CN106910314B - A kind of personalized fall detection method based on the bodily form - Google Patents
A kind of personalized fall detection method based on the bodily form Download PDFInfo
- Publication number
- CN106910314B CN106910314B CN201710063265.XA CN201710063265A CN106910314B CN 106910314 B CN106910314 B CN 106910314B CN 201710063265 A CN201710063265 A CN 201710063265A CN 106910314 B CN106910314 B CN 106910314B
- Authority
- CN
- China
- Prior art keywords
- user
- data
- stage
- fall detection
- vsa
- 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)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The purpose of the present invention is to overcome the deficiency in the prior art, and design personalized and really practical high-precision fall detection method propose use the feature selection approach based on the bodily form thus.A kind of personalized fall detection method based on the bodily form, it is divided into four-stage, the number of first stage record experiment user, the essential informations such as height and weight, and the acceleration transducer and gyroscope embedded using smart phone records its X under required movement, Y, the acceleration and angular speed information in three directions of Z axis calculate 13 common time domain features according to these data informations after carrying out data prediction.Second stage is gathered user for 3-5 class using k means clustering method according to height and body-mass index (body-mass index, that is, weight kilogram counts square divided by height rice number) information of user.Deng.
Description
Technical field
The present invention relates to human body fall detection technical fields.
Background technique
With the acceleration evolution of social senilization, the physically and mentally healthy problem of the elderly obtains people and more pays close attention to.It falls
It is one of the key factor for causing the elderly's disability even dead, much because dead caused by falling and indirect because of falling this
Body, but because caused by not given treatment to timely after falling.Therefore fall detection system is protected as the safety of the elderly
Barrier, higher detection accuracy are the emphasis that we study.
Different fall detection technologies is had proposed and had studied at present, but is not widely used.Main cause is
These fall detectors of all users use unified disaggregated model, do not account for the movement difference between user, cause to examine
It is lower to survey precision.Therefore, in order to solve this problem, the relevant technologies propose personalized fall detection system-
Chameleon.By combining the threshold strategies based on group with individual adjusting thresholds strategy, the threshold value based on weight is established
Model is extracted, solves low precision problem caused by fixed threshold.However, the system only considers gender and the age of user, and
It does not account for realizing more accurate fall detector using the body profile parameter of different users.In addition, adaptive technique is only fitted
For early stage.It does not account for after user determined threshold value, some variations occur for the behavior of user, and threshold value cannot be again
The case where change.Thus while the algorithm based on threshold value is relatively easy, but it can not fully achieve the tumble of property one by one
Detection system.The method based on threshold value is removed in fall detection technology, there are also the methods of pattern-recognition, it is relative to threshold test
Method is more accurate.Although having the personalized fall detection method based on pattern-recognition at present, however, what this technology used
It is video mode, it limits the freedom of user, is very easy to leakage privacy of user, and cannot use in the dark.Therefore,
Given this disadvantage, wearable fall detection mode are more big well-established.The technology on wearable device by being embedded in acceleration
Degree meter, the sensors such as gyroscope realize the real time monitoring to monitoring individual.In the pattern-recognition fall detection based on smart phone
In technology, it will usually select optimal characteristic of division using feature selection approach to improve characteristic of division.But it is only capable of into
The tumble event of function identification about 90%.
Based on this, the present invention is filtered out during different body parameter groups fall using feature selecting algorithm and is changed significantly
Character subset, using this feature trained group disaggregated model, realize based on bodily form feature selecting personalized falls
Detection method.
Summary of the invention
The purpose of the present invention is to overcome the deficiency in the prior art, design personalized and really practical high-precision fall detection
Method proposes use the feature selection approach based on the bodily form thus.This is because we have found for different users he
Optimal detection feature be different.But the computation complexity of individualized feature is selected but to make this for each specific user
Method is difficult to realize.
For this purpose, the technical solution provided are as follows:
A kind of personalized fall detection method based on the bodily form, which is characterized in that it is divided into four-stage,
First stage records the number of experiment user, the essential informations such as height and weight, and embedded using smart phone
Acceleration transducer and gyroscope record its X under required movement, Y, the acceleration and angular speed information in three directions of Z axis,
It carries out data prediction and calculates 13 common time domain features according to these data informations later.
(body-mass index, that is, weight kilogram number is divided by height according to the height and body-mass index of user for second stage
Square of rice number) information, user is gathered for 3-5 class using k means clustering method.
Phase III combines the 13 most common time domain features for each user being calculated in the first stage
Cluster situation with the user in second stage, using packaged type feature selecting algorithm calculate the optimal motion characteristic of every class user to
Amount.First three above-mentioned stage is the data prediction of experiment user, predominantly obtains the optimal motion characteristic vector of every class user.
Fourth stage is the scene of practical application.The user according to obtained in second stage classifies situation and in the phase III
The optimal motion characteristic vector of obtained every class user, using the method for support vector machines combine body inclination angle based on threshold value and
The duration judgement of improper body posture carries out fall detection again, continues to optimize classification finally by on-line learning algorithm
Parameter realizes high-precision personalized fall detection algorithm.
The first stage specific implementation step:
The number of user 1-1) is recorded, the essential informations such as height and weight will be equipped with acceleration transducer and top later
The smart phone of the data acquisition program of spiral shell instrument is placed in the waist pocket of user, and direction is random.Pass through 6 normal behaviours
(walking is jogged, and jump goes upstairs, goes downstairs, sits down) (forward direction is fallen, backward to fall, laterally for the tumble behavior different with 3 kinds
Fall) carry out data acquisition.Because human motion frequency is generally below 20Hz, therefore according to Nyquist Sampling Theorem, data are acquired
Frequency is 50Hz.The number information of user is wherein indicated using N, H indicates the height data of user, and W indicates the weight letter of user
Breath, ax, ay, azIndicate the X of collected acceleration transducer, the acceleration information of tri- axis of Y, Z, px, py, pzExpression collects
Gyroscope X, the angular velocity data of tri- axis of Y, Z.
1-2) using Kalman filtering algorithm to 1-1) in ax, ay, azAnd px, py, pzSix groups of data are pre-processed, and are disappeared
Influence except noise to acceleration transducer and gyroscope measured value, to improve the accuracy of data.
1-3) based on the acceleration and angular velocity data after 1-2) filtering processing, the resultant acceleration of every user is calculated
ValueWith close angle velocity amplitudeAnd extract 13 correlated characteristic F
={ f1,f2,…,fi,…fn(n=13), specific features are described as follows:
f1: aveVSA, i.e. the mean value of VSA.
f2: aveVSP, i.e. the mean value of VSP.
f3: Δ VSA, i.e. VSA maximin absolute value of the difference.
f4: Δ VSP, i.e. VSP maximin absolute value of the difference.
f5: Δ tVSA, i.e. time difference corresponding to VSA maximin.
f6: Δ tVSP, i.e. time difference corresponding to VSP maximin.
f7: stdVSA, i.e. the standard deviation of VSA.
f8: stdVSP, i.e. the standard deviation of VSP.
f9: EVSA, i.e., movable energy value, wherein
f10: S, i.e. slope, wherein
f11: Δ φ, i.e. body minimax inclination angle difference, wherein φ (t) indicates t moment body and ground angle,
f12: Δ pitch, the minimax difference of pitch angle, wherein
f13: Δ roll, the minimax difference of roll angle, wherein
The second stage specific implementation step:
2-1) as 1-1) obtained in user's height data H and weight data W calculate user's body performance figure BMI=H/
W2, and the physiological characteristic vector of user is expressed as FB=(H, BMI).
2-2) based on the user's physiological characteristic vector F being calculated in 2-1)B, user is gathered using K mean cluster algorithm
For 3~5 classes.
The phase III specific implementation step:
3-1) to 1-3) in motion characteristic set F={ f1,f2,…,fi,…fn(n=13), owned by traversal
Motion characteristic proper subclass set G={ F1,F2,…,Fj,…Fm(m=8191).
3-2) randomly select sample for every group of user 80 percent passes through 3-1 respectively) in motion characteristic proper subclass
Feature in set G uses support vector machines train classification models, is tested using remaining 20 percent sample.Make
The method verified with right-angled intersection, seeks each feature vector FjThe disaggregated model trained is attainable for every group of user institute
Precision.Finally choose the optimal motion characteristic vector F that can make to reach the highest feature vector of precision as the groupA.If in case of
Dry feature vector reaches precision highest simultaneously, then feature quantity is least as optimal characteristics vector F in selected characteristic vectorA。
The fourth stage specific implementation step:
4-1) in practice, affiliated group is matched according to specific user's physiological characteristic, affiliated group pair is extracted to it
The optimal characteristics vector answered, and disaggregated model is constructed using support vector machines, carry out first fall detection judgement.
4-2) in order to further improve the performance of fall detection system, will add two features here: most lifelong
Then the duration at body inclination angle and improper body posture judge again using thresholding algorithm to reduce False Rate.Choosing
The reason of selecting the two features is that the value size of their threshold values is unrelated with the body parameter of user.Only when this improper appearance
The tilt angle of gesture and duration are just finally judged as falling when being more than the threshold value of our settings.
By on-line study machine after characteristic of division building fall detection model 4-3) selected according to user's group
The daily behavior of system study specific user and judge that data constantly adjusts and Optimum Classification parameter, really realization personalization with
Adaptive fall detection.
The daily behavior of on-line study mechanism study specific user and judge that data constantly adjust and Optimum Classification
Parameter really realizes personalized and adaptive fall detection, and algorithm is as follows:
Input: user's height H and weight W
Output: tumble behavior occurs for user
(1) the specific body parameter inputted according to user, including height H and weight W, the body quality for calculating user refer to
Number BMI, and the affiliated group of the user is matched with BMI according to H, it executes step (2).
(2) X, the Y of user, the acceleration information and angular velocity data of Z axis are acquired in real time, and extracting according to acquisition data should
The optimal motion characteristic vector value of group corresponding to user executes step (3).
(3) fall detection detection judgement is carried out using the method for support vector machines (SVM), if judging result is normal behaviour
It thens follow the steps (5), it is no to then follow the steps (4).
(4) judging result is further detected for the case where improper behavior, passes through and extracts the most lifelong of user
The duration at body inclination angle and improper body posture judge again using thresholding algorithm to reduce False Rate.Judge non-
Whether the tilt angle of normal posture and duration are more than threshold value that we are arranged, if so then execute step (5), and export use
Tumble behavior occurs for family, no to then follow the steps (5).
(5) learn daily behavior and the tumble behavior of specific user, continuous adjustment and optimization by on-line study mechanism
Sorting parameter executes step (2).
The present invention is grouped according to the similitude of the optimal characteristics of all experiment users first, and calculates each group most
Excellent feature vector.Experiment shows that the detection accuracy of this method significantly improves.In addition, we experimentally found that based on user's
The similitude classification of the bodily form (i.e. height and body-mass index) can produce and be based on personal optimal characteristics similitude classification side
The almost the same result of method.Therefore, we indicated that the feature selecting based on the bodily form can effectively improve detection accuracy.In reality
In, by according to user's bodily form extract it belonging to the corresponding optimal characteristics vector of grouping, use support vector machines and threshold value
Judge that the mode combined carries out fall detection judgement, further improves detection accuracy in this way.It is learned secondly by online
Habit mode constantly adjusts and Optimum Classification parameter, really realizes personalized and adaptive fall detection.
Detailed description of the invention
A kind of personalized fall detection method module map based on the bodily form of Fig. 1
1 flow chart of Fig. 2 algorithm
Specific embodiment
(case)
Entire scheme is divided into four-stage, as shown in Figure 1: the number of first stage record experiment user, height and weight
Etc. essential informations, and using smart phone embed acceleration transducer and gyroscope record its X, Y, Z axis under required movement
The acceleration and angular speed information in three directions calculates 13 often according to these data informations after carrying out data prediction
With time domain characteristic of division.Second stage is according to height and the body-mass index (body-mass index, that is, weight kilogram number of user
Divided by square of height rice number) information, user is gathered for 3-5 class using k means clustering method.Phase III combines the first stage
In user in the 13 most common time domain features and second stage of each user that are calculated cluster situation, make
The optimal motion characteristic vector of every class user is calculated with packaged type feature selecting algorithm.First three stage is that the data of experiment user are pre-
Processing, predominantly obtains the optimal motion characteristic vector of every class user, and fourth stage is the scene of practical application.According to second-order
User obtained in section classifies the optimal motion characteristic vector of every class user obtained in situation and phase III, using support to
The method of amount machine combines the duration judgement at body inclination angle and improper body posture based on threshold value to carry out inspection of falling again
It surveys, continues to optimize sorting parameter finally by on-line learning algorithm and realize high-precision personalized fall detection algorithm.
1, first stage specific implementation step:
The number of user 1-1) is recorded, the essential informations such as height and weight will be equipped with acceleration transducer and top later
The smart phone of the data acquisition program of spiral shell instrument is placed in the waist pocket of user, and direction is random.Pass through 6 normal behaviours
(walking is jogged, and jump goes upstairs, goes downstairs, sits down) (forward direction is fallen, backward to fall, laterally for the tumble behavior different with 3 kinds
Fall) carry out data acquisition.Because human motion frequency is generally below 20Hz, therefore according to Nyquist Sampling Theorem, data are acquired
Frequency is 50Hz.The number information of user is wherein indicated using N, H indicates the height data of user, and W indicates the weight letter of user
Breath, ax, ay, azIndicate the X of collected acceleration transducer, the acceleration information of tri- axis of Y, Z, px, py, pzExpression collects
Gyroscope X, the angular velocity data of tri- axis of Y, Z.
1-2) using Kalman filtering algorithm to 1-1) in ax, ay, azAnd px, py, pzSix groups of data are pre-processed, and are disappeared
Influence except noise to acceleration transducer and gyroscope measured value, to improve the accuracy of data.
1-3) based on the acceleration and angular velocity data after 1-2) filtering processing, the resultant acceleration of every user is calculated
ValueWith close angle velocity amplitudeAnd extract 13 correlated characteristic F
={ f1,f2,…,fi,…fn(n=13), specific features are described as follows:
f1: aveVSA, i.e. the mean value of VSA.
f2: aveVSP, i.e. the mean value of VSP.
f3: Δ VSA, i.e. VSA maximin absolute value of the difference.
f4: Δ VSP, i.e. VSP maximin absolute value of the difference.
f5: Δ tVSA, i.e. time difference corresponding to VSA maximin.
f6: Δ tVSP, i.e. time difference corresponding to VSP maximin.
f7: stdVSA, i.e. the standard deviation of VSA.
f8: stdVSP, i.e. the standard deviation of VSP.
f9: EVSA, i.e., movable energy value, wherein
f10: S, i.e. slope, wherein
f11: Δ φ, i.e. body minimax inclination angle difference, wherein φ (t) indicates t moment body and ground angle,
f12: Δ pitch, the minimax difference of pitch angle, wherein
f13: Δ roll, the minimax difference of roll angle, wherein
2, second stage specific implementation step:
2-1) as 1-1) obtained in user's height data H and weight data W calculate user's body performance figure BMI=H/
W2, and the physiological characteristic vector of user is expressed as FB=(H, BMI).
2-2) based on the user's physiological characteristic vector F being calculated in 2-1)B, user is gathered using K mean cluster algorithm
For 3~5 classes.
2-3) in 2-2) in use FBIt is described as follows as the reason of cluster foundation: the current research in relation to fall detection
In, influence of the different motion characteristics for final accuracy rate does not obtain abundant consideration, we have chosen quotes in recent years
13 class motion characteristic used in relatively high research is measured, based on the data acquired in 1-1), is calculated for every user
Arrived 1-3) shown in motion characteristic, and be found through experiments that for every user all there is one group of specific accuracy rate most
High motion characteristic, this group of motion characteristic is defined as the optimal motion characteristic vector F of user by weA.Our experiment is simultaneously
It demonstrates in fall monitoring system, if we can calculate an optimal motion characteristic vector to every user,
Detection accuracy is available to be obviously improved.But in practical applications, for every user train an optimal motion characteristic to
A series of problems, such as amount is faced with cold start-up, data acquisition and computing resource is limited, therefore, if certain generic features can be passed through
Classify to user, and calculate its distinctive optimal motion characteristic vector for every a kind of user, is to solve actual application problem
An approach.In order to find this generic features, we are firstly the need of providing the classification being of practical significance for fall detection.
So on the basis of previous experiments, it is believed that the similarity of the optimal characteristics vector between two users is higher, they are just
More it is possible that in one kind, experiment user is divided into 3-5 class, each group of optimal motion characteristic vector eventually by the similarity
gFAEqually obtained via experimental calculation.It is demonstrated experimentally that by the rule classification, final accuracy in detection be every user
Calculate FAObtained accuracy no significant difference.But this classifying rules is not suitable for practical application.By to every class
The observation of user's various features (including gender, age, height, weight etc.), it has been found that height H and body in each group group
Performance figure BMI difference is smaller, and group difference is larger.Our physiological characteristic vectors based on user as a result, use K mean value
User is equally polymerized to n class by cluster, it is found that the cluster result coincide substantially with the result according to the grouping of optimal motion characteristic vector,
Therefore in practical applications, we can be according to physiological characteristic vector FBClassify to user, and the optimal movement for calculating every class is special
Vector is levied, when a new user arrives, only need to can correspond to its optimal movement according to the physiological characteristic vector of its own
Feature vector.
2-4) in 2-3) in user gathered be described as follows for the reason of 3 to 5 class: according to above-mentioned analysis it is found that based on user
In the classification of optimal motion characteristic vector, granularity of classification is thinner, then every group of accessible precision reality that is higher, but passing through us
It issues after examination and approval now, if granularity of classification is too thin, the diversity factor of the physiological characteristic vector in every class and between all kinds of is not obvious enough, it is difficult to use
Physiological characteristic vector alternative acts feature vector is classified.We are tested by multiple groups and are found, 3 to 5 classes are most suitably adapted for reality
The classification of border application.
3, phase III specific implementation step:
3-1) to 1-3) in motion characteristic set F={ f1,f2,…,fi,…fn(n=13), owned by traversal
Motion characteristic proper subclass set G={ F1,F2,…,Fj,…Fm(m=8191).
3-2) randomly select sample for every group of user 80 percent passes through 3-1 respectively) in motion characteristic proper subclass
Feature in set G uses support vector machines train classification models, is tested using remaining 20 percent sample.Make
The method verified with right-angled intersection, seeks each feature vector FjThe disaggregated model trained is attainable for every group of user institute
Precision.Finally choose the optimal motion characteristic vector F that can make to reach the highest feature vector of precision as the groupA.If in case of
Dry feature vector reaches precision highest simultaneously, then feature quantity is least as optimal characteristics vector F in selected characteristic vectorA。
4, fourth stage specific implementation step:
4-1) in practice, affiliated group is matched according to specific user's physiological characteristic, affiliated group pair is extracted to it
The optimal characteristics vector answered, and disaggregated model is constructed using support vector machines, carry out first fall detection judgement.
4-2) in order to further improve the performance of fall detection system, will add two features here: most lifelong
Then the duration at body inclination angle and improper body posture judge again using thresholding algorithm to reduce False Rate.Choosing
The reason of selecting the two features is that the value size of their threshold values is unrelated with the body parameter of user.Only when this improper appearance
The tilt angle of gesture and duration are just finally judged as falling when being more than the threshold value of our settings.
By on-line study machine after characteristic of division building fall detection model 4-3) selected according to user's group
The daily behavior of system study specific user and judge that data constantly adjusts and Optimum Classification parameter, really realization personalization with
Adaptive fall detection (being specifically shown in algorithm 1).
Personalized fall detection algorithm of the algorithm 1 based on the bodily form (detailed process is shown in Fig. 2)
Input: user's height H and weight W
Output: tumble behavior occurs for user
(1) the specific body parameter inputted according to user, including height H and weight W, the body quality for calculating user refer to
Number BMI, and the affiliated group of the user is matched with BMI according to H, it executes step (2).
(2) X, the Y of user, the acceleration information and angular velocity data of Z axis are acquired in real time, and extracting according to acquisition data should
The optimal motion characteristic vector value of group corresponding to user executes step (3).
(3) fall detection detection judgement is carried out using the method for support vector machines (SVM), if judging result is normal behaviour
It thens follow the steps (5), it is no to then follow the steps (4).
(4) judging result is further detected for the case where improper behavior, passes through and extracts the most lifelong of user
The duration at body inclination angle and improper body posture judge again using thresholding algorithm to reduce False Rate.Judge non-
Whether the tilt angle of normal posture and duration are more than threshold value that we are arranged, if so then execute step (5), and export use
Tumble behavior occurs for family, no to then follow the steps (5).
(5) learn daily behavior and the tumble behavior of specific user, continuous adjustment and optimization by on-line study mechanism
Sorting parameter executes step (2).
Innovative point of the present invention
1. this project find and demonstrate user optimal detection feature it is related to its bodily form feature (height and weight)
Property.
2. this project proposes the personalized fall detection method based on the bodily form, filtered out using feature selecting algorithm
The character subset that different body parameter groups are changed significantly during falling, using this feature trained group disaggregated model,
Realize the personalized fall detection method based on bodily form feature selecting.
Annotation: the present invention in related term and following data can be found in for previous major technique.
[1]Ren L,Shi W.Chameleon:personalised and adaptive fall detection of
elderly people in home-based environments[J].International Journal of Sensor
Networks,2016,20(3):163-176.
[2]Yu M,Yu Y,Rhuma A,et al.An online one class support vector
machine-based person-specific fall detection system for monitoring an elderly
individual in a room environment[J].IEEE journal of biomedical and health
informatics,2013,17(6):1002-1014.
[3]Baek W S,Kim D M,Bashir F,et al.Real life applicable fall
detection system based on wireless body area network[C]//2013 IEEE 10th
Consumer Communications and Networking Conference(CCNC).IEEE,2013:62-67.
[4]Pierleoni P,Belli A,Palma L,et al.A high reliability wearable
device for elderly fall detection[J].IEEE Sensors Journal,2015,15(8):4544-
4553.
[5]Shan S,Yuan T.A wearable pre-impact fall detector using feature
selection and support vector machine[C]//IEEE 10th INTERNATIONAL CONFERENCE
ON SIGNAL PROCESSING PROCEEDINGS.IEEE,2010:1686-1689.
[6]Liang S,Ning Y,Li H,et al.Feature selection and predictors of
falls with foot force sensors using KNN-based algorithms[J].Sensors,2015,15
(11):29393-29407.
[7]Kansiz A O,Guvensan M A,Turkmen H I.Selection of time-domain
features for fall detection based on supervised learning[C]//Proceedings of
the World Congress on Engineering and Computer Science,San Francisco,CA,
USA.2013:23-25.
Claims (5)
1. a kind of personalized fall detection method based on the bodily form, which is characterized in that it is divided into four-stage,
First stage records number, height and the weight essential information of experiment user, and the acceleration embedded using smart phone
Sensor and gyroscope record its X under required movement, Y, and the acceleration and angular speed information in three directions of Z axis is being counted
13 common time domain features are calculated according to these data informations after Data preprocess；
For second stage according to the height and body-mass index information of user, being gathered user using k means clustering method is 3~5
Class；
Phase III combines the 13 most common time domain features of each user being calculated in the first stage and the
User in two-stage clusters situation, and the optimal motion characteristic vector of every class user is calculated using packaged type feature selecting algorithm；
First three above-mentioned stage is the data prediction of experiment user, predominantly obtains the optimal motion characteristic vector of every class user；
Fourth stage is the scene of practical application: the user according to obtained in second stage classifies situation and to be obtained in the phase III
Every class user optimal motion characteristic vector, using the method for support vector machines combine body inclination angle based on threshold value and it is non-just
The duration judgement of normal body posture carries out fall detection again, continues to optimize sorting parameter finally by on-line learning algorithm
Realize high-precision personalized fall detection algorithm；
The first stage specific implementation step:
Number, height and the weight essential information of user 1-1) are recorded, acceleration transducer and gyroscope will be installed later
The smart phone of data acquisition program is placed in the waist pocket of user, and direction is random；Not by 6 normal behaviours and 3 kinds
Same tumble behavior carries out data acquisition, and 6 normal behaviours are walking, jog, jump, going upstairs, going downstairs, sitting down,
Described 3 kinds different tumble behaviors are to falling, falling backward, laterally tumble for before, and acquisition data frequency is 50Hz, wherein using
N indicates the number information of user, and H indicates the height data of user, and W indicates the weight information of user, ax, ay, azExpression collects
Acceleration transducer X, the acceleration information of tri- axis of Y, Z, px, py, pzIndicate the X of collected gyroscope, Y, Z tri-
The angular velocity data of axis；
1-2) using Kalman filtering algorithm to 1-1) in ax, ay, azAnd px, py, pzSix groups of data are pre-processed, and elimination is made an uproar
Influence of the sound to acceleration transducer and gyroscope measured value, to improve the accuracy of data；
1-3) be based on 1-2) filtering processing after acceleration and angular velocity data, be calculated every user resultant acceleration value and
Close angle velocity amplitude, resultant acceleration valueClose angle velocity amplitudeAnd extract ten
Three correlated characteristic F={ f1,f2,…,fi,…fn, n=13 is 13 common time domain features, specific features explanation
It is as follows:
f1: aveVSA, i.e. the mean value of VSA；
f2: aveVSP, i.e. the mean value of VSP；
f3: Δ VSA, i.e. VSA maximin absolute value of the difference；
f4: Δ VSP, i.e. VSP maximin absolute value of the difference；
f5: Δ tVSA, i.e. time difference corresponding to VSA maximin；
f6: Δ tVSP, i.e. time difference corresponding to VSP maximin；
f7: stdVSA, i.e. the standard deviation of VSA；
f8: stdVSP, i.e. the standard deviation of VSP；
f9: EVSA, i.e., movable energy value, wherein
f10: S, i.e. slope, wherein
f11: Δ φ, i.e. body minimax inclination angle difference, wherein φ (t) indicates t moment body and ground angle,
f12: Δ pitch, the minimax difference of pitch angle, wherein
f13: Δ roll, the minimax difference of roll angle, wherein
2. the method as described in claim 1, which is characterized in that the second stage specific implementation step:
2-1) as 1-1) obtained in user's height data H and weight data W calculate user's body performance figure BMI=H/W2, and
The physiological characteristic vector of user is expressed as FB=(H, BMI)；
2-2) based on the user's physiological characteristic vector F being calculated in 2-1)B, being gathered user using K mean cluster algorithm is 3~5
Class.
3. the method as described in claim 1, which is characterized in that the phase III specific implementation step:
3-1) to 1-3) in motion characteristic set F={ f1,f2,…,fi,…fn, n=13 obtains all movements by traversal
Feature proper subclass set G={ F1,F2,…,Fj,…Fm, m=8191；
3-2) randomly select sample for every group of user 80 percent passes through 3-1 respectively) in motion characteristic proper subclass set G
In feature use support vector machines train classification models, tested using remaining 20 percent sample；Use ten
The method of word cross validation seeks each feature vector FjThe disaggregated model trained is for every group of attainable essence of user institute
Degree；Finally choose the optimal motion characteristic vector F that can make to reach the highest feature vector of precision as the groupA；In case of several
Feature vector reaches precision highest simultaneously, then feature quantity is least as optimal characteristics vector F in selected characteristic vectorA。
4. the method as described in claim 1, which is characterized in that the fourth stage specific implementation step:
4-1) in practice, affiliated group is matched according to specific user's physiological characteristic, it is corresponding that affiliated group is extracted to it
Optimal characteristics vector, and disaggregated model is constructed using support vector machines, carry out first fall detection judgement；
4-2) in order to further improve the performance of fall detection system, two features are added: final body inclination angle and improper
Then the duration of body posture judge again using thresholding algorithm to reduce False Rate；Select the two features
The reason is that the value size of their threshold values is unrelated with the body parameter of user；Only when the tilt angle of this improper posture and
Duration is just finally judged as falling when being more than the threshold value of setting；
By on-line study mechanism after characteristic of division building fall detection model 4-3) selected according to user's group
It practises the daily behavior of specific user and judges that data constantly adjust and Optimum Classification parameter, real realization are personalized and adaptive
The fall detection answered.
5. method as claimed in claim 4, which is characterized in that the daily behavior of the on-line study mechanism study specific user
It is constantly adjusted and Optimum Classification parameter with data are judged, really realizes personalized and adaptive fall detection, algorithm is such as
Under:
Input: user's height H and weight W
Output: tumble behavior occurs for user
(1) the specific body parameter inputted according to user, including height H and weight W, calculate the body-mass index of user
BMI, and the affiliated group of the user is matched with BMI according to H, it executes step (2)；
(2) X, the Y of user, the acceleration information and angular velocity data of Z axis are acquired in real time, and extract the user according to acquisition data
The optimal motion characteristic vector value of corresponding group executes step (3)；
(3) fall detection detection judgement is carried out using the method for support vector machines, executes step if judging result is normal behaviour
Suddenly (5), it is no to then follow the steps (4)；
(4) judging result is further detected for the case where improper behavior, the final body by extracting user inclines
The duration at angle and improper body posture judge again using thresholding algorithm to reduce False Rate；Judge improper
Whether the tilt angle of posture and duration are more than threshold value that we are arranged, if so then execute step (5), and export user's hair
Raw tumble behavior, it is no to then follow the steps (5)；
(5) learn daily behavior and the tumble behavior of specific user, continuous adjustment and Optimum Classification by on-line study mechanism
Parameter executes step (2).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710063265.XA CN106910314B (en) | 2017-02-03 | 2017-02-03 | A kind of personalized fall detection method based on the bodily form |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710063265.XA CN106910314B (en) | 2017-02-03 | 2017-02-03 | A kind of personalized fall detection method based on the bodily form |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106910314A CN106910314A (en) | 2017-06-30 |
| CN106910314B true CN106910314B (en) | 2019-01-25 |
Family
ID=59208839
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710063265.XA Active CN106910314B (en) | 2017-02-03 | 2017-02-03 | A kind of personalized fall detection method based on the bodily form |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106910314B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108021888B (en) * | 2017-12-05 | 2021-09-24 | 电子科技大学 | Fall detection method |
| CN109009145A (en) * | 2018-07-24 | 2018-12-18 | 西安工程大学 | A kind of tumble judgment method based on wearable device |
| CN109166275B (en) * | 2018-09-25 | 2020-08-11 | 山东科技大学 | Human body falling detection method based on acceleration sensor |
| CN110558936A (en) * | 2019-08-09 | 2019-12-13 | 南昌大学 | risk assessment method for balancing ability of old people |
| CN111401507B (en) * | 2020-03-12 | 2021-01-26 | 大同公元三九八智慧养老服务有限公司 | Adaptive decision tree fall detection method and system |
| CN111709277B (en) * | 2020-04-29 | 2024-07-19 | 平安科技（深圳）有限公司 | Human body fall detection method, device, computer equipment and storage medium |
| CN114997245A (en) * | 2022-07-19 | 2022-09-02 | 广州帝隆科技股份有限公司 | Artificial intelligence-based anti-falling method and device |
| CN117589128B (en) * | 2024-01-17 | 2024-04-12 | 山东交通学院 | Method, equipment and medium for identifying road cone attitude change |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9036902B2 (en) * | 2007-01-29 | 2015-05-19 | Intellivision Technologies Corporation | Detector for chemical, biological and/or radiological attacks |
| 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 |
| CN204971278U (en) * | 2015-08-24 | 2016-01-20 | 华南理工大学 | Senile dementia monitor system based on health service robot |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| BE1017989A5 (en) * | 2008-02-08 | 2010-03-02 | Verhaert New Products & Services Nv | WELFARE DETECTOR. |
| CN203163864U (en) * | 2013-04-11 | 2013-08-28 | 广东乐心医疗电子股份有限公司 | Body weight scale with body mass index display function |
| CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
| CN105989694A (en) * | 2015-02-05 | 2016-10-05 | 江南大学 | Human body falling-down detection method based on three-axis acceleration sensor |
| CN105590408B (en) * | 2016-02-06 | 2018-05-04 | 高强 | A kind of tumble detection method for human body and protective device |
-
2017
- 2017-02-03 CN CN201710063265.XA patent/CN106910314B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9036902B2 (en) * | 2007-01-29 | 2015-05-19 | Intellivision Technologies Corporation | Detector for chemical, biological and/or radiological attacks |
| 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 |
| CN204971278U (en) * | 2015-08-24 | 2016-01-20 | 华南理工大学 | Senile dementia monitor system based on health service robot |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106910314A (en) | 2017-06-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106910314B (en) | A kind of personalized fall detection method based on the bodily form | |
| Quaid et al. | Wearable sensors based human behavioral pattern recognition using statistical features and reweighted genetic algorithm | |
| Gao et al. | Adaptive weighted imbalance learning with application to abnormal activity recognition | |
| Nickel et al. | Authentication of smartphone users based on the way they walk using k-NN algorithm | |
| Chernbumroong et al. | Activity classification using a single wrist-worn accelerometer | |
| CN110287825B (en) | Tumble action detection method based on key skeleton point trajectory analysis | |
| CN106237604A (en) | Wearable device and the method utilizing its monitoring kinestate | |
| CN110390565A (en) | The method and system of intelligent gateway adaptive management are realized by AI edge calculations | |
| CN108244744A (en) | A kind of method of moving state identification, sole and footwear | |
| Muaaz et al. | Influence of different walking speeds and surfaces on accelerometer-based biometric gait recognition | |
| Zhao et al. | Recognition of human fall events based on single tri-axial gyroscope | |
| CN109009145A (en) | A kind of tumble judgment method based on wearable device | |
| CN111582361A (en) | Human behavior recognition method based on inertial sensor | |
| Garcia-Ceja et al. | An improved three-stage classifier for activity recognition | |
| Ahmad et al. | Seeking optimum system settings for physical activity recognition on smartwatches | |
| Luqian et al. | Human activity recognition using time series pattern recognition model-based on tsfresh features | |
| Kong et al. | Design of computer interactive system for sports training based on artificial intelligence and improved support vector | |
| CN114881079A (en) | Human body movement intention abnormity detection method and system for wearable sensor | |
| Bai et al. | Application and research of MEMS sensor in gait recognition algorithm | |
| Cao et al. | ActiRecognizer: Design and implementation of a real-time human activity recognition system | |
| Zhang et al. | An efficient feature selection method for activity classification | |
| Alvee et al. | Application of machine learning classifiers for predicting human activity | |
| Dwiyantoro et al. | A simple hierarchical activity recognition system using a gravity sensor and accelerometer on a smartphone | |
| Mekruksavanich et al. | Deep Pyramidal Residual Network for Indoor-Outdoor Activity Recognition Based on Wearable Sensor. | |
| Youn et al. | New gait metrics for biometric authentication using a 3-axis acceleration |
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 |