CN108021888A - A kind of fall detection method - Google Patents
A kind of fall detection method Download PDFInfo
- Publication number
- CN108021888A CN108021888A CN201711268665.0A CN201711268665A CN108021888A CN 108021888 A CN108021888 A CN 108021888A CN 201711268665 A CN201711268665 A CN 201711268665A CN 108021888 A CN108021888 A CN 108021888A
- Authority
- CN
- China
- Prior art keywords
- acceleration
- svm
- value
- signal vector
- tumble
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention discloses a kind of fall detection method, the present invention is primarily based on training sample training decision threshold：After carrying out median filter process to training sample, then acceleration signal vector magnitude is extracted, and extract characteristic value：Include the peak value of acceleration signal vector magnitude, the minimum value of acceleration signal vector magnitude and the difference of maximum, the standard deviation of acceleration signal vector magnitude, and relative angle changing value；The method for being then based on K means clusters trains two points of decision thresholds of each characteristic value；When carrying out fall detection in real time again, treat detection object original acceleration information carry out the pretreatment identical with training sample after extract corresponding four characteristic values, then decision process step by step is carried out to each characteristic value, acquisition fall detection result.The present invention can be used for the real-time monitoring to old group tumble situation, its computer complexity is low, be that the tumble state that can be achieved to carrier is detected in real time in existing human body portable equipment.
Description
Technical field
The invention belongs to field of computer technology, and in particular to a kind of fall detection based on acceleration transducer.
Background technology
In fall detection processing, relatively common detection mode has：The fall detection of view-based access control model, falling based on sound
Detection and sensor-based fall detection.
Wherein, the fall detection mode of view-based access control model mainly realizes tumble shape by the visual information of human motion process
The detection of condition, during the fall detection of view-based access control model, catches the movement vision information of human body by camera, and
Graphical analysis and processing are carried out to each two field picture in the information of seizure, realize the identification to human body in image, finally,
Analyzed by the drastically degree changed to human body in different two field pictures to realize the judgement to tumble situation.But base
Can only realize the fall detection to video area in the fall detection of video, for video area outside region can not fall
Detection.
Then it is that the change of acoustic information during being fallen by human body is realized in the fall detection mode based on sound
Judgement to tumble situation.But in everyday environments, since the source of acoustic information in environment is various, cause to be based on sound
Fall detection be highly susceptible to disturb, therefore the fall detection based on acoustic information is more as a kind of auxiliary for improving accuracy of detection
Means.
In sensor-based fall detection mode, its motion sensor device worn by human body, realization is based on
The fall detection of body motion information.Common motion sensor includes acceleration transducer, gyroscope, pressure sensor etc.,
During human body is fallen, change drastically can occur for its athletic posture, by above-mentioned motion sensor to human body
The information such as acceleration, angular speed, plantar pressure during attitudes vibration are acquired and analyze, and can realize and fall to human body
The detection of situation.Such as document " tumble detection method for human body [J] computer engineering and science based on acceleration transducer,
2017,39(2):In fall detection method based on acceleration transducer disclosed in 330-335 ", it uses support vector machines pair
The threshold value selection of acceleration signature is optimized；Document " human body fall detection system design [J] meters based on smart mobile phone
Calculation machine engineering and design, 2014,35 (4):In fall detection method disclosed in 1465-1470 ", it has used acceleration transducer
With two kinds of motion sensors of gyroscope, realized by the situation of change of acceleration and angular speed during analysis tumble to falling
The accurate detection of process.But the detection efficiency and accuracy of detection of existing sensor-based fall detection method all up for into
One step improves.
The content of the invention
The goal of the invention of the present invention is：A kind of detection efficiency and the more preferable base of accuracy of detection are provided based on decision Tree algorithms
In the fall detection method of sensor.
The fall detection method of the present invention comprises the following steps：
Step 1：Fall detection threshold value is set：
101：Gather training sample set, three-dimensional (x, y, z direction of principal axis) acceleration letter that the training sample carries for human body
Cease one section of discrete original acceleration sequence of acquisition terminal collection；
102：Data prediction is carried out to training sample set：
Median filter process is carried out to original acceleration information, then extracts acceleration signal vector magnitude, each sampled point
Acceleration signal vector magnitude is the evolution of the quadratic sum of each dimension component of the three-dimensional acceleration information of current sampling point；
103：Extract the characteristic data set of training sample：
Based on acceleration signal vector magnitude, the characteristic data set of training sample, including acceleration signal vector width are extracted
The peak value SVMtop of degree, the minimum value of acceleration signal vector magnitude and the difference DELTA SVM of maximum, acceleration signal vector width
The standard deviation sigma (SVM) of degree, and relative angle changing value Δ θ, the relative angle changing value Δ θ are the maximum of training sample
With the difference of minimum angle-of-incidence；
104：Method based on K-means clusters trains two points of decision thresholds of each characteristic value, obtains corresponding relative angle
Changing value Δ θ, peak value SVMtop, the first, second and third of difference DELTA SVM and standard deviation sigma (SVM) and four threshold values, two points of judgements
Represent when whether leading decision object is tumble state；Wherein, first, second and third and four the preferred value of threshold value be respectively：1.191、
3.274、2.945、0.148。
Step 2：Detect the tumble state of object to be detected：
201：Gathered by the three-dimensional acceleration information acquisition terminal of the carrying of detection object one section it is discrete original plus
Velocity series, as original data to be tested；
202：Median filter process is carried out to original data to be tested, then extracts the acceleration signal vector of object to be detected
Amplitude, and the acceleration signal vector magnitude is based on, extract the characteristic data set of object to be detected, including relative angle change
It is worth Δ θ, peak value SVMtop, difference DELTA SVM and standard deviation sigma (SVM)；
203：The tumble state of object to be detected is judged step by step：
Judge whether current relative angle changing value Δ θ is more than first threshold, if it is not, then judging that current object to be detected is
Non- tumble；If so, continue to judge whether present peak value SVMtop is more than second threshold, if it is not, then judging current object to be detected
For non-tumble；If so, then continue to judge whether current difference Δ SVM is more than the 3rd threshold value, if it is not, then judgement is current to be detected right
As for non-tumble；If so, then continue to judge whether current standard deviation σ (SVM) is less than the 4th threshold value, if it is not, then judging currently to treat
Detection object is non-tumble；If so, current object to be detected is then judged to fall.
In conclusion by adopting the above-described technical solution, the beneficial effects of the invention are as follows：The present invention can be used for old age
The real-time monitoring of colony's tumble situation, understands and rescue the timely of Falls Among Old People situation with facilitating.The computer of the present invention is answered
Miscellaneous degree is low, matches with the computing resource of human body portable equipment, in existing human body portable equipment (such as mobile phone)
Realize that the tumble state to carrier is detected in real time, and the accuracy rate detected meets primary demand, its availability is high.
Brief description of the drawings
Fig. 1 is embodiment flow chart；
Fig. 2 is unfiltered walking acceleration plots；
Fig. 3 is to have filtered walking acceleration plots；
Fig. 4 is state accelerating curve comparison diagram of jogging；
Fig. 5 is fall detection illustraton of model.
Embodiment
To make the object, technical solutions and advantages of the present invention clearer, with reference to embodiment and attached drawing, to this hair
It is bright to be described in further detail.
Referring to Fig. 1, the present invention is realized based on decision tree in the processing of fall detection, passes through training dataset realization pair first
The structure of fall detection model；Then the data characteristics for detection object being treated by constructed fall detection model is fallen
Judge, obtain fall detection as a result, it is mainly concerned with following five steps：
1st, acceleration information acquisition：For the acceleration information (training and detection judge) used in fall detection scheme,
By can the acceleration transducer that is carried (such as Android intelligent terminal) of carried terminal be acquired and store can carried terminal institute
The acceleration information of collection；
2nd, acceleration information pre-processes：The preprocessing process of information is used to carry out the noise included in acceleration information
Filter, in the case of ensureing that data characteristics is immovable, the acceleration signal progress to collection is a degree of smooth, and eliminates
Difference caused by the directionality of acceleration transducer.
3rd, the extraction of data characteristics：In order to realize the fall detection based on acceleration information, from pretreated acceleration
Four characteristic information (SVMtop features, Δ SVM features, σ (SVM) features and Δ θ features) composition characteristic data are extracted in signal
Collection.
4th, the structure of decision model：After the extraction work of data characteristics is completed, rule are differentiated to falling with sorting algorithm
Then it is configured, so as to complete the structure of fall detection model.
5th, tumble result is judged：After the structure of fall detection model is completed, the characteristic data set based on object to be detected
Fall detection judgement is carried out, obtains testing result.
Various processes are implemented as follows：
1st, acceleration information acquisition.
In present embodiment, the acquisition terminal used for smart mobile phone, three dimension acceleration sensor built in it and
Three-dimensional gyroscope can realize the multiple functions such as motion measurement, angle detecting.Wearing position selects the waist of human body, so as to more have
Handled beneficial to the fall detection completed to old group.Due to the diversity of human body behavior, will include in gatherer process to people
The normal walking of body, climb building, jog, jump, forward direction is fallen, it is backward fall, the three-dimensional acceleration under lateral a variety of situations such as fall
Degree information is acquired.
2nd, acceleration information pre-processes.
(1) noise filters.
Since acceleration of motion component, the acceleration of gravity of human body can be included in the acceleration information of carried terminal collection
The information such as component and acceleration analysis noise, before handling acceleration information, if not to there are the acceleration of noise
Degree information is filtered processing, then follow-up analysis work will be had some impact on.
The filtering process to original acceleration information is realized using median filter in present embodiment, is passed through
After median filter is filtered processing to original three-dimensional acceleration signal sequence, acceleration signal will obtain a degree of
Smoothly, so as to achieve the purpose that to eliminate the noise in acceleration signal, Fig. 2 and Fig. 3 respectively show the walking states before filtering
Accelerating curve and filtered walking states accelerating curve, are found, change of the median filter to acceleration by contrasting
Curve plays certain smoothing effect.Wherein, the principle of medium filtering is：The window of certain length is selected to streak signal successively
Sequence, in the region that window streaks, the element of window middle position is replaced with the intermediate value that the window includes element, from
And realize filtering.And the filter effect of median filter is influenced by filter window size, when window is excessive, medium filtering
Effect may be deteriorated, therefore, when carrying out medium filtering to three-dimensional acceleration information, the filter window size of median filter
It is preferably arranged to 3.
(2) Data Synthesis.
During acceleration collection is carried out, occurrence and the three-dimensional acceleration of acceleration pass three dimension acceleration sensor
The direction of sensor, which exists, to be closely connected, and the difference of sensor orientation will cause X-axis, Y-axis, the difference of acceleration information on Z axis
It is different, therefore, for the acceleration information of three-dimensional acceleration sensor collection, by the signal vector amplitude for calculating acceleration information
(Signal Vector Magnitude, SVM) come realize to acceleration information synthesis processing, so as to eliminate acceleration sensing
The directionality of device is to influence caused by the subsequent treatment of acceleration information.
The calculation formula of acceleration signal vector magnitude SVM isWherein SVMtRepresent t
Acceleration signal vector magnitude after the synthesis of moment acceleration, ax,t、ay,t、az,tT moment is represented respectively, and x, y, z direction of principal axis adds
Speed.
After Data Synthesis being carried out based on acceleration signal vector magnitude SVM to the acceleration information of collection, resultant acceleration
Will be unrelated with the direction of acceleration transducer equipment, it reacts the drastically program of human motion brief acceleration change.Fig. 4 is shown
The change curve of human body acceleration information under the state of jogging, using acceleration signal vector magnitude SVM to three-dimensional acceleration
After information is synthesized, its resultant acceleration information is as shown in FIG..By accelerating to three-dimensional acceleration change curve and synthesis
Knowable to the changing rule of degree change curve is analyzed, fluctuation situation and the acceleration before synthesis of the accelerating curve after synthesis
Curve is basically identical, it can be seen that, more accurately reflected based on the acceleration information after vector acceleration amplitude SVM processing
The acceleration change situation of human body during exercise.
3rd, the extraction of data characteristics.
In present embodiment, resultant acceleration peak value SVM will be extracted from three-dimensional acceleration informationtop, synthesis plus
Tetra- speed difference Δ SVM, resultant acceleration standard deviation sigma (SVM) and body obliquity changes delta θ features are as tumble inspection scheme
Feature vector.
(1)SVMtopFeature.
Resultant acceleration peak value SVMtopRepresent the maximum in the resultant acceleration curve of human motion, it reflects appearance
It is worth the drastically degree of moment movement velocity change.During the tumble of human body, the acceleration of human body can produce change drastically,
Its resultant acceleration will reflect increase drastically, therefore, can be realized to human body tumble situation based on resultant acceleration peak value
Preliminary judgement.
SVM corresponding to Different activity states is concentrated to training datatopAverage value is calculated, its result of calculation such as table 1
It is shown.By the way that the data in table are observed and can be obtained, people daily walking, climb building, jog during synthesis accelerate
Spend peak value SVMtopNot over 3g, and the resultant acceleration peak value of tumble state and jump state has generally reached more than 4g, by
This finds out that it is obvious poor to exist between the resultant acceleration peak value under resultant acceleration peak value and non-tumble state under tumble state
It is different.
Table 1
| Concrete behavior | SVMtopAverage value |
| Walking | 1.6753 |
| Climb building | 1.7531 |
| Jog | 2.8305 |
| Jump | 4.2406 |
| Forward direction is fallen | 4.2105 |
| It is lateral to fall | 4.8931 |
| It is backward to fall | 4.5014 |
(2) Δ SVM features.
The difference of the maxima and minima of resultant acceleration during the difference DELTA SVM expression human motions of resultant acceleration,
It has reflected the drastically degree of human motion.During the daily behavioral activity of human body such as walks and climbs building, its behavior
Action to tend towards stability, therefore the difference of the maxima and minima corresponding to its resultant acceleration curve is relatively small.And for
Tumble behavior, because be able to can reflect in resultant acceleration curve with the change dramatically of acceleration during tumble
The appearance of one very big peak value, at this time, the resultant acceleration that the difference of resultant acceleration should be greater than under daily common behavior are poor
Value.
Table 2 below shows the Δ SVM average values corresponding to the different motion state based on training dataset, from table
Middle analysis can obtain, and the resultant acceleration difference DELTA SVM under tumble state is much larger than the Δ SVM under daily common behavior.
Table 2
| Concrete behavior | Δ SVM average values |
| Walking | 1.0554 |
| Climb building | 1.1698 |
| Jog | 2.7350 |
| Jump | 4.3067 |
| Forward direction is fallen | 3.9406 |
| It is lateral to fall | 4.6055 |
| It is backward to fall | 4.0968 |
(3) σ (SVM) feature.
After falling occurs in human body, and because tumble injury causes to stand up, at this time, a period of time occurs in human body
Inactive state, the resultant acceleration curve under remaining static there will not be big ups and downs, and therefore, the present invention is by human body
The resultant acceleration standard deviation in T seconds (T is experience preset value) section after tumble is expressed as σ (SVM), is examined as falling
The feature surveyed weighs the fluctuation situation of resultant acceleration after human body is fallen, so as to fulfill the static shape after falling to human body
The judgement of state, wherein, the calculation formula of standard deviation sigma (SVM) is：Wherein N
Represent the number of the acceleration signal vector magnitude SVM in T seconds sections, SVMμRepresent N number of SVMiAverage.
σ (SVM) average value under Different activity states is shown in table 3 below, wherein, σ (SVM) average value passes through
Training dataset is calculated.Data analysis in table can obtain, the resultant acceleration mark after human body is fallen under inactive state
Quasi- difference σ (SVM) not less than 0.1, compared to walking, upstairs, daily behavior state, its σ (SVM) such as jog there is significant difference.
Table 3
(4) Δ θ features.
When human body is fallen, the relative angle of human body can all change, and therefore, the present invention is by before and after human motion
The changing value of relative angle is expressed as Δ θ, and as change of the feature for reacting human body behavior posture.
By by the inclination maximum θ during human motionmaxSubtract minimum angle-of-incidence θminJust the change of pitch angle of human body is obtained
It is worth Δ θ, table 4 below is the corresponding Δ θ average values of human body Different activity states, wherein, Δ θ average values are calculated by training dataset
Obtain.By analyzing data in table, human body is in the case where falling, its body change of pitch angle reflected
It will be greater than the change of pitch angle under daily behavior state.
Table 4
| Concrete behavior | Δ θ average values |
| Walking | 0.7052 |
| Climb building | 0.7404 |
| Jog | 0.9888 |
| Jump | 1.0166 |
| Forward direction is fallen | 1.5769 |
| It is lateral to fall | 1.4594 |
| It is backward to fall | 1.5596 |
4th, the structure of decision model.
(1) characteristic threshold value is set.
When being configured using sorting algorithm to tumble decision rule, need to be carried out after discretization just for continuous type feature
Can complete the calculating of information gain-ratio, the discretization mode of generally use is threshold value division, frequently with mode have artificial observation
Method and information gain-ratio method.Since the introducing of potential error will be caused during artificial observation method selected threshold, thus it is special by calculating
The point of information gain-ratio maximum in section is levied as threshold point, so as to avoid the introducing of potential error.
When carrying out the calculating of classification thresholds based on information gain-ratio, its time complexity calculated is O (n), and wherein n is
The number of characteristic value in section, due to that can be related to the calculating of the secondary information gain-ratios of O (n) in which, and information gain-ratio
Substantial amounts of logarithm operation can cause the increase of operand in calculation formula, therefore, will have when which calculates threshold value compared with the matter of fundamental importance
Calculation amount.Therefore it is of the invention when the continuous feature to acceleration carries out threshold value selection, using the method for K-means clusters to threshold of classifying
Value is calculated, its calculating process is：Two classification carry out all characteristic values in characteristic interval using K-means clusters, and
Threshold point using the midpoint in classification results on cluster centre line as interval division, is realized to section by this threshold point
Classification.Wherein, the time complexity of K-means clustering algorithms is O (k*m*n*t), and k represents the cluster target of K-means clusters
Individual, m represent the dimension of feature, and n represents the number of characteristic value in characteristic interval, and t represents the iteration in K-means cluster process
Number.When carrying out two classification to characteristic interval using K-means clustering algorithms, the dimension m of feature is 1, K-means clusters
Cluster target k is that the iterations of 2, K-means clusters is constant, therefore, is clustered using K-means and carries out characteristic interval
When threshold value is chosen, its corresponding time complexity is O (n).Compared to based on information gain-ratio carry out threshold point calculating method,
Although with it with identical time complexity o (n), K-means clusters are corresponding to be calculated K-means clustering algorithms every time
There was only the simple plus and minus calculation of distance between element in journey, and the calculating process of information gain-ratio can be related to substantial amounts of logarithm fortune
Calculate, therefore, carry out threshold value selection to section using K-means clusters by with of a relatively high computational efficiency.
Corresponding threshold point and information gain-ratio and use when table 5 below is to using K-means cluster progress threshold value selections
Information gain-ratio threshold point corresponding when carrying out threshold value selection and information gain-ratio are contrasted.Data point in table
Analyse and understand, when the threshold value for carrying out acceleration signature is chosen, select the information corresponding to the threshold value of K-means clustering algorithms selection
Ratio of profit increase is sufficiently close to the information gain-ratio corresponding to the threshold value chosen based on information gain-ratio, it is believed that it has almost phase
Same classifying quality, and in terms of operation efficiency, substantial amounts of logarithm operation can cause computing in the calculation formula of information gain-ratio
The increase of amount, therefore, is clustered using K-means and carries out threshold value selection by with certain odds for effectiveness.
Table 5
| Feature | Existing threshold value | K-means threshold values | Existing information ratio of profit increase | K-means information gain-ratios |
| SVMtop | 3.355 | 3.274 | 0.529 | 0.505 |
| ΔSVM | 3.112 | 2.945 | 0.517 | 0.481 |
| σ(SVM) | 0.134 | 0.148 | 0.487 | 0.442 |
| Δθ | 1.235 | 1.191 | 0.869 | 0.831 |
According to it is above-mentioned 4 features in acceleration are carried out classification thresholds calculating as a result, existing by SVMtop、ΔSVM、σ
(SVM) and the classification thresholds of Δ θ are respectively set to the classification thresholds point that K-means clustering algorithms are calculated, and are respectively
3.274th, 2.945,0.148 and 1.191.Because human body is during tumble, its resultant acceleration peak value SVMtopCan drastically it increase
Greatly, therefore, as the SVM of human body behaviortopDuring more than threshold value 3.274, then it is assumed that the doubtful of human body is fallen, and works as SVMtop
During less than threshold value, then it is assumed that human body is in non-tumble state.Similarly, when Δ SVM is more than threshold value 2.945, then it is assumed that human body goes out
The situation of existing doubtful tumble, it is on the contrary then be in non-tumble state.For σ (SVM), because human body is when tumble can occur one section
Between inactive state, therefore, when σ (SVM) is less than threshold value 0.148, it is believed that human body it is doubtful in fall after inactive state, then
Human body is doubtful at this time falls.For Δ θ, because in front and rear body posture of falling change dramatically can occur for human body, Δ θ will
The intensity of variation of body obliquity is reflected, when Δ θ is more than threshold value 1.191, then it is assumed that human body is doubtful to be fallen, it is on the contrary then
Think that human body is in non-tumble state.
(2) fall detection model is built.
Referring to Fig. 5, tumble state is judged by Δ θ features first, due to human body during tumble body
Angle of inclination acute variation will occur, therefore, when Δ θ is less than or equal to threshold value 1.191, then it is assumed that user, which is in, non-to fall
State, when Δ θ is more than threshold value 1.191, then it is assumed that human body is doubtful to be occurred falling, but can not be carried out to tumble situation accurate
Really judge, therefore, will further be judged.
Due to resultant acceleration peak value SVM of the human body under tumble statetopOccur that one sharp raises, therefore, when
SVMtopDuring less than or equal to threshold value 3.274, then it is assumed that current human is in normally performed activity state, does not fall, when
SVMtopDuring more than threshold value 3.274, then it is assumed that human body is doubtful this moment is fallen.Because during human body is fallen, people's
Resultant acceleration, which occurs, to be sharply increased, and human action is relatively gentle before and after falling, its resultant acceleration is relatively small, because
This, when meeting SVMtopAfter feature and Δ θ features, it can be sentenced by tumble situations of the resultant acceleration difference DELTA SVM to user
It is disconnected, when Δ SVM is less than or equal to threshold value 2.945, then it is assumed that human body does not occur tumble situation, when a threshold is exceeded, then also needs
Continue the judgement of next step.
When the motion state of human body meets SVMtopAfter feature, Δ θ features and Δ SVM features, finally by σ (SVM) to
The tumble state at family is judged, because the inactive state of a period of time occurs in human body after tumble, σ (SVM) will reflect
Human body resultant acceleration peak value SVMtopAcceleration change situation after appearance, when σ (SVM) is greater than or equal to threshold value 0.148,
Then think human body in SVMtopIt is active after appearance, can determine that current human is in non-tumble state, when σ (SVM) is less than
During threshold value, then the motion state of current human meets SVMtop, Δ SVM, σ (SVM) and tetra- features of Δ θ, therefore can determine whether current
Falling occurs in human body.
5th, the real-time detection of fall detection.
The fall detection process of fall detection model based on the present invention is：
Original acceleration information is obtained based on the acceleration information acquisition terminal that human body carries；
Medium filtering and Data Synthesis processing are carried out to it, obtains acceleration signal vector magnitude；
Extract four characteristic information (SVMtop features, Δ SVM features, σ (SVM) features and Δ θ features) composition characteristic numbers
According to collection；
Decision process step by step is carried out to characteristic data set, obtains fall detection as a result, at different levels be followed successively by：The judgement of Δ θ features,
SVMtopFeature judgement, the judgement of Δ SVM features and the judgement of σ (SVM) feature.
Table 6 gives the experiment results of the fall detection of the present invention, passes through the accuracy in checkout procedure and wrong report
Rate come embody the present invention fall detection performance.
Table 6
Can be seen that the fall detection scheme of the present invention from the verification result in table 6 can complete to most of tumble feelings
The detection of condition, wherein, the Detection accuracy highest of lateral tumble behavior, has reached 92.5%, and forward direction is fallen and backward tumble
Detection accuracy relative reduction.During in view of being acquired to test data set, artificial conscious tumble can cause to fall
The opposite reduction of action so that the test data of collection with data during true tumble there are certain deviation, real
In the case of tumble, the corresponding resultant acceleration peak value SVM of gathered datatop, resultant acceleration difference DELTA SVM, resultant acceleration mark
The performance of this four features of quasi- difference σ (SVM) and body obliquity changes delta θ will become apparent from, therefore can be determined that, really fall
Under situation, fall detection scheme of the invention is by the detection accuracy with higher.In terms of rate of false alarm, test phase is to people
The acceleration information of the basic daily behavior of body is gathered, including walk, climb building, jog, the behavior state such as jump,
The data of collection are brought into fall detection scheme as test data and are detected, walking is can be seen that from the result of detection
Process upstairs has very low rate of false alarm, and jog and jump state under False Rate averagely reach 16.25%, direct shadow
Ring the accuracy in detection for the system that arrived, it is contemplated that fall detection is mainly directed towards old group, and old group is in daily life
Pao Tiaodeng strenuous exercises are seldom carried out, therefore, specially treated need not be carried out for influence of both states to detection accuracy.
Comprehensive all test results show that the correct detection number of fall detection scheme is 292 times, and erroneous judgement number is 15 times, is failed to report
Number is 13 times, and comprehensive accuracy has reached 91.25%, and therefore, the accuracy rate of fall detection scheme of the invention meets to fall
The primary demand of detection, possesses availability in actual application.
The above description is merely a specific embodiment, any feature disclosed in this specification, except non-specifically
Narration, can be replaced by other alternative features that are equivalent or have similar purpose；Disclosed all features or all sides
Method or during the step of, in addition to mutually exclusive feature and/or step, can be combined in any way.
Claims (3)
1. a kind of fall detection method, it is characterised in that comprise the following steps：
Step 1：Fall detection threshold value is set：
101：Training sample set is gathered, the training sample is the one of the three-dimensional acceleration information acquisition terminal collection of human body carrying
The discrete original acceleration sequence of section；
102：Data prediction is carried out to training sample set：
Median filter process is carried out to original acceleration information, then extracts acceleration signal vector magnitude, the acceleration of each sampled point
Spend evolution of the signal vector amplitude for the quadratic sum of each dimension component of the three-dimensional acceleration information of current sampling point；
103：Extract the characteristic data set of training sample：
Based on acceleration signal vector magnitude, the characteristic data set of training sample is extracted, including acceleration signal vector magnitude
Peak value SVMtop, the minimum value of acceleration signal vector magnitude and the difference DELTA SVM of maximum, acceleration signal vector magnitude
Standard deviation sigma (SVM), and relative angle changing value Δ θ, the relative angle changing value Δ θ are for the maximum of training sample and most
The difference of small inclination；
104：Method based on K-means clusters trains two points of decision thresholds of each characteristic value, obtains corresponding relative angle change
It is worth first, second and third and four threshold values of Δ θ, peak value SVMtop, difference DELTA SVM and standard deviation sigma (SVM), two points of judgements represent
When whether leading decision object is tumble state；
Step 2：Detect the tumble state of object to be detected：
201：The one section of discrete original acceleration gathered by the three-dimensional acceleration information acquisition terminal of the carrying of detection object
Sequence, as original data to be tested；
202：Median filter process is carried out to original data to be tested, then extracts the acceleration signal vector width of object to be detected
Degree, and the acceleration signal vector magnitude is based on, extract the characteristic data set of object to be detected, including relative angle changing value
Δ θ, peak value SVMtop, difference DELTA SVM and standard deviation sigma (SVM)；
203：The tumble state of object to be detected is judged step by step：
Judge whether current relative angle changing value Δ θ is more than first threshold, if it is not, then judging that current object to be detected falls to be non-
；If so, continue to judge whether present peak value SVMtop is more than second threshold, if it is not, then judging that current object to be detected is non-
Fall；If so, then continue to judge whether current difference Δ SVM is more than the 3rd threshold value, if it is not, then judging that current object to be detected is
Non- tumble；If so, then continue to judge whether current standard deviation σ (SVM) is less than the 4th threshold value, if it is not, then judging current to be detected
Object is non-tumble；If so, current object to be detected is then judged to fall.
2. the method as described in claim 1, it is characterised in that described first, second and third and four the preferred value of threshold value be respectively：
1.191、3.274、2.945、0.148。
3. method as claimed in claim 1 or 2, it is characterised in that the filter window size of the median filter process is preferred
It is arranged to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711268665.0A CN108021888B (en) | 2017-12-05 | 2017-12-05 | Fall detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711268665.0A CN108021888B (en) | 2017-12-05 | 2017-12-05 | Fall detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108021888A true CN108021888A (en) | 2018-05-11 |
| CN108021888B CN108021888B (en) | 2021-09-24 |
Family
ID=62078669
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711268665.0A Active CN108021888B (en) | 2017-12-05 | 2017-12-05 | Fall detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108021888B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110327050A (en) * | 2019-05-05 | 2019-10-15 | 北京理工大学 | People's tumble state embedded intelligence detection method for wearable equipment |
| CN110443798A (en) * | 2018-12-25 | 2019-11-12 | 电子科技大学 | A kind of self-closing disease detection method based on magnetic resonance image, apparatus and system |
| CN111436944A (en) * | 2020-04-20 | 2020-07-24 | 电子科技大学 | Falling detection method based on intelligent mobile terminal |
| CN112561024A (en) * | 2020-12-23 | 2021-03-26 | 中移（杭州）信息技术有限公司 | Fall detection method and system, communication equipment and storage medium |
| CN113435244A (en) * | 2021-05-14 | 2021-09-24 | 杭州电子科技大学 | Fall-down behavior detection method based on fuzzy entropy |
| CN115024717A (en) * | 2022-08-09 | 2022-09-09 | 广东百年医疗健康科技发展有限公司 | Fall detection method, device, equipment and storage medium |
| CN115192049A (en) * | 2022-09-14 | 2022-10-18 | 深圳市心流科技有限公司 | Myoelectricity sampling frequency adjusting method of intelligent artificial limb |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007184715A (en) * | 2006-01-05 | 2007-07-19 | Softbank Mobile Corp | Information communication terminal |
| US20090147991A1 (en) * | 2007-12-07 | 2009-06-11 | Tom Chau | Method, system, and computer program for detecting and characterizing motion |
| CN101711401A (en) * | 2007-04-19 | 2010-05-19 | 皇家飞利浦电子股份有限公司 | Fall detection system |
| CN102186420A (en) * | 2008-10-16 | 2011-09-14 | 皇家飞利浦电子股份有限公司 | Fall detection system |
| CN102841818A (en) * | 2011-06-23 | 2012-12-26 | 西部数据技术公司 | Electronic device avoiding false detection of free fall event based on detected operating mode |
| CN102855321A (en) * | 2012-09-07 | 2013-01-02 | 陕西师范大学 | Human behavior recognition method |
| US20130054180A1 (en) * | 2011-08-29 | 2013-02-28 | James R. Barfield | Method and system for detecting a fall based on comparing data to criteria derived from multiple fall data sets |
| US20130120152A1 (en) * | 2011-11-14 | 2013-05-16 | Vigilo Networks, Inc. | Method and system for fall detection of a user |
| CN103308069A (en) * | 2013-06-04 | 2013-09-18 | 电子科技大学 | Falling-down detection device and method |
| US9005141B1 (en) * | 2007-10-12 | 2015-04-14 | Biosensics, L.L.C. | Ambulatory system for measuring and monitoring physical activity and risk of falling and for automatic fall detection |
| CN105749548A (en) * | 2015-01-05 | 2016-07-13 | 索尼公司 | Information Processing Device, Information Processing Method, And Program |
| CN106599914A (en) * | 2016-12-07 | 2017-04-26 | 广东工业大学 | Multi-sensor fused wristed falling detection method and device |
| CN106910314A (en) * | 2017-02-03 | 2017-06-30 | 同济大学 | A kind of personalized fall detection method based on the bodily form |
| CN106981174A (en) * | 2017-04-27 | 2017-07-25 | 南京邮电大学 | A kind of Falls Among Old People detection method based on smart mobile phone |
-
2017
- 2017-12-05 CN CN201711268665.0A patent/CN108021888B/en active Active
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007184715A (en) * | 2006-01-05 | 2007-07-19 | Softbank Mobile Corp | Information communication terminal |
| CN101711401A (en) * | 2007-04-19 | 2010-05-19 | 皇家飞利浦电子股份有限公司 | Fall detection system |
| US9005141B1 (en) * | 2007-10-12 | 2015-04-14 | Biosensics, L.L.C. | Ambulatory system for measuring and monitoring physical activity and risk of falling and for automatic fall detection |
| US20090147991A1 (en) * | 2007-12-07 | 2009-06-11 | Tom Chau | Method, system, and computer program for detecting and characterizing motion |
| CN102186420A (en) * | 2008-10-16 | 2011-09-14 | 皇家飞利浦电子股份有限公司 | Fall detection system |
| CN102841818A (en) * | 2011-06-23 | 2012-12-26 | 西部数据技术公司 | Electronic device avoiding false detection of free fall event based on detected operating mode |
| US20130054180A1 (en) * | 2011-08-29 | 2013-02-28 | James R. Barfield | Method and system for detecting a fall based on comparing data to criteria derived from multiple fall data sets |
| US20130120152A1 (en) * | 2011-11-14 | 2013-05-16 | Vigilo Networks, Inc. | Method and system for fall detection of a user |
| CN102855321A (en) * | 2012-09-07 | 2013-01-02 | 陕西师范大学 | Human behavior recognition method |
| CN103308069A (en) * | 2013-06-04 | 2013-09-18 | 电子科技大学 | Falling-down detection device and method |
| CN105749548A (en) * | 2015-01-05 | 2016-07-13 | 索尼公司 | Information Processing Device, Information Processing Method, And Program |
| CN106599914A (en) * | 2016-12-07 | 2017-04-26 | 广东工业大学 | Multi-sensor fused wristed falling detection method and device |
| CN106910314A (en) * | 2017-02-03 | 2017-06-30 | 同济大学 | A kind of personalized fall detection method based on the bodily form |
| CN106981174A (en) * | 2017-04-27 | 2017-07-25 | 南京邮电大学 | A kind of Falls Among Old People detection method based on smart mobile phone |
Non-Patent Citations (6)
| Title |
|---|
| DONGHA LIM 等: "Fall-Detection Algorithm Using 3-Axis Acceleration:Combination with Simple Threshold and Hidden Markov Model", 《JOURNAL OF APPLIED MATHEMATICS》 * |
| IVAN MIGUEL PIRES 等: "Data Fusion on Motion and Magnetic Sensors embedded on Mobile Devices for the Identification of Activities of Daily Living", 《ARXIV》 * |
| WEI ZHUANG 等: "A novel wearable smart button system for fall detection", 《AIP CONFERENCE PROCEEDINGS》 * |
| 张娇 等: "基于二分 K均值的 SVM决策树的高维数据分类方法", 《赤峰学院学报（自然科学版）》 * |
| 李政: "一种基于三轴加速度的跌倒检测方法", 《电子科技》 * |
| 王振 等: "数据融合的跌倒检测系统设计", 《传感器与微系统》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443798A (en) * | 2018-12-25 | 2019-11-12 | 电子科技大学 | A kind of self-closing disease detection method based on magnetic resonance image, apparatus and system |
| CN110327050A (en) * | 2019-05-05 | 2019-10-15 | 北京理工大学 | People's tumble state embedded intelligence detection method for wearable equipment |
| CN111436944A (en) * | 2020-04-20 | 2020-07-24 | 电子科技大学 | Falling detection method based on intelligent mobile terminal |
| CN112561024A (en) * | 2020-12-23 | 2021-03-26 | 中移（杭州）信息技术有限公司 | Fall detection method and system, communication equipment and storage medium |
| CN113435244A (en) * | 2021-05-14 | 2021-09-24 | 杭州电子科技大学 | Fall-down behavior detection method based on fuzzy entropy |
| CN115024717A (en) * | 2022-08-09 | 2022-09-09 | 广东百年医疗健康科技发展有限公司 | Fall detection method, device, equipment and storage medium |
| CN115192049A (en) * | 2022-09-14 | 2022-10-18 | 深圳市心流科技有限公司 | Myoelectricity sampling frequency adjusting method of intelligent artificial limb |
| CN115192049B (en) * | 2022-09-14 | 2022-12-09 | 深圳市心流科技有限公司 | Myoelectricity sampling frequency adjusting method of intelligent artificial limb |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108021888B (en) | 2021-09-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108021888A (en) | A kind of fall detection method | |
| Quaid et al. | Wearable sensors based human behavioral pattern recognition using statistical features and reweighted genetic algorithm | |
| EP3058441B1 (en) | Calculating pace and energy expenditure from athletic movement attributes | |
| CN104729507B (en) | A kind of gait recognition method based on inertial sensor | |
| CN106981174A (en) | A kind of Falls Among Old People detection method based on smart mobile phone | |
| Jensen et al. | Classification of kinematic swimming data with emphasis on resource consumption | |
| CN110245718A (en) | A kind of Human bodys' response method based on joint time-domain and frequency-domain feature | |
| CN105268171B (en) | gait monitoring method, device and wearable device | |
| CN107212890A (en) | A kind of motion identification and fatigue detection method and system based on gait information | |
| WO2012119126A2 (en) | Apparatus, system, and method for automatic identification of sensor placement | |
| CN103955699A (en) | Method for detecting tumble event in real time based on surveillance videos | |
| Ghazali et al. | Common sport activity recognition using inertial sensor | |
| CN106725495A (en) | A kind of fall detection method, apparatus and system | |
| CN108717548B (en) | Behavior recognition model updating method and system for dynamic increase of sensors | |
| CN107588784A (en) | A kind of state recognition and the method, apparatus and terminal distinguished | |
| CN110811578A (en) | Step counting device and step counting method thereof, controller and readable storage medium | |
| CN104269025A (en) | Wearable type single-node feature and position selecting method for monitoring outdoor tumble | |
| CN114881079A (en) | Human body movement intention abnormity detection method and system for wearable sensor | |
| KR101793934B1 (en) | Method and apparatus for automatically classifying types of weight training workouts | |
| CN117133057A (en) | Physical exercise counting and illegal action distinguishing method based on human body gesture recognition | |
| Yadav et al. | SPECIAL SESSION ON RECENT ADVANCES IN COMPUTATIONAL INTELLIGENCE & TECHNOLOGYS (SS_10_RACIT) Development of Generic Human Motion Simulation Categorization using Inception based CNN | |
| KR102194313B1 (en) | Apparatus and method for identifying individuals by performing neural network analysis for various detection information | |
| KR20210046121A (en) | Apparatus and method for identify patients with parkinson's disease and patients with podarthritis by performing neural network analysis by various detection information | |
| KR101483218B1 (en) | Activity diagnosis apparatus | |
| KR20210011097A (en) | Apparatus and method for classification of gait type by performing neural network analysis for various detection information |
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 |