corresponds to device location category label as p in SiFeature inwardsForm sample
Subset Si, meet
Step (1-9), to sample set SiIn each feature vectorBehavior classification a when flag data acquiresj,
aj∈ A obtains Behavioral training vectorki∈ { 1,2 ..., Ki, all training vectors after label are constituted
Behavioral training collection KiFor sample set SiThe number of middle feature vector；
Step (1-10), utilizes Behavioral training collectionTrain classification models obtain corresponding position piBehavior disaggregated model
cai；
To step (1-8)-(1-10), i=1,2 ..., I is taken to be repeated in respectively, I indicates device location predetermined
Class number, training obtain the behavior disaggregated model ca under distinct device position1, ca2..., caI。
Further, the sensor raw data in the step (1-1) includes different mobile device built-in sensors
And the data on same sensor different directions；Sensor raw data, that is, X that each sampled point obtainsl=[x1, x2...,
xP]∈RP, P is sensor raw data dimension, and l ∈ { 1,2 ..., L }, L are the total sampling length of sensing data.
Preferably, the data prediction in the step (3) includes the following steps:
Step (3-1) is corrected and filters to sensor raw data；
Step (3-2) calculates according to correction and filtered data and generates data Yl, Yl=[y1, y2..., yQ]∈RQ, Q
To generate data dimension；By sensor raw data and data merging is generated, obtains Zl, Zl=[Xl, Yl]∈RP+Q, l ∈ 1,
2 ..., L }；Data group on all sampled points is combined into sensor data set, i.e. [Z1；Z2；...；ZL]；
Continuous sensor data set is divided into data segment according to time window by the segmentation of step (3-3) data.
Pretreatment in step (1-1) also uses the above process.
It is by feature vector V that device location classification is obtained in the step (4)pInput position disaggregated model cp, cpFor training
Obtained decision-tree model；The feature vector of input selects qualified branch according to the splitting condition of decision tree step by step,
It is being classified at leaf node as a result, i.e. position pi, pi∈P。
Acquisition behavior classification is by feature vector V in the step (5)aInput behavior disaggregated model cai, caiIt is trained
The support vector cassification model arrived is made of g (J) a support vector machines, function g (x) by model multi-class classification policy
It determines, J is class object, that is, behavior classification number；The corresponding Optimal Separating Hyperplane expression formula of each support vector machines is f (x),
Functional value f (V) is calculated according to the feature vector V of input, is obtained classification results (f (V) >=0 or f (V) < 0)；To it is all support to
The classification results of amount machine are weighted processing, and peak is taken to export as behavior classification results.
Feature set F in step (1-3) is made of one-dimensional corresponding time domain every in data segment, frequency domain or time and frequency domain characteristics.
The device location disaggregated model of training is decision-tree model in step (1-5), and the classifier is by one group " if yes "
The classifying rules of form constitutes tree structure；In training, the highest feature of each split vertexes selection evaluation function value is made
For splitting condition.
Feature selecting in step (1-6) selects the feature set F for the input of position disaggregated model from feature set Fp,FpIt is screened using embedded mode, i.e. Decision-Tree Classifier Model c of the training for position identificationpAfterwards, by decision tree classification
Model feature composition characteristic collection F selected to usep；
Feature selecting in step (1-7) selects the feature set F for the input of behavior disaggregated model from feature set Fa,FaUsing filtering-encapsulation two-stage process screening；Filter process selection evaluation function successively comments all features
Valence and sequence, screening and assessment function result are highest M first；Encapsulation process passes through the output of combination supporting vector machine disaggregated model
As a result, search meets the minimal characteristic set of output threshold condition, as the feature set F inputted for behavior disaggregated modela；
The behavior disaggregated model of training is supporting vector machine model in step (1-10), and the model is by g (J) a supporting vector
Machine composition；Each support vector machines according to the training data of input construct largest interval Optimal Separating Hyperplane, the Optimal Separating Hyperplane by
FormulaDefinition, wherein x is the feature vector of input, xiIt is supporting vector, yiIt is branch
Hold the corresponding result label of vector, yiValue meets yi∈ { -1,1 }, n are supporting vector number, aiIt is the branch being calculated with b
Hold vector parameter, K (x, xi) it is kernel function for input vector to be mapped to high dimension linear space.
Beneficial effects of the present invention are mainly manifested in: the present invention provides one kind to be based on mobile device built-in sensors data
Human bodys' response method, this method can identify carrying of the mobile device with respect to human body or using position, and then select needle
To the Activity recognition model of the position.Compared with prior art, method provided by the invention passes through above-mentioned layering identification model
Building and suitable feature selection approach can adapt to different in daily life under the premise of guaranteeing that computation complexity is moderate
Position of mobile equipment and direction, realize the Human bodys' response of high-accuracy.
Detailed description of the invention
Fig. 1 is the flow chart of mobile device-based Human bodys' response method of the invention；
Fig. 2 is device location disaggregated model of the invention and behavior disaggregated model training flow chart；
Fig. 3 is feature selecting flow chart of the invention.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
Referring to Fig.1~Fig. 3, a kind of mobile device-based Human bodys' response method, comprising the following steps:
Step (1), training device location disaggregated model cpWith the behavior disaggregated model ca based on distinct device positioni, cai
∈ C, C are the set being made of behavior disaggregated model, C={ ca1, ca2..., caI, behavior disaggregated model caiWith device location
piWith one-to-one relationship, device location pi∈ P, P are the set being made of device location predetermined, P={ p1,
p2..., pI, I is device location class number predetermined；
Step (2) acquires sensor raw data by mobile device built-in sensors in real time；
Step (3) carries out data prediction to the data that mobile device built-in sensors acquire in real time, obtains current time
The corresponding data segment of window；
Step (4), according to position identification feature collection Fp, feature extraction is carried out to the data segment of acquisition, obtains the data segment
Corresponding feature vectorNpIt is characterized collection FpIn number of features；By feature vector Vp
Input equipment position disaggregated model cp, obtain device location classification pi；
Step (5), according to Activity recognition feature set Fa, feature extraction is carried out to the data segment of acquisition, obtains the data segment
Corresponding feature vectorNaIt is characterized collection FaIn number of features；According to step (4)
The position classification p of acquisitioniSelect corresponding behavior disaggregated model cai, by feature vector VaInput behavior disaggregated model cai, output
Behavior classification aj, aj∈ A, A are the set being made of behavior predetermined, A={ a1, a2..., aJ, J is predetermined
Behavior class number, ajAs final recognition result；
Training device location disaggregated model c in the step (1)pBehavior classification mould with training based on distinct device position
Type caiSpecific step is as follows:
Step (1-1) acquires distinct device position and different human body behavior feelings by mobile device built-in sensors offline
Sensor raw data under condition；
Step (1-2) carries out data prediction to the sensor raw data of acquisition, obtains corresponding different time window
Data segment；
Step (1-3) carries out feature extraction to each data segment, obtains the corresponding feature of the data segment according to feature set F
Vector Vk=[v1, v2..., vN]∈RN, N is characterized the number of features in collection F, k ∈ { 1,2 ..., K }, and all data segments extract
Feature vector collectively form sample set S, S=[V1；V2；...；VK], feature vector number K is data segment number；
Step (1-4), to each feature vector V in sample set Sk, flag data acquire when device location classification pi,
pi∈ P is obtained position training vector k ∈ { 1,2 ..., K }, and all training vectors after label constitute position
Set training set Tp,
Step (1-5), utilizes training set TpTraining device location disaggregated model, obtains device location disaggregated model cp；
Step (1-6), according to disaggregated model cp, the Partial Feature composition position in feature set F for the model is selected to know
Other feature set Fp；
Step (1-7) carries out feature selecting to feature set F, and selection is wherein suitable for the Partial Feature of Activity recognition model
Form Activity recognition feature set Fa；
Step (1-8), Screening Samples, which integrate, corresponds to device location category label as p in SiFeature inwardsForm sample
Subset Si, meet
Step (1-9), to sample set SiIn each feature vectorBehavior classification a when flag data acquiresj,
aj∈ A obtains Behavioral training vectorki∈ { 1,2 ..., Ki, all training vectors after label are constituted
Behavioral training collection KiFor sample set SiThe number of middle feature vector；
Step (1-10), utilizes Behavioral training collectionTrain classification models obtain corresponding position piBehavior disaggregated model
cai；
To step (1-8)-(1-10), i=1,2 ..., I is taken to be repeated in respectively, training obtains under distinct device position
Behavior disaggregated model ca1, ca2..., caI。
Further, the sensor raw data in the step (2) and step (1-1) includes built in different mobile devices
Data on sensor and same sensor different directions；Sensor raw data, that is, X that each sampled point obtainsl=[x1,
x2..., xP]∈RP, P is sensor raw data dimension, and l ∈ { 1,2 ..., L }, L are the total sampling length of sensing data.
Data prediction in step (3) and step (1-2) includes the following steps:
Step (3-1) is corrected and filters to sensor raw data；
Step (3-2) calculates according to correction and filtered data and generates data Yl, Yl=[y1, y2..., yQ]∈RQ, Q
To generate data dimension；By sensor raw data and data merging is generated, obtains Zl, Zl=[Xl, Yl]∈RP+Q, l ∈ 1,
2 ..., L }；Data group on all sampled points is combined into sensor data set, i.e. [Z1；Z2；...；ZL]；
Continuous sensor data set is divided into data segment according to time window by the segmentation of step (3-3) data.
It is by feature vector V that device location classification is obtained in step (4)pInput position disaggregated model cp, cpIt is obtained for training
Decision-tree model；The feature vector of input selects qualified branch according to the splitting condition of decision tree step by step, in leaf
It is being classified at node as a result, i.e. position pi, pi∈P。
Acquisition behavior classification is by feature vector V in step (5)aInput behavior disaggregated model cai, caiTraining obtains
Support vector cassification model is made of g (J) a support vector machines, and function g (x) is determined by the multi-class classification policy of model,
J is class object, that is, behavior classification number；The corresponding Optimal Separating Hyperplane expression formula of each support vector machines is f (x), according to defeated
The feature vector V entered calculates functional value f (V), obtains classification results (f (V) >=0 or f (V) < 0)；To all support vector machines
Classification results are weighted processing, and peak is taken to export as behavior classification results.
Feature set F in step (1-3) is made of one-dimensional corresponding time domain every in data segment, frequency domain or time and frequency domain characteristics.
The device location disaggregated model of training is decision-tree model in step (1-5), and the classifier is by one group " if yes "
The classifying rules of form constitutes tree structure；In training, the highest feature of each split vertexes selection evaluation function value is made
For splitting condition.
Feature selecting in step (1-6) selects the feature set F for the input of position disaggregated model from feature set Fp,FpIt is screened using embedded mode, i.e. Decision-Tree Classifier Model c of the training for position identificationpAfterwards, by decision tree point
The feature composition characteristic collection F used selected by class modelp。
Feature selecting in step (1-7) selects the feature set F for the input of behavior disaggregated model from feature set Fa,FaUsing filtering-encapsulation two-stage process screening；Filter process selection evaluation function successively comments all features
Valence and sequence, screening and assessment function result are highest M first；Encapsulation process passes through the output of combination supporting vector machine disaggregated model
As a result, search meets the minimal characteristic set of output threshold condition, as the feature set F inputted for behavior disaggregated modela。
The behavior disaggregated model of training is supporting vector machine model in step (1-10), and the model is by g (J) a supporting vector
Machine composition；Each support vector machines according to the training data of input construct largest interval Optimal Separating Hyperplane, the Optimal Separating Hyperplane by
FormulaDefinition, wherein x is the feature vector of input, xiIt is supporting vector, yiIt is branch
Hold the corresponding result label of vector, yiValue meets yi∈ { -1,1 }, n are supporting vector number, aiIt is the branch being calculated with b
Hold vector parameter, K (x, xi) it is kernel function for input vector to be mapped to high dimension linear space.
Mobile device of the present invention mainly includes smart phone, Intelligent flat computer etc., the tool that the present embodiment uses
Body mobile device is 4 mobile phone of millet for running Android platform.Select acceleration transducer, the gyro built in 4 mobile phone of millet
Instrument, baroceptor, light sensor and range sensor are acquired for data, and sample frequency is set as 50Hz.Data acquisition
When, the factors such as the specific implementation of equipment placement direction and behavior are not required.The collected data of each sampled point
Format is Xl=[x1, x2..., xP]∈RP, P=9, l ∈ { 1,2 ..., L }, L is the total sampling length of sensing data, L=
485050.The nine dimensions sensor raw data includes three-dimensional acceleration information, three-dimensional gyro data, one-dimensional air pressure transmission
Sensor data, one-dimensional light sensor data and one-dimensional range-sensor data.
The present embodiment acquires sensor raw data using 4 cell phone apparatus of millet, for device location disaggregated model and
The training and assessment of behavior disaggregated model.The number for participating in data sampling is 10 people, including 3 women and 7 males；It adopts
The behavior of collection include sit, stand, walking, running, upstairs, go downstairs 6 kinds of classifications.Above-mentioned 6 kinds of behavior classifications are denoted as a respectively1, a2..., a6,
Constituting action set A={ a1, a2..., a6}.The position of mobile equipment of acquisition includes in pocket of trousers, in coat pocket, both shoulders packet
In, in shoulder bag, hold in side, be placed in one's ear, hold in front, place 8 kinds of classifications on the table.Above-mentioned 8 kinds of equipment
Position is denoted as p respectively1, p2..., p8, constitution equipment location sets P={ p1, p2..., p8}。
As shown in Fig. 2, in the training process of disaggregated model, by the collected multiple sensors initial data of mobile device
It needs by further pretreatment operation.These operations include the correction to acceleration information, to acceleration transducer, gyro
The filtering and noise reduction of instrument and baroceptor data calculates according to correction and filtered data and generates data, to continuous data
Data segmentation.
Wherein, generate data include three-dimensional acceleration of gravity data, three-dimensional linear acceleration data and it is one-dimensional plus
Velocity amplitude data.Specific implementation includes that low-pass filtering is carried out to three-dimensional acceleration data, and low-pass filtering is calculated such as formula:
afiltered(l)=afiltered(l-1)+α×(aunfiltered(l)-afiltered(l-1)) (1)
Wherein l indicates current sampling point, afilteredIt (l) is filtered acceleration value, aunfilteredIt (l) is before filtering
Acceleration value, factor alpha are cutoff frequency, are set as 0.05Hz.
It is three-dimensional gravity acceleration information, remaining high frequency section, i.e. a by the data obtained after low-pass filterhigh
(l)=a (l)-afilteredIt (l) is linear acceleration data.
The calculating that data further include one-dimensional acceleration amplitude data is generated, i.e., ax、ay、
azFor the 3-axis acceleration data at current sampling point.Therefore, the form for generating data is Yl=[y1, y2..., yQ]∈RQ, Q
=7.By correction, filtered sensor acquisition data and data merging is generated, obtains the data Z on each sampled pointl, Zl=
[Xl, Yl]∈RP+Q, P+Q=16, l ∈ { 1,2 ..., L }.
Data segmentation is handled using slip window sampling the continuous data in each dimension of sensor data set, is adopted
It is 2 seconds with length, i.e., the sliding window that is 100 comprising data sampling points is set as 50% between two adjacent windows and covers
Lid.Collected continuous data is divided into the data segment of floor (T/50-1) number by the slip window sampling, wherein floor ()
For downward bracket function.Segmentation obtains summary data field number K=9551.
Feature set F is by mean value, standard deviation, variance, the degree of bias, kurtosis, maximum value, minimum value, mean square deviation, interquartile range, flat
Equal absolute deviation, 11 category feature of coverage area are constituted, and every category feature generates 16 spies to P+Q=16 data axis of data segment
Sign, therefore symbiosis is at N=11 × 16=176 features.
Enter feature extraction process by pretreated data.Feature extraction is the feature meter to data contained by data segment
It calculates.To each data segment of above-mentioned acquisition, calculate separately the N item feature in feature set F, obtain each data segment calculate feature to
Measure Vk=[v1, v2..., vN]∈RN, k ∈ { 1,2 ..., K }.The feature vector of all acquisitions forms sample set S=[V1；
V2；...；VK]。
By each feature vector V in above-mentioned sample set SkWith position mark piForm training vectorpi∈P.By training vectorComposing training collection For
The training of device location disaggregated model.
Training device location disaggregated model is using C4.5 Method And Principle training Decision-Tree Classifier Model.In training, each
Split vertexes select the highest feature of information gain-ratio as splitting condition.The calculation of information gain-ratio is as follows:
In above formula, molecule indicates to divide bring information gain, the i.e. reduction of comentropy, Entropy (T) table by training set
Show the comentropy gathered before division,Gather after expression division.Wherein Entropy is collection
The comentropy of conjunction, piIt is the ratio for belonging to classification i in training set T, m is the number of target category；V (f) is the possible institute of feature f
There is value set, v is the value set for wherein meeting splitting condition；TvIt is the son that characteristic value meets splitting condition in training set T
Collection,It is TvThe ratio denominator term of middle feature vector number and the feature vector number in former set T indicates division Information Meter,
Wherein；TiIt is the corresponding feature vector set of target category i,It is the corresponding feature vector number of target category i and former set
The ratio of feature vector number in T.
According to above method principle, training forms Decision-Tree Classifier Model, and the size of the decision tree is 137, uses feature
Number is 69, number of features in about feature set F69 above-mentioned feature composition characteristic collection, this feature collection are used for
The feature set F of device location identificationp, as shown in Figure 3.
It is assessed using the device class model that training set obtains training.Assessment has used ten folding cross-validation methods,
The average recognition accuracy of obtained device location is 97.2%.
By each feature vector V and behavior category label a in above-mentioned sample set SjComposition training is inwardsaj∈ A, 6 kinds of human body behavior classifications when corresponding data acquires.By training vectorIt constitutes
Training setFor feature selecting.Feature selecting includes filtering and encapsulating two step process.Filtering
To each single item feature calculation information gain-ratio in feature set F, information gain-ratio such as formula (2) is described, big according to information gain-ratio
Small to be ranked up, the forward M item feature of selected and sorted takes M=100.Then, combination supporting vector machine model was packaged
Journey further screens filtered feature set.Encapsulation process comprehensively considers the accuracy rate and recall rate of classification, using branch
The F-Measure of vector machine classification results is held as evaluation criterion, calculation is as follows:
Wherein the accurate rate of P presentation class model output, R presentation class model export recall rate.It is filtered all special
The F-Measure obtained when sign input support vector machines is 93.12%.After the preceding filtering to sequential search strategy search one by one
Feature set, when disaggregated model output meet setting threshold value, i.e. when F-Measure is more than 88.46%, output signature search
As a result, the result includes 49 features.These features constitute the feature set F for being used for Human bodys' responsea。
To sample set S, screening position classification is piFeature vector, form sample set Si, meetI ∈ 1,
2 ..., 8 }.To sample set SiIn each feature vectorBehavior classification a when flag data acquiresj, aj∈ A is obtained
Behavioral training vectorki∈ { 1,2 ..., Li, all training vector constituting action training sets after label LiFor feature vector number in behavior training set.Utilize Behavioral training collectionInstruction
Practice disaggregated model, obtains behavior disaggregated model cai.Training behavior disaggregated model caiIt is realized using support vector machine method.
The supporting vector machine model target category is human body behavior, including 6 kinds of target categories, using one-to-one multi-class
Classification policy, respective functionTherefore 15 support vector machines of training is needed to constitute complete mould of classifying more
Type.Each support vector machines selects the corresponding data of two kinds of behavior classifications, according to the training data of input calculate supporting vector and
Parameter constructs largest interval Optimal Separating Hyperplane, and the Optimal Separating Hyperplane is by formulaIt is fixed
Justice, wherein x is the feature vector of input, and wherein x is the feature vector of input, xiIt is supporting vector, yiIt is that supporting vector is corresponding
As a result label, yiValue meets yi∈ { -1,1 }, n are supporting vector number, aiIt is the supporting vector parameter being calculated, K with b
(x, xi) it is kernel function of the non-linear low-dimensional maps feature vectors for that will input to higher dimensional space realization linear separability；Core letter
Number uses Radial basis kernel function, by formulaIt defines, wherein xcFor kernel function center, σ is the width of function
Parameter is spent, σ=1 is taken.
I=1,2 ..., 8 is taken to above-mentioned supporting vector machine model respectively, is repeated in, and then training obtains 8 kinds of different positions
Set the behavior disaggregated model ca in the case of classification1, ca2..., ca8.Using training set to training obtain based on different location
The assessment of behavior disaggregated model, as a result such as table 1:
| Behavior disaggregated model based on different location | Behavior classification accuracy |
| Equipment is located at the behavior disaggregated model ca in pocket of trousers1 | 96.4% |
| Equipment is located at the behavior disaggregated model ca of coat pocket2 | 98.0% |
| Equipment is located at the behavior disaggregated model ca in knapsack3 | 98.0% |
| Equipment is located at the behavior disaggregated model ca in shoulder bag4 | 98.7% |
| Behavior disaggregated model ca when holding using equipment5 | 97.5% |
| Behavior disaggregated model ca when handheld device is conversed6 | 98.6% |
| Behavior disaggregated model ca of handheld device when side7 | 95.4% |
| Behavior disaggregated model ca when equipment is located on table8 | - |
Table 1 is assessed based on the behavior disaggregated model of different location
The behavior classification method of differentiation is not added to position using training set and behavior classification method provided by the invention carries out
Assessment, such as table 2, it can be seen that layering behavior classification method provided by the invention has a distinct increment to Activity recognition accuracy rate.
Whether table 2 distinguishes two kinds of classification method accuracys rate comparison of device location
The mould as shown in Figure 1, above-mentioned trained device location disaggregated model and the behavior based on distinct device position are classified
Type has built the Activity recognition model of layering.In real-time identification process, specifically locate in advance comprising real-time data acquisition, data
Reason, feature extraction and device location identification and Activity recognition process.
Under at an arbitrary position, pass through acceleration transducer, gyroscope, barometer, the light sensor built in mobile device
With range sensor with the frequency collection real time data of 50Hz.
The data of acquisition are calculated by Data correction, filtering, generation data, are preset when the data points of sampling meet
Window considerations when, that is, when meeting new sampled point number and reaching 50, together with 50 groups of samples before the window at sampling
The window that length is 100, forms the data segment of current window.
According to feature set Fp, feature extraction is carried out to the data segment of acquisition, obtains the corresponding feature vector of the data segmentNp=69；By feature vector VpInput equipment position disaggregated model cp。cpFor training
The Decision-Tree Classifier Model finished, the feature vector V of inputpIt is selected step by step according to the splitting condition of decision tree qualified
Branch, the final result classified at leaf node, i.e. position pi, pi∈P。
According to feature set Fa, feature extraction is carried out to the data segment of acquisition, obtains the corresponding feature vector of the data segmentNa=49；By feature vector VaInput and device location classification piCorresponding behavior point
Class model cai。caiFor the support vector cassification model that training obtains, it is made of 15 support vector machines, each supporting vector
Machine classifies to two kinds of behavior classifications.Optimal Separating Hyperplane expression formula corresponding to each support vector machines, according to the spy of input
Levy vector VaCalculate functional value f (Va), obtain classification results (f (Va) >=0 or f (Va) < 0), then to all classification results
It votes, number of votes obtained is highest after adding up exports as behavior classification results.
It, can be with it should be noted that the behavior classification in the embodiment of the present invention includes but are not limited to above-mentioned situation
Other a variety of mobile devices are adapted to carry, using position, can also be to other a variety of behavior category classifications, it can be according to actual
It is determined using needing.
Claims (9)
1. a kind of mobile device-based Human bodys' response method, it is characterised in that: the following steps are included:
Step (1), training device location disaggregated model cpWith the behavior disaggregated model ca based on distinct device positioni, cai∈ C, C
It is the set being made of behavior disaggregated model, C={ ca1, ca2..., caI, behavior disaggregated model caiWith device location piHave
One-to-one relationship, device location pi∈ P, P are the set being made of device location predetermined, P={ p1, p2...,
pI, I is device location class number predetermined；
Step (2) acquires sensor raw data by mobile device built-in sensors in real time；
Step (3) carries out data prediction to the data that mobile device built-in sensors acquire in real time, obtains actual time window
Corresponding data segment；
Step (4), according to position identification feature collection Fp, feature extraction is carried out to the data segment of acquisition, it is corresponding to obtain the data segment
Feature vectorNpIt is characterized collection FpIn number of features；By feature vector VpInput is set
Standby position disaggregated model cp, obtain device location classification pi；
Step (5), according to Activity recognition feature set Fa, feature extraction is carried out to the data segment of acquisition, it is corresponding to obtain the data segment
Feature vectorNaIt is characterized collection FaIn number of features；It is obtained according to step (4)
Position classification piSelect corresponding behavior disaggregated model cai, by feature vector VaInput behavior disaggregated model cai, export behavior class
Other aj, aj∈ A, A are the set being made of behavior predetermined, A={ a1, a2..., aJ, J is behavior class predetermined
Other number, ajAs final recognition result；
Training device location disaggregated model c in the step (1)pWith behavior disaggregated model ca of the training based on distinct device positioni
The step of it is as follows:
Step (1-1), in the case of acquiring distinct device position and different human body behavior offline by mobile device built-in sensors
Sensor raw data；
Step (1-2) carries out data prediction to the sensor raw data of acquisition, obtains the data of corresponding different time window
Section；
Step (1-3) carries out feature extraction to each data segment, obtains the corresponding feature vector of the data segment according to feature set F
Vk=[v1, v2..., vN]∈RN, N is characterized the number of features in collection F, k ∈ { 1,2 ..., K }, the spy that all data segments extract
Sign vector collectively forms sample set S, S=[V1；V2；...；VK], feature vector number K is data segment number；
Step (1-4), to each feature vector V in sample set Sk, flag data acquire when device location classification pi, pi∈
P obtains position training vectorK ∈ { 1,2 ..., K }, all training vectors after label constitute position instruction
Practice collection Tp,
Step (1-5), utilizes training set TpTraining device location disaggregated model, obtains device location disaggregated model cp；
Step (1-6), according to disaggregated model cp, select to form position identification feature for the Partial Feature of the model in feature set F
Collect Fp；
Step (1-7) carries out feature selecting to feature set F, and selection is wherein formed suitable for the Partial Feature of Activity recognition model
Activity recognition feature set Fa；
Step (1-8), Screening Samples, which integrate, corresponds to device location category label as p in SiFeature vectorForm sample set
Si, meet
Step (1-9), to sample set SiIn each feature vectorBehavior classification a when flag data acquiresi, aj∈ A,
Obtain Behavioral training vector, ki∈ { 1,2 ..., Ki, all training vector constituting actions instruction after label
Practice collection Ti a, KiFor sample set SiThe number of middle feature vector；
Step (1-10) utilizes Behavioral training collection Ti aTrain classification models obtain corresponding position piBehavior disaggregated model cai；
To step (1-8)-(1-10), i=1,2 ..., I is taken to be repeated in respectively, I indicates device location classification predetermined
Number, training obtain the behavior disaggregated model ca under distinct device position1, ca2..., CaI。
2. mobile device-based Human bodys' response method as described in claim 1, it is characterised in that: in step (1-1)
Sensor raw data include data on different mobile device built-in sensors and same sensor different directions；Each
Sensor raw data, that is, X that sampled point obtainsl=[x1, x2..., xP]∈RP, P be sensor raw data dimension, l ∈ 1,
2 ..., L }, L is the total sampling length of sensing data.
3. mobile device-based Human bodys' response method as claimed in claim 1 or 2, it is characterised in that: the step
(3) data prediction in includes the following steps:
Step (3-1) is corrected and filters to sensor raw data；
Step (3-2) calculates according to correction and filtered data and generates data Yl,
Yl=[y1, y2..., yQ]∈RQ, Q is to generate data dimension；By sensor raw data and data merging is generated, is obtained
Zl, Zl=[Xl, Yl]∈RP+Q, l ∈ { 1,2 ..., L }；Data group on all sampled points is combined into sensor data set, i.e. [Z1；
Z2；...；ZL]；
Continuous sensor data set is divided into data segment according to time window by the segmentation of step (3-3) data.
4. mobile device-based Human bodys' response method as claimed in claim 1 or 2, it is characterised in that: the step
(4) it is by feature vector V that device location classification is obtained inpInput position disaggregated model cp, cpThe decision tree mould obtained for training
Type；The feature vector of input selects qualified branch according to the splitting condition of decision tree step by step, obtains at leaf node
Classification as a result, i.e. position pi, pi∈P。
5. mobile device-based Human bodys' response method as claimed in claim 1 or 2, it is characterised in that: the step
(5) acquisition behavior classification is by feature vector V inaInput behavior disaggregated model cai, caiThe support vector machines point obtained for training
Class model is made of g (J) a support vector machines, and function g (x) determines that J is class object by the multi-class classification policy of model
That is the number of behavior classification；The corresponding Optimal Separating Hyperplane expression formula of each support vector machines is f (x), according to the feature of input to
It measures V and calculates functional value f (V), obtain classification results, f (V) >=0 or f (V) < 0；To the classification results of all support vector machines into
Row weighting processing, takes peak to export as behavior classification results.
6. mobile device-based Human bodys' response method as described in claim 1, it is characterised in that: in step (1-3)
Feature set F by data segment per one-dimensional corresponding time domain, frequency domain or time and frequency domain characteristics form.
7. mobile device-based Human bodys' response method as described in claim 1, it is characterised in that: in step (1-5)
Trained device location disaggregated model is decision-tree model, which is made of the classifying rules of one group of " if yes " form
Tree structure；In training, each split vertexes select the highest feature of evaluation function value as splitting condition.
8. mobile device-based Human bodys' response method as described in claim 1, it is characterised in that: in step (1-6)
Feature selecting selected from feature set F for position disaggregated model input feature set Fp,FpUsing embedded mode
Screening, i.e. Decision-Tree Classifier Model c of the training for position identificationpAfterwards, the feature used as selected by Decision-Tree Classifier Model
Composition characteristic collection Fp；
Feature selecting in step (1-7) selects the feature set F for the input of behavior disaggregated model from feature set Fa,FaUsing filtering-encapsulation two-stage process screening；Filter process selection evaluation function successively evaluates all features
And sequence, screening and assessment function result are highest M first；Encapsulation process passes through the output knot of combination supporting vector machine disaggregated model
Fruit, search meets the minimal characteristic set of output threshold condition, as the feature set F inputted for behavior disaggregated modela。
9. mobile device-based Human bodys' response method as claimed in claim 2, it is characterised in that: in step (1-10)
Trained behavior disaggregated model is supporting vector machine model, which is made of g (J) a support vector machines；Each supporting vector
Machine constructs largest interval Optimal Separating Hyperplane according to the training data of input, and the Optimal Separating Hyperplane is by formulaDefinition, wherein x is the feature vector of input, xiIt is supporting vector, yiIt is to support
The corresponding result label of vector, yiValue meets yi∈ { -1,1 }, n are supporting vector number, aiIt is the support being calculated with b
Vector parameter, K (x, xi) it is kernel function for input vector to be mapped to high dimension linear space.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201511009411.8A CN105678222B (en) | 2015-12-29 | 2015-12-29 | A kind of mobile device-based Human bodys' response method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201511009411.8A CN105678222B (en) | 2015-12-29 | 2015-12-29 | A kind of mobile device-based Human bodys' response method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105678222A CN105678222A (en) | 2016-06-15 |
| CN105678222B true CN105678222B (en) | 2019-05-31 |
Family
ID=56297832
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201511009411.8A Active CN105678222B (en) | 2015-12-29 | 2015-12-29 | A kind of mobile device-based Human bodys' response method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105678222B (en) |
Families Citing this family (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107643908A (en) * | 2016-07-20 | 2018-01-30 | 中兴通讯股份有限公司 | Voice application trigger control method, device and terminal |
| CN106295675B (en) * | 2016-07-25 | 2019-09-27 | 南京邮电大学 | A kind of Activity recognition method based on smart phone of high accuracy |
| CN106123911A (en) * | 2016-08-06 | 2016-11-16 | 深圳市爱康伟达智能医疗科技有限公司 | A kind of based on acceleration sensor with the step recording method of angular-rate sensor |
| CN107918688B (en) * | 2016-10-10 | 2020-02-28 | 深圳云天励飞技术有限公司 | Scene model dynamic estimation method, data analysis method and device and electronic equipment |
| CN106709471A (en) * | 2017-01-05 | 2017-05-24 | 宇龙计算机通信科技(深圳)有限公司 | Fall detection method and device |
| CN107092894A (en) * | 2017-04-28 | 2017-08-25 | 孙恩泽 | A kind of motor behavior recognition methods based on LSTM models |
| CN107495972A (en) * | 2017-08-14 | 2017-12-22 | 哈尔滨工业大学（威海） | One kind falls down detection algorithm |
| CN108196998A (en) * | 2018-01-02 | 2018-06-22 | 联想(北京)有限公司 | A kind of state identification method, mobile equipment and server |
| CN108245172B (en) * | 2018-01-10 | 2020-04-07 | 山东大学 | Human body posture recognition method free of position constraint |
| CN108392208B (en) * | 2018-02-23 | 2020-12-15 | 大同公元三九八智慧养老服务有限公司 | Falling judgment method based on multi-stage model |
| CN108921047B (en) * | 2018-06-12 | 2021-11-26 | 江西理工大学 | Multi-model voting mean value action identification method based on cross-layer fusion |
| CN108985195A (en) * | 2018-06-29 | 2018-12-11 | 平安科技（深圳）有限公司 | Activity recognition method, apparatus, computer equipment and storage medium |
| CN109784418B (en) * | 2019-01-28 | 2020-11-17 | 东莞理工学院 | Human behavior recognition method and system based on feature recombination |
| CN109889672B (en) * | 2019-04-01 | 2020-12-25 | 中国科学技术大学 | Friend recommendation method based on mobile phone sensor |
| CN111814523A (en) * | 2019-04-12 | 2020-10-23 | 北京京东尚科信息技术有限公司 | Human body activity recognition method and device |
| CN110427969B (en) * | 2019-07-01 | 2020-11-27 | 创新先进技术有限公司 | Data processing method and device and electronic equipment |
| CN110458033B (en) * | 2019-07-17 | 2023-01-03 | 哈尔滨工程大学 | Human body behavior sequence identification method based on wearable position sensor |
| CN110598744A (en) * | 2019-08-12 | 2019-12-20 | 浙江大学 | Real-time human body behavior recognition system and method based on inertial sensor and Edge TPU |
| CN111582361A (en) * | 2020-05-06 | 2020-08-25 | 浙江大学 | Human behavior recognition method based on inertial sensor |
| CN112471680A (en) * | 2020-12-21 | 2021-03-12 | 浙江工业大学 | Intelligent safety helmet capable of identifying working state of electric power worker and working method of intelligent safety helmet |
| CN112782365B (en) * | 2020-12-30 | 2022-08-23 | 江苏智闻智能传感科技有限公司 | Mixed gas identification method based on multistage classifier |
| CN115731602A (en) * | 2021-08-24 | 2023-03-03 | 中国科学院深圳先进技术研究院 | Human body activity recognition method, device, equipment and storage medium based on topological representation |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8165405B2 (en) * | 2006-12-18 | 2012-04-24 | Honda Motor Co., Ltd. | Leveraging temporal, contextual and ordering constraints for recognizing complex activities in video |
| CN103268495B (en) * | 2013-05-31 | 2016-08-17 | 公安部第三研究所 | Human body behavior modeling recognition methods based on priori knowledge cluster in computer system |
| CN103886323B (en) * | 2013-09-24 | 2017-02-15 | 清华大学 | Behavior identification method based on mobile terminal and mobile terminal |
| CN105046215B (en) * | 2015-07-07 | 2018-11-16 | 中国科学院上海高等研究院 | The posture Activity recognition method not influenced by individual wearing position and wearing mode |
-
2015
- 2015-12-29 CN CN201511009411.8A patent/CN105678222B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN105678222A (en) | 2016-06-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105678222B (en) | A kind of mobile device-based Human bodys' response method | |
| Yen et al. | Human daily activity recognition performed using wearable inertial sensors combined with deep learning algorithms | |
| Henpraserttae et al. | Accurate activity recognition using a mobile phone regardless of device orientation and location | |
| CN105310696B (en) | A kind of fall detection model building method and corresponding fall detection method and device | |
| CN106096662B (en) | Human motion state identification based on acceleration transducer | |
| Uddin et al. | Random forests based recognition of human activities and postural transitions on smartphone | |
| Mekruksavanich et al. | Exercise activity recognition with surface electromyography sensor using machine learning approach | |
| CN106228200A (en) | A kind of action identification method not relying on action message collecting device | |
| CN108171278B (en) | Motion pattern recognition method and system based on motion training data | |
| CN103400123A (en) | Gait type identification method based on three-axis acceleration sensor and neural network | |
| Ahmed et al. | An approach to classify human activities in real-time from smartphone sensor data | |
| CN105023022A (en) | Tumble detection method and system | |
| CN112464738A (en) | Improved naive Bayes algorithm user behavior identification method based on mobile phone sensor | |
| CN110532898A (en) | A kind of physical activity recognition methods based on smart phone Multi-sensor Fusion | |
| WO2023035093A1 (en) | Inertial sensor-based human body behaviour recognition method | |
| Al-Ghannam et al. | Prayer activity monitoring and recognition using acceleration features with mobile phone | |
| CN110674875A (en) | Pedestrian motion mode identification method based on deep hybrid model | |
| CN106295675A (en) | A kind of Activity recognition method of based on smart mobile phone of high accuracy | |
| Hung et al. | Activity recognition with sensors on mobile devices | |
| CN109770913A (en) | A kind of abnormal gait recognition methods based on reverse transmittance nerve network | |
| Wang et al. | Energy expenditure estimation during normal ambulation using triaxial accelerometry and barometric pressure | |
| Harasimowicz et al. | Accelerometer-based human activity recognition and the impact of the sample size | |
| Zhang et al. | An efficient feature selection method for activity classification | |
| CN111772639B (en) | Motion pattern recognition method and device for wearable equipment | |
| CN105760819A (en) | Daily activity recognition method based on acceleration signal |
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 |