CN106875630B - A kind of wearable fall detection method and system based on hierarchical classification - Google Patents
A kind of wearable fall detection method and system based on hierarchical classification Download PDFInfo
- Publication number
- CN106875630B CN106875630B CN201710144772.6A CN201710144772A CN106875630B CN 106875630 B CN106875630 B CN 106875630B CN 201710144772 A CN201710144772 A CN 201710144772A CN 106875630 B CN106875630 B CN 106875630B
- Authority
- CN
- China
- Prior art keywords
- tumble
- sample
- weighting
- classification model
- sample set
- 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
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0407—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis
- G08B21/043—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis detecting an emergency event, e.g. a fall
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0438—Sensor means for detecting
- G08B21/0446—Sensor means for detecting worn on the body to detect changes of posture, e.g. a fall, inclination, acceleration, gait
Landscapes
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Gerontology & Geriatric Medicine (AREA)
- Business, Economics & Management (AREA)
- Emergency Management (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Psychiatry (AREA)
- Psychology (AREA)
- Social Psychology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of wearable fall detection method and system based on hierarchical classification, the fall detection method include: to acquire the daily behavior data of user;It the processing such as synthesized, filtered to the daily behavior data and generate initial data;The time and frequency domain characteristics of the initial data are extracted using sliding window mechanism, generate sample, and the sample is combined into sample set;Each of sample set sample is identified using first layer oneclass classification model, the result after identification is sent to the weighting two classification model of the second layer;The weighting two classification model of the second layer is responsible for generating weighting tumble sample, and be passed to the regular two classification model of third layer to weight assignment processing;Whether the regular two classification model of third layer meets tumble rule according to the weighting tumble sample, judges whether user occurs tumble behavior.The present invention realizes the accurate judgement to user's tumble behavior by above method.
Description
Technical field
The present invention relates to general fit calculations and field of health care, and in particular to a kind of wearable tumble based on hierarchical classification
Detection method and system.
Background technique
January in 2016,22 portion of Ren society press spokesman Li Zhong was pointed out, ended 2014,60 years old Chinese or more elderly population
Reach 2.1 hundred million, account for the ratio 15.5% of total population, and combine national standard, 60 years old or more elderly population are considered as old-age group up to 10%
Change society.With the increasingly increase at age, the physiological organism function of the elderly gradually fails, increasingly slow to the reaction of accident
Slowly, it is easier to fall.It falls and has become the first injury cause of the death of Elder crowd, incidence is high, and damage is serious, to a
People, family or society bring great burden, have gradually become the social common problem concerning senior health and fitness.Therefore,
How tumble behavior is detected with carrying out real-time and precise, becomes a highly important social concern.
Numerous and complicated numerous wearable device increasingly pours in daily life, is widely used in health supervision, movement is protected
It is strong to wait fields.Wearable device acquires data using embedded microsensor, can effectively excavate the daily behavior of user.Together
When, wearable device has many advantages, such as that cheap, configuration is simple, portable, has weight for the challenge of reply social senilization
The realistic meaning wanted.Therefore, tool of the present invention using wearable device as research fall detection.
It falls and is used as a kind of typical abnormal behaviour, have the characteristics that its own.As shown in Figure 1, falling under normal conditions
Include three states: weightlessness hits, is static.When tumble just starts, the both feet of people can gradually leave ground and the effect in gravity
Lower freedom is fallen downwards, and people is now in state of weightlessness to a certain degree.When shock acts generation, the downward speed of body reaches
Maximum value has been arrived, has been hit suddenly with ground or other objects at this time, so that resultant acceleration moment has reached peak.?
It hits in some period after occurring, regardless of the severity fallen, people can be in a kind of opposing stationary state.This
Outside, the features such as tumble is also often accompanied with the time-constrain between the variation and each state of human body direction.Such as: human body
The variation of direction refers to that the direction of human body will be different with before hitting after shock movement occurs.
Fall detection method based on wearable device can be divided into threshold method, machine learning method and threshold method and machine learning
The combined method of method.Threshold method identified by comparing the size relation of some or several features and respective threshold it is current whether
Some state in tumble, and then judge whether current behavior is tumble.Such as patent CN201610058318.4 passes through three
Axle sensor real-time monitoring human body active state information, and vector sum is calculated, it is compared with preset threshold value, from
And judge whether to be fallen;Patent CN201610062316.2 uses the method based on characteristic quantity threshold value, calculates and closes acceleration
Spend characteristic quantity A, accumulated angle velocity characteristic amount W and similarity characteristic quantity S, with the threshold comparison that is obtained after signal vector mould integral into
Row judgement.Machine learning method regards fall detection as a typical classification problem, simultaneously based on training data learning classification model
For fall detection.Such as patent CN201610152570.1 be based on Kalman filtering and k nearest neighbor (KNearestNeighbor,
KNN) algorithm carries out classification model construction to human body active state, identifies that the type of sports of human body, judgment module are made whether " to fall
Decision ";Patent CN201610083726.5 uses support vector machines (SupportVectorMachine, SVM) algorithm structure
Build classifier;Tumble sample and daily routines behavior sample composing training collection are obtained, classifier is trained.Based on combination side
The fall detection of method often combines threshold method and machine learning method, judges tumble behavior, such as patent
CN201010285585.8 carries out secondary judgement after threshold decision, using one-class support vector machine, to judge whether it is
It falls.
The missing inspection consequence of the problem of fall detection is a cost-sensitive, i.e. tumble behavior will be extremely serious, it is desirable that mould
The lower false dismissed rate of type.On the other hand, frequent false alarm can cause the dislike of user, reduce its trust to detection system
Degree, is unfavorable for the practical application of method, the false alarm rate of modulus type is as low as possible again.Although fall detection has more side
Method, but existing method is difficult to meet the requirement of low false dismissed rate and low false alarm rate simultaneously.The reason of causing this problem mainly has three
A aspect: 1, having method and do not comprehensively consider model false dismissed rate and false alarm rate, but uses single judgment criteria (such as essence
Degree);2, using conventional machine learning classification method, the particularity for this abnormal behaviour that do not account for falling;3, due to part
The transients of daily behavior (such as run, go downstairs) and the similarity of tumble behavior are higher, and additive noise etc. is to data
It influences, reduces the accuracy rate of model inspection.
Summary of the invention
As healthy and safe important leverage, the consequence of tumble leak detection is often fatal, and frequent false alarm
User can be caused to the dislike of system.For the false dismissed rate and false alarm rate that fall detection is effectively reduced, increase fall detection method pair
The separating capacity of tumble behavior, while influence of the noise data to model is filtered, the invention proposes one kind to be based on hierarchical classification
Wearable fall detection method, wherein the fall detection method include:
Step 1, the daily behavior data of wearable motion sensor acquisition user are utilized;
Step 2, the daily behavior data of acquisition are synthesized, operation is filtered, generate initial data;
Step 3, it extracts the time and frequency domain characteristics of the initial data using sliding window mechanism, generates sample, and by the sample
It is combined into sample set;
Step 4, each of sample set sample is identified using the oneclass classification model of first layer, will be identified
As a result it is combined into tumble sample set for the tumble sample group of " tumble ", and the tumble sample set is sent to two class of weighting of the second layer
Disaggregated model;
Step 5, the weighting two classification model of the second layer is responsible for carrying out the tumble samples all in the tumble sample set
Weight assignment processing generates weighting tumble sample, and the weighting tumble sample is sent to the regular two classification mould of third layer
Type;
Step 6, whether the regular two classification model of third layer meets tumble rule, judgement according to the weighting tumble sample
Whether user occurs tumble behavior, goes to step 7 if being judged as tumble behavior, otherwise goes to step 1;
Step 7, corresponding alarm mechanism is triggered, if you need to continue to test, then 1 is gone to step, otherwise terminates.
The wearable fall detection method based on hierarchical classification, wherein
By learning to previously given sample set, the oneclass classification model is established;And with the oneclass classification model pair
Each of previously given sample set sample is identified, previously given tumble sample set is generated;
By learning to the previously given tumble sample set, the weighting two classification model is established;
By learning to previously given tumble rule, the rule two classification model is established.
The wearable fall detection method based on hierarchical classification, wherein the oneclass classification model is that supporting vector data are retouched
Model is stated, which generates a hypersphere according to the previously given sample set, and should by judgement
Whether sample is located within the hypersphere, is tumble sample by the specimen discerning if the sample is located within the hypersphere.
The wearable fall detection method based on hierarchical classification, wherein the weighting two classification model is to weight transfinite
Habit machine model, to distribute different weights from non-tumble sample to the tumble sample in the tumble sample set.
The wearable fall detection method based on hierarchical classification, wherein the tumble rule specifically,
A) weightless, during weightlessness, the value of resultant acceleration is gradually reduced by acceleration of gravity and is intended to zero;
B) hit, shock act occur when before, the downward speed of body has had reached maximum value, at this time when with ground
Face or other objects are hit suddenly, so that resultant acceleration moment has reached a peak value more than twice of acceleration of gravity
Peak, it is zero that speed, which die-offs, at this time;
C) static, the X of accelerometer, Y, Z axis reading and resultant acceleration reading are in steady state.
The present invention also provides a kind of wearable fall detection system based on hierarchical classification, wherein the fall detection system packet
It includes:
Data acquisition module, with the daily behavior data of wearable motion sensor acquisition user;
Data processing module generates original for the daily behavior data of acquisition to be synthesized, be filtered with operation
Data;
Sample generation module extracts the time and frequency domain characteristics of the initial data using sliding window mechanism, generates sample, and will
The sample is combined into sample set;
First layer identification module, which includes oneclass classification model, for every in the sample set
One sample identified, the tumble sample group that recognition result is " tumble " is combined into tumble sample set, and by the tumble sample
Collection is sent to second layer weighting block;
Second layer weighting block, which includes weighting two classification model, for the tumble sample
It concentrates all tumble samples to be weighted allocation processing, generates weighting tumble sample, and the weighting tumble sample is sent to
Third layer judgment module;
Third layer judgment module, which includes regular two classification model, for being fallen according to the weighting
Whether sample meets tumble rule, judges whether user occurs tumble behavior, 7 is gone to step if being judged as tumble behavior, instead
Then return to data acquisition module, continue acquire user daily behavior data;
Alarm triggering module if you need to continue to test, then returns to data acquisition module for triggering corresponding alarm mechanism,
The daily behavior data for continuing acquisition user, otherwise terminate.
The wearable fall detection system based on hierarchical classification, wherein
By learning to previously given sample set, the oneclass classification model is established;And with the oneclass classification model pair
Each of previously given sample set sample is identified, previously given tumble sample set is generated;
By learning to the previously given tumble sample set, the weighting two classification model is established;
By learning to previously given tumble rule, the rule two classification model is established.
The wearable fall detection system based on hierarchical classification, wherein the oneclass classification model is that supporting vector data are retouched
Model is stated, which generates a hypersphere according to the previously given sample set, and should by judgement
Whether sample is located within the hypersphere, is tumble sample by the specimen discerning if the sample is located within the hypersphere.
The wearable fall detection system based on hierarchical classification, wherein the weighting two classification model is to weight transfinite
Habit machine model, to distribute different weights from non-tumble sample to the tumble sample in the tumble sample set.
The wearable fall detection system based on hierarchical classification, wherein the tumble rule specifically,
A) weightless, during weightlessness, the value of resultant acceleration is gradually reduced by acceleration of gravity and is intended to zero;
B) hit, shock act occur when before, the downward speed of body has had reached maximum value, at this time when with ground
Face or other objects are hit suddenly, so that resultant acceleration moment has reached a peak value more than twice of acceleration of gravity
Peak, it is zero that speed, which die-offs, at this time;
C) static, the X of accelerometer, Y, Z axis reading and resultant acceleration reading are in steady state.
As it can be seen from the above scheme it is an advantage of the current invention that compared with the prior art, hierarchical classification provided by the invention falls
Detection method is using wearable sensing data as target object, is gradually solved using layer architecture each during fall detection
Item problem.By being several single goals by the challenge neutralizing of fall detection this multiple target (high detection rate, low false alarm rate)
The simple problem that (minimum surrounds ball, F-Score highest, and false alarm rate is minimum) can independently solve, recycles strategy of dividing and rule, most
The all multi-models constructed at last are interrelated with the framework being layered, and finally meet the aggregate demand of fall detection task.
Detailed description of the invention
Fig. 1 is the three phases acceleration schematic diagram of tumble behavior;
Fig. 2 is that the present invention is based on the wearable fall detection method frame diagrams of hierarchical classification;
Fig. 3 is that the present invention is based on the wearable fall detection flow charts of hierarchical classification;
Fig. 4 is SVDD model schematic of the present invention;
Fig. 5 is DT model schematic of the present invention.
Specific embodiment
To allow features described above and effect of the invention that can illustrate more clearly understandable, special embodiment below, and cooperate
Bright book attached drawing is described in detail below.
As healthy and safe important leverage, the consequence of tumble leak detection is often fatal, and frequent false alarm
User can be caused to the dislike of system.For the false dismissed rate and false alarm rate that fall detection is effectively reduced, increase fall detection method pair
The separating capacity of tumble behavior, while influence of the noise data to model is filtered, the invention proposes one kind to be based on hierarchical classification
Wearable fall detection method frame, as shown in Fig. 2, the first layer building includes the minimum sphere of all tumble samples, precisely
Lock onto target domain (distribution space of tumble sample);For the uneven degree fallen in hypersphere with non-tumble sample distribution, lead to
It crosses and different weights is respectively set to different classes of sample, the disaggregated model of the second layer building weighting, which improves the whole of two classes, to be known
Other ability;It is the sample of " tumble " for recognition result in hypersphere, third layer inspection " tumble " sample occurs whether front and back meets
Tumble dependency rule, to reduce the false alarm rate of tumble.
Firstly, acquiring user's daily behavior data, and pass through pretreatment, feature extraction by wearable motion sensor
The daily behavior data processing is sample set, fallen comprising tumble sample with non-in the sample set in this stage by equal pretreatment early periods
Sample, the tumble sample include determining tumble sample and doubtful tumble sample, and with the oneclass classification model of first layer to this
Each of sample set sample is identified, and the tumble sample that recognition result is " tumble " is sent to the weighting of the second layer
Two classification model is further processed.Wherein the method for building up of the oneclass classification model is, to previously given sample set into
Row study.Each of previously given sample set sample is identified with the oneclass classification model built up, is generated preparatory
Given tumble sample set, so that second layer weighting two classification model is learnt.The previously given sample set can be big data
The lower sample set of statistics, next weighting two classification model and rule two classification model establish thinking be also in this way,
The oneclass classification model for identification result be " tumbles " tumble data sample, as entirely wearable fall detection method frame
The first layer of frame, it is therefore an objective to reduce tumble false dismissed rate;
Secondly, the weighting two classification model of the second layer receives the tumble sample, and to all tumble samples
It is weighted allocation processing, generates weighting tumble sample, it is therefore an objective to so that all tumble samples have higher identification capability,
And the regular two classification model that the weighting tumble sample is sent to third layer is further processed.Wherein two class of weighting
The method for building up of disaggregated model is to learn to the previously given tumble sample set.The weighting two classification model is as whole
The second layer of a wearable fall detection method frame, it is therefore an objective to improve overall accuracy of identification;
Finally, the regular two classification model of third layer is analyzed and determined and is used according to the weighting tumble sample and rule of falling
Whether family occurs tumble behavior, triggers corresponding alarm mechanism if being judged as tumble.The wherein rule two classification model
Method for building up is to learn to previously given tumble rule.The rule two classification model is as entire wearable tumble inspection
Survey the third layer of method frame, it is therefore an objective to the testing result for filtering, smoothly falling, so that the false alarm rate fallen further decreases.
Wherein the analytical judgment process is the adjacent sample judged in the weighting tumble sample, if meet the tumble rule (such as:
Human body is generally in lying status after shock, and hitting front and back human body direction would generally change, in a period of time after tumble
Human body is in relative static conditions), it is judged as that tumble behavior occurs for user if meeting, if being unsatisfactory for being judged as that user does not have
Tumble behavior occurs.
Furthermore, it is understood that the wearable fall detection flow chart based on hierarchical classification is as shown in figure 3, mainly include following step
It is rapid:
Step 1, the daily behavior data of the wearable motion sensor acquisition user such as accelerometer, gyroscope are utilized;
Step 2, it the processing operations such as synthesized, filtered to the daily behavior data of acquisition, generate initial data;
Step 3, it extracts the time and frequency domain characteristics of the initial data using sliding window mechanism, generates sample, and by the sample
It is combined into sample set, it should be noted that the sample that this step generates includes tumble sample and non-tumble sample, the tumble sample
It can be used for training pattern or test;
Step 4, each of sample set sample is identified using the oneclass classification model of first layer, will be identified
As a result it is combined into tumble sample set for the tumble sample group of " tumble ", and the tumble sample set is sent to two class of weighting of the second layer
Disaggregated model, wherein 5 are gone to step if recognition result is " tumble ", conversely, 1 is then gone to step, the structure of the oneclass classification model
The process of building can be, based on the tumble sample in previous sample set, learning training be carried out, in the first layer building oneclass classification mould
Type;
It step 5, is the tumble sample of " tumble ", two class of weighting of the second layer for recognition result in first layer (step 4)
Disaggregated model receives the tumble sample set, and is weighted at distribution to the tumble samples all in the tumble sample set
Reason generates weighting tumble sample, and the regular two classification model that the weighting tumble sample is sent to third layer is carried out into one
Step processing;
Step 6, for the weighting tumble sample of sound field in the second layer (step 5), the regular two classification mould of third layer
Whether type meets tumble rule according to the weighting tumble sample, analyzes and determines whether user occurs tumble behavior, falls if being judged as
Then go to step 7, it is on the contrary then go to step 1;
Step 7, corresponding alarm mechanism is triggered, if you need to continue to test, then 1 is gone to step, otherwise terminates.
Wherein the oneclass classification model can be Support Vector data description model, the Support Vector data description model according to
The previously given sample set generates a hypersphere, and by judging whether the sample is located within the hypersphere, if the sample
It is then tumble sample by the specimen discerning within the hypersphere;The weighting two classification model is to weight the learning machine that transfinites
Model, to distribute different weights from non-tumble sample to the tumble sample in the tumble sample set.
In the fall detection frame based on hierarchical classification, it is related to following three models:
1, oneclass classification model.The present invention mainly constructs a Support Vector data description (Support Vector
Domain Description, SVDD) model.Asking a center is a, and radius is the hypersphere of R, is including as much as possible fall
The while of falling training sample, requires the radius of a ball small as far as possible;
2, two classification model is weighted.The present invention mainly constructs one and weights learning machine (Weighted ELM) mould that transfinites
Type.Learnt based on the training set fallen with non-tumble data nonbalance by distributing different weights to different classes of sample
One single layer BP network model, so that F value (F-Score) value of model is high as far as possible;
3, rule-based two classification model.The present invention passes through judgement for the particularity for this abnormal behaviour of falling
Human body is generally in lying status after shock, and hitting front and back human body direction would generally change, in a period of time after tumble
Human body is in relative static conditions etc. and constructs a rule set, to filter out doubtful tumble as much as possible (puppet is fallen);
The present invention is not limited to the above method, any kind sorting algorithm is can be used in the SVDD of first layer;The second layer
Any weighting two classification algorithm can be used in WeightedELM;Decision can be used in the rule-based two classification model of third layer
Set (Decision Tree, DT), random forest (Random Forest, RF) or other rule-based sorting algorithms;Second
The F-Score of layer can use ROC, the module of other two class data set entirety recognition capabilities of measurement such as G-Means.
From top to bottom include three models based on three layers of fall detection methods: oneclass classification model SVDD weights two classes
It is situated between in detail below disaggregated model WeightedELM, the two classification model DT. based on rule of falling to these three models
It continues.
One, oneclass classification model SVDD
The target of SVDD is the accurate distributed areas for obtaining tumble sample, and basic thought is one hypersphere of study, package
Tumble sample as much as possible, while the radius of ball will not be too big.The model schematic of SVDD such as Fig. 4 institute in two-dimensional space
Show, the circle in Fig. 4 indicates that tumble sample, fork fork indicate non-tumble sample.
It gives by N1Training set { the x that a tumble sample is constitutedi∈Rd| i=1 ..., N1, wherein xiIndicate i-th of trained sample
This, characteristic dimension d.The process for solving SVDD is to ask a center for a, and radius is the spherical surface of R.The optimization problem of SVDD is such as
Under:
Wherein ξiIt indicates slack variable, measures sample xiMistake in the training process divides degree, and ξ is indicated by N1A ξiComposition
N1Dimensional vector;T representing matrix transposition;C is punishment parameter, maximizes two for weighing empirical risk minimization and generalization ability
A target.In the forecast period of SVDD, a sample z ∈ R is judgeddWhether it is tumble, that is, sees whether sample z is located at hypersphere
Within, i.e. judgement (z-a)T(z-a)≤R2It is whether true, if so, then the recognition result of SVDD is " tumble " by the specimen discerning
For tumble sample;Conversely, be then " non-tumble " by the specimen discerning being non-tumble sample.It is all by SVDD to be determined as " falling "
Tumble sample can then enter WeightedELM and carry out secondary judgement.
Two, two classification model WeightedELM is weighted
After SVDD, aiming field (distributed areas of tumble sample) has been reduced to a relatively small subregion.
In the training process of hierarchy model, only those the training sample " fallen " is determined as by the SVDD of first layer just can be with
Into the second layer, training weighting two classification model WeightedELM is participated in.However " tumble " here is SVDD to sample
Predict classification, and the true classification of non-sample.Since the training process of SVDD requires to cause comprising sample of falling as much as possible
The non-tumble sample in part is also enclosed in hypersphere (see Fig. 4).The training process of WeightedELM is with all in hypersphere
It falls with non-tumble sample as training set.For all tumble samples for including in aiming field and a small amount of non-tumble sample group
At uneven set, different weights is distributed to different classes of sample first, based on the training sample set building after weighting
WeightedELM guarantees so that the G-Means of model is as high as possible to two categories in aiming field (falling and non-tumble)
Sample identification capability all with higher.
It is assumed that tumble all in aiming field and non-tumble sample xjTotal N2A, the set of composition is by collecting { (xj, tj)∈Rd
×R2| j=1 ..., N2Indicate, wherein tjIndicate j-th of sample xjClass label, value be (1, -1)TOr (- 1,1)T, point
It Biao Shi not sample xjFor tumble or non-tumble sample.The optimization problem of WeightedELM is as follows:
Wherein β is the independent variable to be solved;H (x) is that sample x is mapped to some higher dimensional space from original d dimension space
Mapping function;" one, oneclass classification model SVDD " part is shown in the explanation of punishment parameter;W is N2×N2Diagonal weight matrix, W's
Diagonal entry WjjIndicate sample xjWeight.Compared with the training sample of major class, the training sample of group would generally be according to certain
Kind rule assigns the biggish weight that compares.
Three, the two classification model DT based on rule of falling
The distributed areas of tumble sample can be effectively obtained using SVDD;Mesh can be precisely distinguished using Weighted ELM
Mark the tumble in domain and non-tumble sample;However due to the transients of individual normal behaviours (such as run, go downstairs) with fall
Similarity is higher, and the influence of additive noise data, the false alarm rate of fall detection still may be higher.Therefore, DT is used for
It is for re-filtering to the sample that Weighted ELM recognition result is " tumble ", reduce false alarm rate.
It falls as a kind of special abnormal behaviour, there are some specific rules.According to several stages of tumble behavior
(as shown in Figure 1) as can be seen that conventional tumble behavior occurs, front and back is usual to have following rule: human body is generally in after shock
Human body is in opposing stationary shape in a period of time after human body direction would generally change, fall before and after lying status, shock
State etc., measurement means that the present invention uses (rule of falling) is using the motion sensor acquisition data worn and feel
The motion conditions for knowing human body carry out analysis to data according to the distinctive rule of tumble behavior and whether final decision fall
?.By taking the three axis accelerometer for being worn on waist as an example, the index being commonly used is the X of accelerometer, Y, the original reading of Z axis
Several and resultant acceleration (X, Y, the root mean square of Z axis reading) reading.Specifically the present invention is utilized and is worn on user
One or more motion sensors can be with the motion conditions of effectively perceive user, to detect whether user is fallen.I
The typical characteristics of this abnormal behaviour of falling are introduced by taking the acceleration readings of waist as an example.A) weightless.When tumble just starts, people
Both feet can gradually leave ground and under gravity freely fall downwards, people is now in weightless shape to a certain degree
State.Since only by gravity, the downward speed of body can be gradually increased before making impact with the ground.During weightlessness, close
It is gradually reduced at the value of acceleration by acceleration of gravity (1g) and is intended to zero (see " root mean square " void under Fig. 1 state of weightlessness
Line).B) it hits.Shock act occur when before, the downward speed of body has had reached maximum value, at this time when with ground or
Other objects are hit suddenly, so that it is more than the peak value of 2g (see under Fig. 1 crash situation that resultant acceleration moment, which has reached one,
" root mean square " dotted line) peak, it is zero that speed, which die-offs, at this time.C) it is hitting in some period after occurring, no matter is falling
Severity how, people can be embodied under the stationary state in Fig. 1, acceleration in a kind of opposing stationary state
The X of meter, Y, Z axis reading and resultant acceleration reading are in a kind of relatively stable state.It is also frequently accompanied by addition, falling
The time-constrain between the variation of human body direction and each state the features such as.Such as: the variation of human body direction refers to, hits
It hits the direction of human body after movement occurs and will be different that (see X in Fig. 1, Y, Z axis reading is before and after hitting and occurring before hitting
Sign change situation).Dedicated rule set is constructed using DT, filters out the normal behaviour of doubtful tumble and the sample of noise,
It is final to reduce false alarm rate.The schematic diagram of DT model is as shown in Figure 5.
The following are system embodiment corresponding with above method embodiment, present embodiment can be mutual with above embodiment
Cooperation is implemented.The above-mentioned relevant technical details mentioned in mode of applying are still effective in the present embodiment, in order to reduce repetition, this
In repeat no more.Correspondingly, the relevant technical details mentioned in present embodiment are also applicable in above embodiment.
The present invention also provides a kind of wearable fall detection system based on hierarchical classification, wherein the fall detection system packet
It includes:
Data acquisition module, using the daily behavior data of wearable motion sensor acquisition user;
Data processing module generates original for the daily behavior data of acquisition to be synthesized, be filtered with operation
Data;
Sample generation module extracts the time and frequency domain characteristics of the initial data using sliding window mechanism, generates sample, and will
The sample is combined into sample set;
First layer identification module, which includes oneclass classification model, for every in the sample set
One sample identified, the tumble sample group that recognition result is " tumble " is combined into tumble sample set, and by the tumble sample
Collection is sent to second layer weighting block;
Second layer weighting block, which includes weighting two classification model, for the tumble sample
It concentrates all tumble samples to be weighted allocation processing, generates weighting tumble sample, and the weighting tumble sample is sent to
Third layer judgment module;
Third layer judgment module, which includes regular two classification model, for being fallen according to the weighting
Whether sample meets tumble rule, judges whether user occurs tumble behavior, 7 is gone to step if being judged as tumble behavior, instead
Then return to data acquisition module, continue acquire user daily behavior data;
Alarm triggering module if you need to continue to test, then returns to data acquisition module for triggering corresponding alarm mechanism,
The daily behavior data for continuing acquisition user, otherwise terminate.
The wearable fall detection system based on hierarchical classification, wherein
By learning to previously given sample set, the oneclass classification model is established;And with the oneclass classification model pair
Each of previously given sample set sample is identified, previously given tumble sample set is generated;
By learning to the previously given tumble sample set, the weighting two classification model is established;
By learning to previously given tumble rule, the rule two classification model is established.
The wearable fall detection system based on hierarchical classification, wherein the oneclass classification model is that supporting vector data are retouched
Model is stated, which generates a hypersphere according to the previously given sample set, and should by judgement
Whether sample is located within the hypersphere, is tumble sample by the specimen discerning if the sample is located within the hypersphere.
The wearable fall detection system based on hierarchical classification, wherein the weighting two classification model is to weight transfinite
Habit machine model, to distribute different weights from non-tumble sample to the tumble sample in the tumble sample set.
The wearable fall detection system based on hierarchical classification, wherein the tumble rule specifically,
A) weightless, during weightlessness, the value of resultant acceleration is gradually reduced by acceleration of gravity and is intended to zero;
B) hit, shock act occur when before, the downward speed of body has had reached maximum value, at this time when with ground
Face or other objects are hit suddenly, so that resultant acceleration moment has reached a peak value more than twice of acceleration of gravity
Peak, it is zero that speed, which die-offs, at this time;
C) static, the X of accelerometer, Y, Z axis reading and resultant acceleration reading are in steady state.
Although the present invention is disclosed with above-described embodiment, specific examples are only used to explain the present invention, is not used to limit
The present invention, any those skilled in the art of the present technique without departing from the spirit and scope of the invention, can make some change and complete
It is kind, therefore the scope of the present invention is subject to claims.
Claims (10)
1. a kind of wearable fall detection method based on hierarchical classification, which is characterized in that the fall detection method includes:
Step 1, the daily behavior data of wearable motion sensor acquisition user are utilized;
Step 2, the daily behavior data of acquisition are synthesized, operation is filtered, generate initial data;
Step 3, the time and frequency domain characteristics of the initial data are extracted using sliding window mechanism, generate sample, and combine the sample
At sample set;
Step 4, each of sample set sample is identified using the oneclass classification model of first layer, by recognition result
It is combined into tumble sample set for the tumble sample group of " tumble ", and the tumble sample set is sent to the weighting two classification of the second layer
Model;
Step 5, the weighting two classification model of the second layer is responsible for being weighted the tumble samples all in the tumble sample set
Allocation processing generates weighting tumble sample, and the weighting tumble sample is sent to the regular two classification model of third layer;
Step 6, whether the regular two classification model of third layer meets tumble rule according to the weighting tumble sample, judges user
Whether generation tumble behavior, go to step 7 if being judged as tumble behavior, otherwise go to step 1;
Step 7, corresponding alarm mechanism is triggered, if you need to continue to test, then 1 is gone to step, otherwise terminates.
2. the wearable fall detection method based on hierarchical classification as described in claim 1, which is characterized in that
By learning to previously given sample set, the oneclass classification model is established;And it is pre- to this with the oneclass classification model
It first gives each of sample set sample to be identified, generates previously given tumble sample set;
By learning to the previously given tumble sample set, the weighting two classification model is established;
By learning to previously given tumble rule, the rule two classification model is established.
3. the wearable fall detection method based on hierarchical classification as claimed in claim 2, which is characterized in that the oneclass classification
Model is Support Vector data description model, which generates one according to the previously given sample set
A hypersphere, and judge whether the sample is located within the hypersphere, if the sample is located within the hypersphere, by the sample
It is identified as tumble sample.
4. the wearable fall detection method based on hierarchical classification as claimed in claim 2, which is characterized in that two class of weighting
Disaggregated model is to weight the learning machine model that transfinites, with different from non-tumble sample distribution to the tumble sample in the tumble sample set
Weight.
5. the wearable fall detection method based on hierarchical classification as described in claim 1, which is characterized in that tumble rule
Specifically,
A) weightless, during weightlessness, the value of resultant acceleration is gradually reduced by acceleration of gravity and is intended to zero;
B) hit, shock movement occur before, the downward speed of body has had reached maximum value, at this time when with ground or other
Object is hit suddenly, so that resultant acceleration moment has reached the highest of a peak value more than twice of acceleration of gravity
Value, it is zero that speed, which die-offs, at this time;
C) static, the X of accelerometer, Y, Z axis reading and resultant acceleration reading are in steady state.
6. a kind of wearable fall detection system based on hierarchical classification, which is characterized in that the fall detection system includes:
Data acquisition module, with the daily behavior data of wearable motion sensor acquisition user;
Data processing module generates original number for the daily behavior data of acquisition to be synthesized, be filtered with operation
According to;
Sample generation module, extracts the time and frequency domain characteristics of the initial data using sliding window mechanism, generates sample, and by the sample
Originally it is combined into sample set;
First layer identification module, the first layer identification module include oneclass classification model, for each of the sample set
Sample is identified, the tumble sample group that recognition result is " tumble " is combined into tumble sample set, and the tumble sample set is sent out
It send to second layer weighting block;
Second layer weighting block, which includes weighting two classification model, for in the tumble sample set
All tumble samples are weighted allocation processing, generate weighting tumble sample, and the weighting tumble sample is sent to third
Layer judgment module;
Third layer judgment module, which includes regular two classification model, for according to the weighting tumble sample
Whether this meets tumble rule, judges whether user occurs tumble behavior, turns alarm triggering module if being judged as tumble behavior,
It is on the contrary then call the data acquisition module, continue the daily behavior data for acquiring user;
Alarm triggering module if you need to continue to test, then returns to data acquisition module for triggering corresponding alarm mechanism, continues
The daily behavior data for acquiring user, otherwise terminate.
7. the wearable fall detection system based on hierarchical classification as claimed in claim 6, which is characterized in that
By learning to previously given sample set, the oneclass classification model is established;And it is pre- to this with the oneclass classification model
It first gives each of sample set sample to be identified, generates previously given tumble sample set;
By learning to the previously given tumble sample set, the weighting two classification model is established;
By learning to previously given tumble rule, the rule two classification model is established.
8. the wearable fall detection system based on hierarchical classification as claimed in claim 7, which is characterized in that the oneclass classification
Model is Support Vector data description model, which generates one according to the previously given sample set
A hypersphere, and by judging whether the sample is located within the hypersphere, it, should if the sample is located within the hypersphere
Specimen discerning is tumble sample.
9. the wearable fall detection system based on hierarchical classification as claimed in claim 7, which is characterized in that two class of weighting
Disaggregated model is to weight the learning machine model that transfinites, with different from non-tumble sample distribution to the tumble sample in the tumble sample set
Weight.
10. the wearable fall detection system based on hierarchical classification as claimed in claim 6, which is characterized in that tumble rule
Then specifically,
A) weightless, during weightlessness, the value of resultant acceleration is gradually reduced by acceleration of gravity and is intended to zero;
B) hit, shock act occur when before, the downward speed of body has had reached maximum value, at this time when with ground or
Other objects are hit suddenly, so that resultant acceleration moment has reached a peak value more than twice of acceleration of gravity most
High level, it is zero that speed, which die-offs, at this time;
C) static, the X of accelerometer, Y, Z axis reading and resultant acceleration reading are in steady state.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710144772.6A CN106875630B (en) | 2017-03-13 | 2017-03-13 | A kind of wearable fall detection method and system based on hierarchical classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710144772.6A CN106875630B (en) | 2017-03-13 | 2017-03-13 | A kind of wearable fall detection method and system based on hierarchical classification |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106875630A CN106875630A (en) | 2017-06-20 |
CN106875630B true CN106875630B (en) | 2018-12-04 |
Family
ID=59170123
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710144772.6A Active CN106875630B (en) | 2017-03-13 | 2017-03-13 | A kind of wearable fall detection method and system based on hierarchical classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106875630B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108960056B (en) * | 2018-05-30 | 2022-06-03 | 西南交通大学 | Fall detection method based on attitude analysis and support vector data description |
CN109993065B (en) * | 2019-03-06 | 2022-08-23 | 开易(北京)科技有限公司 | Driver behavior detection method and system based on deep learning |
CN110279420A (en) * | 2019-07-18 | 2019-09-27 | 郑州轻工业学院 | Portable falling detection device and detection method based on extreme learning machine |
CN110659595A (en) * | 2019-09-10 | 2020-01-07 | 电子科技大学 | Tumble type and injury part detection method based on feature classification |
CN112580403A (en) * | 2019-09-29 | 2021-03-30 | 北京信息科技大学 | Time-frequency feature extraction method for fall detection |
CN111710129B (en) * | 2020-06-12 | 2021-06-01 | 电子科技大学 | Real-time pre-collision falling detection method for old people |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102302370A (en) * | 2011-06-30 | 2012-01-04 | 中国科学院计算技术研究所 | Method and device for detecting tumbling |
CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
CN103417219A (en) * | 2013-09-11 | 2013-12-04 | 重庆大学 | Wearable human body falling detection device |
CN104125337A (en) * | 2014-07-22 | 2014-10-29 | 厦门美图移动科技有限公司 | Smart phone falling detection and alarming method |
CN104510475A (en) * | 2014-12-15 | 2015-04-15 | 中国科学院计算技术研究所 | Human body falling-down detection method and system |
CN104637242A (en) * | 2013-11-12 | 2015-05-20 | 广州华久信息科技有限公司 | Elder falling detection method and system based on multiple classifier integration |
CN105310696A (en) * | 2015-11-06 | 2016-02-10 | 中国科学院计算技术研究所 | Fall detection model construction method as well as corresponding fall detection method and apparatus |
CN106388831A (en) * | 2016-11-04 | 2017-02-15 | 郑州航空工业管理学院 | Method for detecting falling actions based on sample weighting algorithm |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI402778B (en) * | 2010-07-22 | 2013-07-21 | Univ China Medical | Fall warning system |
-
2017
- 2017-03-13 CN CN201710144772.6A patent/CN106875630B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102302370A (en) * | 2011-06-30 | 2012-01-04 | 中国科学院计算技术研究所 | Method and device for detecting tumbling |
CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
CN103417219A (en) * | 2013-09-11 | 2013-12-04 | 重庆大学 | Wearable human body falling detection device |
CN104637242A (en) * | 2013-11-12 | 2015-05-20 | 广州华久信息科技有限公司 | Elder falling detection method and system based on multiple classifier integration |
CN104125337A (en) * | 2014-07-22 | 2014-10-29 | 厦门美图移动科技有限公司 | Smart phone falling detection and alarming method |
CN104510475A (en) * | 2014-12-15 | 2015-04-15 | 中国科学院计算技术研究所 | Human body falling-down detection method and system |
CN105310696A (en) * | 2015-11-06 | 2016-02-10 | 中国科学院计算技术研究所 | Fall detection model construction method as well as corresponding fall detection method and apparatus |
CN106388831A (en) * | 2016-11-04 | 2017-02-15 | 郑州航空工业管理学院 | Method for detecting falling actions based on sample weighting algorithm |
Non-Patent Citations (1)
Title |
---|
基于极限学习机的跌倒检测分类识别研究;王之琼等;《中国生物医学工程学报》;20140831;第33卷(第4期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN106875630A (en) | 2017-06-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106875630B (en) | A kind of wearable fall detection method and system based on hierarchical classification | |
CN108244744A (en) | A kind of method of moving state identification, sole and footwear | |
CN105310696B (en) | A kind of fall detection model building method and corresponding fall detection method and device | |
CN101194273B (en) | Spatio-temporal self organising map | |
CN103955699B (en) | A kind of real-time fall events detection method based on monitor video | |
CN108446733A (en) | A kind of human body behavior monitoring and intelligent identification Method based on multi-sensor data | |
CN103211599A (en) | Method and device for monitoring tumble | |
CN104269025B (en) | Wearable single node feature and the position choosing method of monitoring is fallen down towards open air | |
CN107212890A (en) | A kind of motion identification and fatigue detection method and system based on gait information | |
CN106407996A (en) | Machine learning based detection method and detection system for the fall of the old | |
CN106981174A (en) | A kind of Falls Among Old People detection method based on smart mobile phone | |
CN111832608A (en) | Multi-abrasive-particle identification method for ferrographic image based on single-stage detection model yolov3 | |
CN106405384A (en) | Simulation circuit health state evaluation method | |
CN106725495A (en) | A kind of fall detection method, apparatus and system | |
CN107169512A (en) | The construction method of HMM SVM tumble models and the fall detection method based on the model | |
Malshika Welhenge et al. | Human activity classification using long short-term memory network | |
CN111931568A (en) | Human body falling detection method and system based on enhanced learning | |
JP2008086479A (en) | Physical activity measuring system | |
CN110169774A (en) | A kind of moving state identification system and method based on block chain | |
CN106960541A (en) | Adaptive multistage fall detection method | |
CN111582361A (en) | Human behavior recognition method based on inertial sensor | |
CN108182410A (en) | A kind of joint objective zone location and the tumble recognizer of depth characteristic study | |
CN111710129B (en) | Real-time pre-collision falling detection method for old people | |
CN116840835B (en) | Fall detection method, system and equipment based on millimeter wave radar | |
CN106161458A (en) | Network inbreak detection method based on the online extreme learning machine of double weightings |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |