CN108320456A - It is a kind of fusion multisensor the elderly fall down prediction technique and system - Google Patents
It is a kind of fusion multisensor the elderly fall down prediction technique and system Download PDFInfo
- Publication number
- CN108320456A CN108320456A CN201810080170.3A CN201810080170A CN108320456A CN 108320456 A CN108320456 A CN 108320456A CN 201810080170 A CN201810080170 A CN 201810080170A CN 108320456 A CN108320456 A CN 108320456A
- Authority
- CN
- China
- Prior art keywords
- trunk
- prediction
- angle
- elderly
- fusion
- 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.)
- Pending
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)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
- User Interface Of Digital Computer (AREA)
Abstract
The invention discloses a kind of the elderlys of fusion multisensor to fall down prediction technique and system, belong to assistant robot and rehabilitation system technical field, human hands tactile data, human body 3-axis acceleration information, trunk angle information are acquired by three kinds of sensors, it carries out falling down tentative prediction respectively using these three kinematics character parameters during falling over of human body, be merged finally by three kinds of decision level fusion center pair information.By detecting tactile and trunk movable information and carrying out Multi-information acquisition, the current motion state of human body can be quickly detected in 300ms, and rate of accuracy reached has relatively high efficiency and accuracy rate to falling down prediction, can quickly and accurately judge the elderly's physical condition to 95%.
Description
Technical field
The invention belongs to assistant robot and rehabilitation system technical fields, and in particular to a kind of old age of fusion multisensor
People falls down prediction technique and system.
Background technology
Most of handicapped the elderlys need to look after and be nursed in life, expend a large amount of social resources.And with
The increase at age, the elderly's body physiological function is substantially change, and shows the decline of organ aging and function.Wherein originally
The agings such as body-sensing feel, vision cause the elderly to be easy to happen and fall down when use is helped the elderly with row robot, not only the elderly are made to feel
To frightened and anxiety, and its body can be enabled to cause serious damage after falling down, great puzzlement is brought to its life.
Falling down for human body is a process complicated and changeable, including the movement of human limb and the movement of trunk, passes through biography
The hand tactile data and trunk angle information of sensor acquisition people describes the current motion state of human body, passes through vision and body early period
Body state predicts falling down for human body, judges overlong time, and people finally lies low from body is unbalance to making impact with the ground
The entire process of falling down on the ground only has 2~6s.So such method cannot quickly carry out prediction and fall down, by detecting tactile
It goes forward side by side row information fusion with trunk movable information, can realize that accuracy is 95% to detect that human body is current in 300ms
Motion state, predict falling down.Therefore it falls down with higher learning value and answers using this method prediction the elderly
Use meaning.
The relevant technologies in the prior art in terms of falling over of human body prediction are as follows:
1, application No. is 201610202912.6 patent application " one kind falling down prediction technique and system ", disclosed interior
Hold:The image information in front of user corresponding to environment is obtained by monitoring;According to described image information, environment in front of user is determined
In whether include the interested article of user when it includes the interested article of user to determine in front of user environment, in order to prevent
Baby quickly runs to the interested article and causes to fall down, and starts level-one early warning operation, to remind caregiver in time to baby
Youngster protects, so as to reduce the probability that baby falls down.This method and system application are baby, and the elderly and baby
The physical feature of youngster and the upper difference of cognition are too big, therefore are not suitable for the elderly and fall down prediction.
2, it proposes one kind application No. is 201410505266.1 patent application " a kind of falls down supervising device " and falling down prison
Device is controlled, the heart rate of human body is obtained by this kind of device and aerobic intensity value monitors the physical condition of human body, this kind of device
The physical condition of detection human body is needed, and for predicting that the characteristic parameter of falling over of human body only has angle data, for the elderly
For, physical condition is more complicated, and it is too simple judge to only rely on angle data and heart rate.And judgement time mistake
It is long, be not suitable for the elderly and fall down prediction.
Invention content
The purpose of the present invention is to provide a kind of the elderlys of fusion multisensor to fall down prediction technique and system, this method
By detecting tactile and trunk movable information and carrying out Multi-information acquisition, human body can be quickly detected in 300ms and worked as
Preceding motion state, and rate of accuracy reached has relatively high efficiency and accuracy rate to 95% to falling down prediction, it can accurately quickly
The elderly's physical condition is judged.
The present invention is to be achieved through the following technical solutions:
The invention discloses a kind of the elderlys of fusion multisensor to fall down prediction technique, includes the following steps:
Step 1:Acquire signal
By touch sensor, 3-axis acceleration sensor and gyroscope, user's hand tactile data, people are acquired respectively
Body 3-axis acceleration information and trunk angle information;
Step 2:Extract key feature information
By establishing coordinate system between ground and human body, the elderly hand haptic force, three axis of trunk during falling down are chosen
Characteristic parameter in three kinds of dynamic-change informations of acceleration and trunk angle is as the key feature for falling down prediction;
Step 3:Key feature information fusion judges
To the key characterization parameter of the hand haptic force, trunk 3-axis acceleration and trunk angle extracted, use respectively
Corresponding prediction algorithm of falling down carries out tentative prediction to the trend of falling down of user, and obtains three kinds of tentative prediction as a result, then
Decision level fusion is carried out to three kinds of results of tentative prediction again, finally determines whether there is and falls down possibility.
User's hand tactile data, human body 3-axis acceleration information and trunk angle information are why selected, is
Because:First, people falls the damage that bigger is caused to body in order to prevent when falling down, often with the grasping hand of four limbs instinct
Neighbouring support directly falls down born energy to reduce body.Therefore hand haptic force is chosen as judgement falling over of human body
One of motion feature.Second, since haptic force can not characterize trunk feature, and human body is when falling down, human body
Trunk 3-axis acceleration can change obviously, therefore choose motion feature of the three axis resultant acceleration of trunk as judgement falling over of human body
One of.Third, in normal activity, large range of variation can also occur human body in space for trunk, only add by three axis of trunk
Velocity estimated, which is fallen down, will produce erroneous judgement, and trunk angle can be to describe the position and attitude of trunk in space, therefore select
Take trunk angle as one of the motion feature of judgement falling over of human body.Therefore, relative motion feature packet during falling down
It includes:3-axis acceleration variation, the variation of trunk angle change, hand grip
Further, hand is acquired by the array tactile sensor being made of 8 road PVDF piezoelectric film sensor units
Haptic force, and establish feature vector R=[ai,bi,ci,di,ei,fi,gi,hi], the tactile that element is acquired by each road sensor
Power numerical value Fi, i=1,2,3...8
Body trunk coordinate system Oxyz and earth axes OXYZ are established, acceleration change is metastomium along the x-axis direction
ax, the acceleration of metastomium along the y-axis direction is ay, the acceleration of metastomium along the z-axis direction is az, then
Resultant acceleration is:
Define rotational angle theta of the trunk relative to earth axes OXYZ1, θ2, θ3, wherein θ1It is angle of heel around x-axis rotation;θ2Around
Y-axis rotation is pitch angle, θ3It is spin angle around z-axis rotation.
Further, the prediction algorithm of falling down in step 3 includes:Based on hand touch sensor fall down tentative prediction,
Falling down tentative prediction, tentative prediction fallen down based on gyroscope based on trunk 3-axis acceleration sensor.
Further, the tentative prediction of falling down based on hand touch sensor, concrete operations are:
Touch sensor shares 8 tunnels, sets effective signal threshold value FT, work as Fi> FT, input signal is effective.Under proper motion,
Its useful signal number of active lanes N=4;When falling down generation, useful signal number N > 4.Build haptic force feature vector R=[ai,
bi,ci,di,ei,fi,gi,hi], setting threshold value NT, as N > NT, it is determined with and falls down, and is fallen to decision level fusion center input prediction
Possibility T.
Further, the tentative prediction of falling down based on trunk 3-axis acceleration sensor, concrete operations are:
The first step:Every section of T of characterization is determined with maximum distance methodsThe characteristic acceleration value a of interior motion state characteristicc, definition
One reference tape B of resultant acceleration is B=[b1,b2], setting B=[9.5m/s2,11.5m/s2]。
T is determined by maximum distance methodsInterior characteristic acceleration value ac.Define the resultant acceleration value a of t momentiAt a distance from B:
d(ai, B)=| ai-b1|+|ai-b2|
Then TsInterior m resultant acceleration value aiWith the maximum distance of B:
max[d(ai, B) | i=1,2 ..m]
Second step:By acCodomain carry out break sign, generate the element { c of sliding window SiI=1,2...n.
acCodomain θ be defined as:
θ=[θ0,θ1)∪[θ1,θ2)∪···∪[θk-1,θk),θ0=0, θk→+∞
Symbol finite aggregate is defined as:
S={ s1,s2,···,sk},k≥1
It is with S correspondences by each section of codomain θ:The k=8 of symbol S.
Further, the foundation that HMM is carried out to falling over of human body process, by the mistake of initial equilibrium conditions to out-of-balance condition
Cheng Jianli mold segments λP=(M, N, π, A, B).
Institute established model λPThe parameter of=(M, N, π, A, B) is described as follows:
①M:Implicit motion state quantity during falling down.M=3.
②N:Observation quantity, N=8.
③π:Initial probability distribution.π={ πi, i=1,2, M, wherein:
④A:User fall down during hidden state-transition matrix, A={ aij, i, j=1,2, M, wherein
⑤B:Observing matrix.The relationship between each hidden state of process and observation is fallen down in description.B={ bjk, j=1,
2, M, k=1,2, N, wherein:
Final HMM is:
The each sample for the different motion processes that statistical sample is concentrated is subjected to feature extraction and obtains acceleration time sequence
Then row are used as observation sequence O, calculate output probability P (O | λp), prediction discrimination threshold is fallen down using statistical method determination,
Entire judgement time t < 300ms.
Criterion is:As P > P1When, it is believed that there is risk of falls, and judging result is inputted to decision level fusion center, it is defeated
It is output probability P obtained by HMM model to enter value.
Further, tentative prediction is fallen down based on gyroscope, the angular speed rotated around x-axis is obtained by gyroscope
ω1With the angular velocity omega rotated around y-axis2To calculate the angle of heel and pitch angle of body trunk.
It calculates as follows:
Wherein, n1=int (t1/ T), n2=int (t2/T)。
It takes and falls down set of data samples ΩF, calculate people's first impact peak value of distance during unbalance in each sample
The inclination angle theta of 300ms, is recorded as set omegaθ1.Then proper motion mode data sample set Ω is takenO, sample is calculated each
Moment trunk deviates the inclination angle of Z axis, is denoted as set omegaθ2。
Threshold θ is then obtained by the calculating process of SVM (support vector machines) methodT
The angle of heel θ of trunkYAnd pitching angle thetaPAbsolute value in there are one be more than θTWhen, which, which has been determined, falls
It is possible, and fall down possibility Z to the input of decision level fusion center.
Further, in step 3, determined to three kinds of results of tentative prediction using BP neural network blending algorithm
Plan grade merges, i.e., three kinds of results of tentative prediction is input to BP neural network, decision level fusion is carried out, to realize that the elderly falls
Accurate prediction.
BP neural network blending algorithm why is selected, is since human motion has certain complexity, three kinds of movements
Characteristic parameter and human motion state are a kind of non-linear relations, and carrying out fusion to information needs higher robustness and fault-tolerant
Rate, and since the process fallen down is complicated and changeable, blending algorithm is needed with self-learning function, therefore, chooses BP neural network
Algorithm is as decision level fusion algorithm.
Further, the used algorithm in decision level fusion center is BP neural network algorithm, and input layer is respectively
Falling down tentative prediction T, tentative prediction fallen down based on trunk 3-axis acceleration sensor based on hand touch sensor
As a result P and tentative prediction result Z is fallen down based on gyroscope.
Output layer directly exports fusion results, chooses t=1 herein, and corresponding neural network output has the general of risk of falls
Rate p'.BP neural network is trained and is tested to obtain the connection weight and threshold value of network, decision level fusion center is established and falls
Prediction model.
Compared with prior art, the present invention has technique effect beneficial below:
The present invention acquires human hands tactile data, human body 3-axis acceleration information, trunk angle by three kinds of sensors
Information carries out falling down tentative prediction respectively using these three kinematics character parameters during falling over of human body, finally by
Three kinds of decision level fusion center pair information merges.Melted by detecting tactile and trunk movable information and carrying out multi information
It closes, the current motion state of human body can be quickly detected in 300ms, and rate of accuracy reached has to 95% to falling down prediction
Relatively high efficiency and accuracy rate can quickly and accurately judge the elderly's physical condition.
Description of the drawings
Fig. 1 is to fall down prediction technique based on combined of multi-sensor information;
Fig. 2 is to fall down tentative prediction algorithm flow chart based on touch sensor;
Fig. 3 is to carry out falling down process tentative prediction identification process figure using HMM methods;
Fig. 4 is to fall down process tentative prediction identification process figure using angular speed;
Fig. 5 is MULTISENSOR DECISION FUSION SYSTEM grade fusion center structure chart;
Fig. 6 is whole hardware realization block diagram.
Specific implementation mode
With reference to specific embodiment, the present invention is described in further detail, it is described be explanation of the invention and
It is not to limit.
Referring to Fig. 1~Fig. 6, the present invention is a kind of Multi-sensor Fusion prediction technique that the elderly falls down, and flow diagram is such as
Shown in Fig. 1, include the following steps:
Human hands tactile data, people are obtained respectively first with touch sensor, 3-axis acceleration sensor, gyroscope
Three axis of body trunk accelerates information and angle information;
Then by extracting relevant feature during falling over of human body, using falling down prediction algorithm accordingly to selected
Haptic signal feature, three axis resultant acceleration feature of human body and trunk angle character are calculated, and tentatively judge whether human body occurs
It falls down;
Multi information decision level fusion is realized finally by BP neural network algorithm, finally carries out prediction judgement to falling down.
1, falling over of human body course motion feature extraction
Relative motion feature includes during falling down:Resultant acceleration variation, the variation of body posture, hand grip
Variation.
Initially set up body trunk coordinate system Oxyz and earth axes OXYZ.Acceleration becomes metastomium along the x-axis direction
Turn to ax, the acceleration of metastomium along the y-axis direction is ay, the acceleration of metastomium along the z-axis direction is az, thenResultant acceleration is:
With trunk corner indicate trunk direction represented by angle, define corner of the trunk relative to earth axes OXYZ
θ1, θ2, θ3, wherein θ1It is angle of heel around x-axis rotation;θ2It is pitch angle, θ around y-axis rotation3It is spin angle around z-axis rotation.It fell down
θ when being fallen laterally in journey1It changes greatly, front and back fall down is then θ2It varies widely.
2, tentative prediction is fallen down based on hand touch sensor
Based on hand tactile data to fall down prediction algorithm flow chart as shown in Figure 2:
Touch sensor number of active lanes for acquiring human hand tactile data is 8 tunnels, and shape is normally moved in the elderly
Under state, the feature vector R=[a of useful signal number of active lanes N=4 structure haptic signal acquisition systemsi,bi,ci,di,ei,fi,
gi,hi], by the research of the behavior posture to people it has been observed that when user falls down, the both hands of people understand the tight of instinct
It holds, can be more than set at this point, there will be over 4 road touch sensors by human hand pressure, and since pressure value is larger
Useful signal decision threshold FT=0.1.So when feature vector R useful signal number N > 4, user's proper motion mould
The tactile useful signal port number setting threshold value N of feature vector under formulaT=4, acquire touching for touch sensor by the way that DSP is arranged
Feel pressure value, programming realizes the processing to haptic signal and obtains the port number of tactile useful signal, as useful signal port number N
> NTShi Jiyou risk of falls, output fall down tentative prediction T to decision level fusion center.
3, tentative prediction is fallen down based on trunk 3-axis acceleration sensor
The acceleration time series sampling period is TsIt is n with length, that is, includes n element { ci(i=1,2 ..., n),
In each element characterize corresponding period TsThe motion feature of trunk on interior human body.The time of each acceleration time series
Length is Ts×n.Acceleration time series is indicated using sliding time window S, that is, there is sampling period TsIt is n with length.Three
The sampling period T=20ms of axle acceleration sensor, then Ts≥T.If Ts=mT, wherein m are integer, and m >=1.Then TsPeriod
Inside contain m resultant acceleration valueThe process of feature extraction is each section of TsM in period
Resultant acceleration value merges the element for extracting the motion feature that one characterizes the period, finally this n element combinations at one
A characterization TsThe acceleration time series of motion feature in × n.
The acceleration time series time span T of this motion process is describeds× n is also certain.Therefore Ts× n=T × m
× n is also certain.When sensor uses cycle T=20ms, is analyzed by many experiments, take m=2, length of time series n
=11, i.e. t=n × T=440ms, and n × m=22.
Followed by feature extraction:
The first step:Every section of T of characterization is determined with maximum distance methodsThe resultant acceleration a of interior motion state characteristicc.Definition, which is closed, to be added
One reference tape B of speed is B=[b1,b2], wherein b1< G, b2> G.B=[9.5m/s are set2,11.5m/s2].Therefore it presses
Maximum distance method determines TsInterior characteristic acceleration value ac.Define the resultant acceleration value a of t momentiAt a distance from B:
d(ai, B)=| ai-b1|+|ai-b2|;
Then TsInterior m resultant acceleration value aiWith the maximum distance of B:
max[d(ai, B) | i=1,2 ..m];
Then acCreate-rule be:
If:d(aj, B) and=max [d (ai, B)], i, j=1,2 ..m, then:ac=aj。
Second step:By acCodomain carry out break sign, generate the element { c of sliding window SiI=1,2...n.ac's
Codomain θ is defined as:
θ=[θ0,θ1)∪[θ1,θ2)∪···∪[θk-1,θk),θ0=0, θk→+∞;
Symbol finite aggregate is defined as:
S={ s1,s2,···,sk},k≥1;
Therefore it is with S correspondences by each section of codomain θ:
Characteristic acceleration acThe rule of symbolism:I-th of element c of current sliding window mouthi, corresponding acceleration is
(ac)i, then:If (ac)i∈[θj-1,θj), 1≤j≤k, then ci=sj.After the test according to the resultant acceleration during falling down
Variation Features take k1=3, k2=3, the k=8 of then symbol S.
HMM (hidden Markov model) is established:
The parameter for the first-order hidden Markov model established in the present invention is described as follows:
①M:Implicit motion state quantity during falling down.User is the u since preceding safe condition1, experience peace
Transition state u between total state and disequilibrium state2, arrive out-of-balance condition uMThis process, M=3.
②N:Observation quantity, i.e. acceleration time series of the acceleration value in user's motion process after extraction
In element species, N=8.
③π:Initial probability distribution.π={ πi, i=1,2, M, wherein:
Because of safe condition u before initial state is always fallen1Start.Therefore π1=1, πi=0, i=1,2, M.
④A:User fall down during hidden state-transition matrix, A={ aij, i, j=1,2, M, wherein
From motion state uiTo state ujTransition probability.
The characteristics of when choosing initial transition matrix A according to initial model, observes selection matrix A with many experiments:
⑤B:Observing matrix.The relationship between each hidden state of process and observation is fallen down in description.B={ bjk, j=1,
2, M, k=1,2, N, wherein:
User is kept in motion ujWhen, it is v after acceleration value is extractedkProbability.
It is according to selection matrix B the characteristics of initial model and the motion state fallen down when choosing initial observation matrix:
Final HMM model is:
The each sample for the different motion processes that statistical sample is concentrated is subjected to feature extraction and obtains acceleration time sequence
Then row are used as observation sequence O, calculate output probability P (O | λp), the threshold fallen down prediction and differentiated is determined using statistical method
Value.Entire judgement time t < 300ms.Criterion is at this time:As P > P1When, possibility is fallen down in prediction, and is exported preliminary pre-
Result P is surveyed to decision level fusion center.
After taking logarithm to the output probability of acquisition, threshold value, setting are calculated using statistical method:logP1=-33.It uses
HMM methods carry out falling down process identification block diagram such as Fig. 3.
4, tentative prediction is fallen down based on gyroscope
The angular velocity omega rotated around x-axis is obtained by gyroscope1With the angular velocity omega rotated around y-axis2To calculate body trunk
Angle of heel and pitch angle.Sampling period T=20ms, angle of heel and pitch angle calculate as follows:
Wherein, n1=int (t1/ T), n2=int (t2/T)。
People falls down process and the tilt threshold of the trunk deviation Z axis of other proper motion patterns is denoted as θT.Fall down with it is low
The angle of heel θ of 300ms or so before gesture object impactYAnd pitching angle thetaPAt least one is more than θT;And in proper motion mode process
In, angle of heel θYAnd pitching angle thetaPRespectively less than θT。
Prediction threshold value is fallen down in determination:It takes and falls down set of data samples ΩF, people is calculated in each sample during disequilibrium
The inclination angle theta of first impact peak value 300ms of distance, including angle of heel and pitch angle, are recorded as set omegaθ1.Then normal fortune is taken
Dynamic model formula set of data samples ΩO, the inclination angle that sample deviates Z axis in each moment trunk is calculated, set omega is denoted asθ2。
Because of Ωθ1And Ωθ2Element is all inclination data, belongs to single-dimensional data, so feature space YnDimension be n=1.
Then the form of optimal separating hyper plane is a point:
Y=b;
Therefore classifying rules is:If y≤b, y ∈ Ωθ2Class;If y >=b, y ∈ Ωθ1Class.Fall down prediction threshold value θT=b.
If some time engraves the angle of heel θ of trunkYAnd pitching angle thetaPAbsolute value in it is at least one be more than θT, then the movement
Process is judged as falling down tendency, and falls down possibility Z to the input of decision level fusion center.
It is as shown in Figure 4 that trunk angle falls down prediction flow chart.
5, the multi-sensor information decision level fusion based on BP neural network
Sample data is normalized, formula is normalized:
Specific design is applied in the BP neural network for falling down prediction model.
1. input layer:Input layer number i=3.Input layer is respectively to be fallen down result T based on tactile judgement, be based on
Trunk 3-axis acceleration sensor falls down tentative prediction result P and falls down tentative prediction result Z based on gyroscope.
2. output layer:Output layer directly exports fusion results, chooses t=1, and corresponding neural network output has risk of falls
Probability p '.
3. the selection of hidden layer number of nodes:Hidden layer number of nodes is chosen using empirical equation:
Take hidden layer number of nodes j=12.
4. connecting the determination of initial weight:Initial weight takes the random number between [- 1,1].
5. falling down prediction neural network model structure:Predictive information emerging system is fallen down reversely to pass using three layers of feedforward error
Broadcast neural network.Using BP algorithm, target is so that the neural network output y=p' and desired value p in experimentmBetween it is equal
Variance e is minimum, i.e.,:
In formula:P' and pmIt is normalized value, m is sample number, and ε is arbitrarily small positive real number, i.e. given allowable error,
Take ε=10-2。
6. falling down the neural network algorithm of predictive information fusion:LM (Levemberg-Marquardt) learning algorithm.
By being trained to BP neural network and test obtains network connection weights and threshold value, establish in information fusion
The heart falls down prediction model.The decision output for falling down prediction emerging system is divided into two kinds:Safe condition falls down tendency.Nerve
The output valve of network is probability value, indicates the possibility size for falling down tendency.When fall reach 30 ° when output valve
It is set as 0.6, output valve when reaching 60 ° that falls is 0.9.So when output probability value is more than 0.6, you can to judge to make
User falls down tendency;When output probability value is less than 0.5, it is possible to determine that user is in safe handling state.By to training
The experimental analysis of sample chooses output probability value 0.5 as the threshold value differentiated, and when more than 0.5, tendency is fallen down in judgement;It is small
In when judgement in safety state.
Fig. 5 is multi-sensor information fusion centre junction composition.
Fig. 6 is whole hardware realization block diagram, and haptic signal, 3-axis acceleration signal and angular velocity signal are by touch sensor
It acquires and is input in microprocessor in real time with MPU-6050 modules, microprocessor carries out processing calculating, final output to signal
Instruction is fallen down in judgement, starts Shatter-resistant device.
In conclusion it is a kind of based on combined of multi-sensor information the invention reside in providing, applied to falling down for the elderly
Prediction technique.This method by touch sensor, 3-axis acceleration sensor and gyroscope acquisition user's hand tactile data,
Trunk 3-axis acceleration information and angle information.After extracting corresponding characteristic information, by falling down pre- measuring and calculating accordingly
Method is predicted falling down, and is merged by information fusion center information, is judged falling down, special to the physical condition of the elderly
Sign judges in advance.It prevents from falling down and the elderly is damaged, Liberation can human resources.With higher learning value and
Application value.
Claims (9)
1. a kind of the elderly of fusion multisensor falls down prediction technique, which is characterized in that include the following steps:
Step 1:Acquire signal
By touch sensor, 3-axis acceleration sensor and gyroscope, respectively the hand tactile data of corresponding acquisition user,
Human body 3-axis acceleration information and trunk angle information;
Step 2:Extract key feature information
By establishing coordinate system between ground and human body, chooses the elderly hand haptic force, three axis of trunk during falling down and accelerate
Characteristic parameter in degree and three kinds of dynamic-change informations of trunk angle is as the key feature for falling down prediction;
Step 3:Key feature information fusion judges
To the key characterization parameter of the hand haptic force, trunk 3-axis acceleration and trunk angle extracted, use respectively corresponding
Fall down prediction algorithm to user fall down trend carry out tentative prediction, obtain three kinds of tentative prediction as a result, then right again
Three kinds of results of tentative prediction carry out decision level fusion, finally determine whether there is and fall down possibility.
2. the elderly of fusion multisensor according to claim 1 falls down prediction technique, which is characterized in that step 2
In, extraction key feature information concrete operations are:
First, relative motion feature during extraction is fallen down:Hand haptic force, trunk 3-axis acceleration and trunk angle;
Secondly, hand haptic force is acquired by the touch sensor that 8 road PVDF piezoelectric film sensor units form, and establishes feature
Vectorial R=[ai,bi,ci,di,ei,fi,gi,hi], the haptic force numerical value F that element is acquired by each road sensori, i=1,2,
3...8;
Body trunk coordinate system Oxyz and earth axes OXYZ are established, acceleration change is a to metastomium along the x-axis directionx, body
The acceleration of cadre position along the y-axis direction is ay, the acceleration of metastomium along the z-axis direction is az, thenIt closes and adds
Speed is:
Define rotational angle theta of the trunk relative to earth axes OXYZ1, θ2, θ3, wherein θ1It is angle of heel around x-axis rotation;θ2Around y-axis
Rotation is pitch angle, θ3It is spin angle around z-axis rotation;
θ when being fallen laterally during falling down1It changes greatly, front and back is then θ when falling down2It varies widely.
3. the elderly of fusion multisensor according to claim 1 falls down prediction technique, which is characterized in that step 3
In, the hand haptic force extracted is carried out falling down tentative prediction, concrete operations are:
Touch sensor shares 8 tunnels, sets effective signal threshold value FT, work as Fi> FT, input signal is effective;Under proper motion, have
Imitate signal path number N=4;When falling down generation, useful signal number N > 4;Build haptic force feature vector R=[ai,bi,ci,
di,ei,fi,gi,hi], setting threshold value NT=4;
Rule of judgment is:As N > NT, it is determined with and falls down possibility, and exports judging result to decision level fusion center.
4. the elderly of fusion multisensor according to claim 1 falls down prediction technique, which is characterized in that step 3
In, the trunk 3-axis acceleration extracted is carried out falling down tentative prediction, concrete operations are:
Trunk 3-axis acceleration data are extracted as a series of acceleration time series Oi, establish the hidden Ma Erke for falling down prediction
The feature extraction of husband model HMM, motion process obtain acceleration time series, are then used as observation sequence O, calculate output probability
P(O|λp), the threshold value P for falling down prediction and differentiating is determined with statistical method1;
Rule of judgment is:As P > P1When, possibility is fallen down in prediction, and exports judging result to decision level fusion center.
5. the elderly of fusion multisensor according to claim 1 falls down prediction technique, which is characterized in that step 3
In, the trunk angle extracted is carried out falling down tentative prediction, concrete operations are:
Calculate trunk angle of heel θYAnd pitching angle thetaP, human body trunk angle samples during falling down are acquired, are looked for using SVM
Go out optimal planar, is set as trunk tilt threshold θT;
Rule of judgment is:Work as θY> θTOr θP> θT, it is determined with and falls down possibility, and exports judging result to decision level fusion center.
6. the elderly of fusion multisensor according to claim 1 falls down prediction technique, which is characterized in that step 3
In, it is that decision level fusion is carried out to three kinds of results of tentative prediction using BP neural network blending algorithm, i.e., by tentative prediction
Three kinds of results are input to BP neural network, carry out decision level fusion, to realize accurate prediction that the elderly falls down.
7. a kind of the elderly for the fusion multisensor realized described in any one of claim 1~6 falls down prediction technique
System, which is characterized in that including:
Signal collecting device, for acquiring hand haptic signal, human body 3-axis acceleration signal and trunk angle signal;
Microprocessor is electrically connected with signal collecting device, carries out processing calculating for the signal to acquisition, and export judgement and fall down
Instruction;
Drive module, the order-driven Shatter-resistant device work sent out for receiving microprocessor;
Shatter-resistant device is worn on user's trunk position.
8. system according to claim 7, which is characterized in that acquisition user's hand tactile data touch sensor by
The array of several PVDF piezoelectric film sensors unit compositions.
9. system according to claim 7, which is characterized in that acquisition human body 3-axis acceleration information and trunk angle
The sensor of information is MPU-6050.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810080170.3A CN108320456A (en) | 2018-01-27 | 2018-01-27 | It is a kind of fusion multisensor the elderly fall down prediction technique and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810080170.3A CN108320456A (en) | 2018-01-27 | 2018-01-27 | It is a kind of fusion multisensor the elderly fall down prediction technique and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108320456A true CN108320456A (en) | 2018-07-24 |
Family
ID=62888037
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810080170.3A Pending CN108320456A (en) | 2018-01-27 | 2018-01-27 | It is a kind of fusion multisensor the elderly fall down prediction technique and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108320456A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109249429A (en) * | 2018-09-25 | 2019-01-22 | 安徽果力智能科技有限公司 | A kind of biped robot's classification of landform system |
CN109259989A (en) * | 2018-11-07 | 2019-01-25 | 西安交通大学 | One kind is helped the elderly with pedestrian's hand vibration comfort regulating device and its adjusting method |
CN109685111A (en) * | 2018-11-26 | 2019-04-26 | 深圳先进技术研究院 | Action identification method, computing system, smart machine and storage medium |
CN110244624A (en) * | 2019-06-20 | 2019-09-17 | 南京机电职业技术学院 | A kind of tumble protective device and control method |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102125495A (en) * | 2010-12-21 | 2011-07-20 | 西安交通大学苏州研究院 | Multifunctional robot with slippage sensor for helping aged people to walk |
EP2811948A1 (en) * | 2012-02-07 | 2014-12-17 | 3D Systems, Inc. | Brace with elongated fenestrations |
WO2015148578A3 (en) * | 2014-03-24 | 2015-11-26 | Alghazi Ahmad Alsayed M | Multi-functional smart mobility aid devices and methods of use |
CN105469544A (en) * | 2015-11-19 | 2016-04-06 | 常州大学 | Hand-myoelectricity-based multi-functional walking stick |
CN105468145A (en) * | 2015-11-18 | 2016-04-06 | 北京航空航天大学 | Robot man-machine interaction method and device based on gesture and voice recognition |
CN105469546A (en) * | 2016-01-14 | 2016-04-06 | 上海大学 | Tumbling alarm system and method |
CN106227999A (en) * | 2016-07-16 | 2016-12-14 | 北京工业大学 | A kind of high-adaptability is fallen behavioral value method |
CN106360900A (en) * | 2016-09-29 | 2017-02-01 | 武汉康益生电子商务有限公司 | Safety crutch for elderly |
-
2018
- 2018-01-27 CN CN201810080170.3A patent/CN108320456A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102125495A (en) * | 2010-12-21 | 2011-07-20 | 西安交通大学苏州研究院 | Multifunctional robot with slippage sensor for helping aged people to walk |
EP2811948A1 (en) * | 2012-02-07 | 2014-12-17 | 3D Systems, Inc. | Brace with elongated fenestrations |
WO2015148578A3 (en) * | 2014-03-24 | 2015-11-26 | Alghazi Ahmad Alsayed M | Multi-functional smart mobility aid devices and methods of use |
CN105468145A (en) * | 2015-11-18 | 2016-04-06 | 北京航空航天大学 | Robot man-machine interaction method and device based on gesture and voice recognition |
CN105469544A (en) * | 2015-11-19 | 2016-04-06 | 常州大学 | Hand-myoelectricity-based multi-functional walking stick |
CN105469546A (en) * | 2016-01-14 | 2016-04-06 | 上海大学 | Tumbling alarm system and method |
CN106227999A (en) * | 2016-07-16 | 2016-12-14 | 北京工业大学 | A kind of high-adaptability is fallen behavioral value method |
CN106360900A (en) * | 2016-09-29 | 2017-02-01 | 武汉康益生电子商务有限公司 | Safety crutch for elderly |
Non-Patent Citations (2)
Title |
---|
佟丽娜: ""基于力量学信息获取系统的人体摔倒过程识别方法研究"", 《中国博士学位论文全文数据库信息科技辑》 * |
刘树栋: ""基于多传感器融合的摔倒监测研究与设计"", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109249429A (en) * | 2018-09-25 | 2019-01-22 | 安徽果力智能科技有限公司 | A kind of biped robot's classification of landform system |
CN109259989A (en) * | 2018-11-07 | 2019-01-25 | 西安交通大学 | One kind is helped the elderly with pedestrian's hand vibration comfort regulating device and its adjusting method |
CN109259989B (en) * | 2018-11-07 | 2019-10-11 | 西安交通大学 | One kind is helped the elderly with pedestrian's hand vibration comfort regulating device and its adjusting method |
CN109685111A (en) * | 2018-11-26 | 2019-04-26 | 深圳先进技术研究院 | Action identification method, computing system, smart machine and storage medium |
CN109685111B (en) * | 2018-11-26 | 2023-04-07 | 深圳先进技术研究院 | Motion recognition method, computing system, intelligent device and storage medium |
CN110244624A (en) * | 2019-06-20 | 2019-09-17 | 南京机电职业技术学院 | A kind of tumble protective device and control method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108320456A (en) | It is a kind of fusion multisensor the elderly fall down prediction technique and system | |
Yuan et al. | A novel GRU-RNN network model for dynamic path planning of mobile robot | |
CN108828944B (en) | Encoder fault diagnosis system and method based on improved PSO and SVM | |
CN107016342A (en) | A kind of action identification method and system | |
CN102945554B (en) | Target tracking method based on learning and speeded-up robust features (SURFs) | |
AU2020103811A4 (en) | Method for recognizing human knee motion postures based on extreme learning machine | |
CN107657277A (en) | A kind of human body unusual checking based on big data and decision method and system | |
CN109325476A (en) | A kind of human body exception attitude detection system and method based on 3D vision | |
CN106725495A (en) | A kind of fall detection method, apparatus and system | |
CN113663312A (en) | Micro-inertia-based non-apparatus body-building action quality evaluation method | |
CN107169512A (en) | The construction method of HMM SVM tumble models and the fall detection method based on the model | |
Wu et al. | Gait phase prediction for lower limb exoskeleton robots | |
CN104269025A (en) | Wearable type single-node feature and position selecting method for monitoring outdoor tumble | |
CN110464315A (en) | It is a kind of merge multisensor the elderly fall down prediction technique and device | |
CN111382699A (en) | Dynamic gesture recognition method based on particle swarm optimization LSTM algorithm | |
Gao et al. | Logic-enhanced adaptive network-based fuzzy classifier for fall recognition in rehabilitation | |
Wang et al. | A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu | |
CN105869354A (en) | Method for old man falling-down detection based on attractor propagation algorithm | |
Shi et al. | Towards HMM based human motion recognition using MEMS inertial sensors | |
Chen et al. | A comparative study of LSTM and phased LSTM for gait prediction | |
CN106227999B (en) | A kind of high-adaptability tumble behavioral value method | |
CN110889330A (en) | BP neural network-based old people tumbling detection method and system | |
Cheng et al. | A fall detection algorithm based on pattern recognition and human posture analysis | |
Wang et al. | A new fall detection method based on fuzzy reasoning for an omni-directional walking training robot | |
Liu et al. | Bidirectional LSTM-based network for fall prediction in a humanoid |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180724 |