CN108509897A - A kind of human posture recognition method and system - Google Patents
A kind of human posture recognition method and system Download PDFInfo
- Publication number
- CN108509897A CN108509897A CN201810269770.4A CN201810269770A CN108509897A CN 108509897 A CN108509897 A CN 108509897A CN 201810269770 A CN201810269770 A CN 201810269770A CN 108509897 A CN108509897 A CN 108509897A
- Authority
- CN
- China
- Prior art keywords
- data
- human body
- sample
- decision
- human
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
The present invention proposes a kind of human posture recognition method and system, belongs to safety engineering field.This approach includes the following steps：(1) linear acceleration and angular acceleration six class parameter of the human body on space X, tri- directions Y, Z is at least acquired；(2) data characteristics that all kinds of parameters are presented is analyzed, is per class t data statistical characteristics of data decimation, as the conditional attribute for judging human body attitude；Wherein, t is the integer more than 1；(3) data statistical characteristics for extracting the data set under all kinds of human body attitudes, constitute sample set, sample set are divided into training sample and test sample；(4) according to training sample and test sample, Decision-Tree Classifier Model is built；(5) it is based on Decision-Tree Classifier Model, real-time grading is carried out to the data acquired, identifies human body attitude.Recognition accuracy of the present invention is high, and recognition speed is fast, and for the adaptable of multi input sample, the output of classification results has high fault tolerance, strong robustness.
Description
Technical field
The invention belongs to safety engineering fields, are related to a kind of identification technology, especially human body attitude identification technology.
Background technology
With the continuous aggravation and the continuous rising of Empty nest elderly quantity of aging of population, people caused by Falls in Old People
Body injury problem becomes increasingly conspicuous.According to statistics, it falls and has become the fourth-largest reason of China's injury scope at present, over-65s
The first reason of injury for aged death, and further increased with the increase tumble death rate at old man's age.Fall in addition to
It directly results in except the elderly's death, also generates a large amount of deformity, reduce the mobility and scope of activities of the elderly, serious shadow
Ring the quality of life and physical and mental health of the elderly.
There are many kinds of the existing methods for studying obtained human motion gesture recognition, and main two kinds of recognition methods are：Base
It is identified in the human body attitude identification of computer vision and the human body attitude based on wearable sensors.Appearance based on computer vision
State identification utilizes some 3D sensors (PrimeSense configured with infrared transmitter, RGB cameras, depth image camera
Or Kinect etc.) three dimensional local information at the positions such as human synovial is obtained, cost is higher, while the image or video generated
Required memory space is larger, can not ensure that data can be effectively transmitted, not repeat.Posture based on wearable sensors
Identification is widely used in the researchs of equipment such as Intelligent bracelet, intelligent running shoes, mainly uses motion sensor and pressure sensing
Device completes the gesture recognition of human body, and this recognition methods acquisition of information is simple, and cost is relatively low, data space demand is small,
Transmission is simple.
The research for carrying out precise real-time detection and alarm to tumble behavior at present mostly utilizes acceleration or gyro data
Time series statistics domain or transform domain feature, use curve similarity comparison algorithm carry out fall detection.This method profit
The weightlessness occurred successively with acceleration time series data during tumble is overweight and its corresponding statistics domain or transform domain
Threshold value carries out fall detection, and the accuracy and robustness of fall detection be not high.Therefore, the human body appearance of accuracy height and strong robustness
State identification technology has extremely strong research significance and application value.
Invention content
The purpose of the present invention is to provide a kind of accuracy height, the human body attitude identification technologies of strong robustness.
In order to achieve the above object, solution of the invention is：
A kind of human posture recognition method, includes the following steps：
(1) linear acceleration and angular acceleration six class parameter of the human body on space X, tri- directions Y, Z is at least acquired；It is described
X, Y, Z-direction are the direction of setting；
(2) data characteristics that all kinds of parameters are presented is analyzed, per class t data statistical characteristics of data decimation, to make
To judge the conditional attribute of human body attitude；Wherein, t is the integer more than 1；
(3) data statistical characteristics for extracting the data set under all kinds of human body attitudes, constitute sample set, by the sample set point
For training sample and test sample；
(4) according to the training sample and the test sample, Decision-Tree Classifier Model is built；
(5) it is based on the Decision-Tree Classifier Model, real-time grading is carried out to the data acquired, identifies human body attitude.
The data acquired described in the step (5) is are acquired and are located according to the step (1) to the step (2)
The data to be identified of reason.
The human body attitude is tumble posture.
The X, Y, Z-direction are mutually perpendicular to；Preferably, the X, Y, Z-direction are respectively using human body to be measured as the water of reference
Flat front-rear direction, horizontal left and right directions and vertical direction；Preferably, the step (1) is using six axle sensors acquisition human body ginseng
Number；It is further preferred that the function of the six axle sensors integrated tri-axial acceleration meter and three-axis gyroscope.
Before the step (2), at least six class parameters that the step (1) acquires are filtered；Preferably,
It is described to be filtered using kalman filter method.
All kinds of human body attitudes in the step (3) include the human body attitude for coming from action sequence；Preferably, described dynamic
Include from standing to tumble, and/or from walking to tumble, and/or from running to tumble, and/or from squatting down tumble as sequence；
Preferably, the action sequence further include by walk, sit elevator, upstairs, downstairs, the continuous action sequence that constitutes of crouching；Further
Preferably, 10 samples are extracted from the action sequence for each human body attitude as training sample, from the action sequence
Select 10 samples as test sample in remaining sample in row.
The data statistical characteristics of data set in the step (3) under all kinds of human body attitudes of extraction, constituting sample set includes：
It is that every class parameter chooses n continuous data points by axis of the time, takes t data statistics characteristic value of every class parameter, altogether n*t
A characteristic value constitutes a sample；The time interval of specimen sample is the sampling time at d consecutive numbers strong point；Wherein, n, t, d are equal
For positive integer, and n ＞ d；Preferably, the sampling time at consecutive numbers strong point is 0.1s.
The Decision-Tree Classifier Model built in the step (4) is that gradient promotes Decision-Tree Classifier Model；Preferably, described
Structure Decision-Tree Classifier Model includes the following steps in step (4)：
(41) initialization gradient promotes Decision-Tree Classifier Model, and estimation makes the constant value of loss function minimization, initialization
It is the tree only there are one root node that gradient afterwards, which promotes Decision-Tree Classifier Model,；
(42) it calculates gradient of the negative gradient of the loss function after the initialization and promotes Decision-Tree Classifier Model
Value, the estimation as residual error；
(43) estimation returns leaf nodes region, with the approximation of regression criterion；
(44) using the value for linearly searching element estimation leaf node region, make loss function minimization；
(45) regression tree is updated；
It is further preferred that the loss function refers to the function of residual error between quantitative prediction value and actual value；
It is further preferred that the residual error subtracts predicted value equal to actual value.
The human posture recognition method further includes after the step (5)：Export the human body of the step (5) identification
Posture；Preferably, it includes when falling, to export tumble signal to recognize human body attitude；Preferably, the human posture recognition method
Further include after the step (5)：In the human body attitude recognized, human body attitude is the output priority highest fallen；Fall
Human body attitude other than is exported according to the logic synthesis for excluding redundancy；It is further preferred that the logic for excluding redundancy
For：
Using the recognition result of a sample as the first recognition result, using the recognition result of l continuous sample as second
Recognition result exports final recognition result in accordance with the following methods based on the recognition result of l continuous sample：
When l the first recognition results include the human body attitude fallen, output human body attitude is to fall to know as final
Other result；Otherwise,
When the l the first recognition results are consistent, output human body attitude is that the consistent recognition result is known as final
Other result；
When the l the first recognition results are inconsistent, the most recognition result of output occurrence number；Know when there are m kinds
Not as a result, and the number that occurs of each recognition result it is equal when, export the recognition result that occurs at first as final identification knot
Fruit.
In the step (5) when structure Decision-Tree Classifier Model, classified using the method reduced decision tree of Feature Selection
Model, Decision-Tree Classifier Model described in the method validation using cross validation.
A kind of human body attitude identifying system for realizing above-mentioned human posture recognition method, including data detector and control
Device；The data detector is at least linear acceleration and angular acceleration six of the acquisition human body on space X, tri- directions Y, Z
Class parameter；The X, Y, Z-direction are the direction of setting, and the parameter of acquisition is sent to the controller；The controller base
In the parameter of data detector acquisition, the human posture recognition method is executed；Preferably, the data detector passes through
The parameter of acquisition is sent to the controller by the mode of bluetooth；Preferably, the data detector senses for six-axle acceleration
Device；Preferably, the controller is host computer；It is further preferred that the host computer is background terminal.
By adopting the above scheme, the beneficial effects of the invention are as follows：The original number that six axle sensor of the present invention couple is acquired
According to Kalman filtering is first carried out, exceptional value and noise jamming are removed, the confidence level of data is reinforced；Then to treated, line accelerates
Degrees of data and angular acceleration data extract data statistical characteristics value, different from only using 3-axis acceleration data as distinguishing rule
Recognition methods, the data of six axis can more effectively reflect the motion characteristic of each human body attitude.In terms of the selection of sample, this hair
It is bright that multiple consecutive numbers strong points more than the sampling interval is selected to constitute individual data collection, extract multiple data characteristics conducts of data set
One sample, the size of sampling interval determine the size of identification frequency.In terms of the selection of grader, the present invention has selected ladder
Degree promotes decision tree (Gradient Tree Boosting) disaggregated model, and classification speed is fast, efficient.In addition, in human body appearance
The output facet of state recognition result uses K (K>1) a recognition result exports the recognition logic of a final result, is provided with
The output priority of emphasis recognition result can export emphasis recognition result, while can be better protected from frequent output in time
As a result, can also prevent the interference that abnormal operation is such as suddenly waved to gesture recognition result.On the whole, the present invention identifies accurate
True rate is high, and recognition speed is fast, and for the adaptable of multi input sample, the output of classification results has high fault tolerance, robustness
By force.
Description of the drawings
Fig. 1 is the flow chart of identification process in human posture recognition method in one embodiment of the invention；
Fig. 2 a are the accelerating curves from standing tumble action sequence in sample set used by this embodiment of the invention
Figure；
Fig. 2 b are the accelerating curves from walking tumble action sequence in sample set used by this embodiment of the invention
Figure；
Fig. 2 c are from running to the accelerating curve of tumble action sequence in sample set used by this embodiment of the invention
Figure；
Fig. 2 d are in sample set used by this embodiment of the invention from the accelerating curve squatted down to tumble action sequence
Figure；
Fig. 3 is the acceleration plots of a continuous action sequence in sample set used by this embodiment of the invention.
Specific implementation mode
The present invention will be further described with reference to the accompanying drawings.
The present invention proposes a kind of human posture recognition method, and Fig. 1 show in the human posture recognition method and identified
The flow chart of journey (i.e. following steps (1)-step (6)).Three sides of human body of this method based on six axle sensor institute continuous acquisitions
Upward linear acceleration and angular acceleration data passes through data Kalman filtering, the extraction of data statistical characteristics amount, decision tree point
Structure, posture synthesis output of class model etc. carry out comprehensive analysis, realize the identification of each posture of human body, especially tumble posture
Identification.The human posture recognition method includes the following steps：
(1) space X, the linear acceleration and angular acceleration data on tri- directions Y, Z that six axle sensors are acquired are used,
6 class data altogether pass through Bluetooth transmission to host computer.The host computer refers to receiving the background terminal of data.
In the present embodiment, which uses JY61 modules.Its range is：Linear acceleration is ± 16g, angular acceleration
For ± 2000 °/s, angle is ± 180 °.Measurement error is：Linear acceleration 0.01g, 0.05 °/s of angular acceleration.Frequency acquisition is
10 times/s.Six axle sensor is different from widely used three-axis sensor, being capable of integrated tri-axial acceleration meter and three axis tops
The function of spiral shell instrument, while collecting linear acceleration data and angular acceleration data on three directions, altogether six class data.
Above-mentioned X, Y, Z-direction are mutually perpendicular to, respectively horizontal front-rear direction, horizontal left and right directions and vertical direction, here
It is front and back and left and right be using human body to be measured as reference.
(2) Kalman filtering processing is carried out in real time to the six class gathered datas for being transmitted to host computer respectively, after being filtered
Data and be stored in backstage.
The kalman filter method used in the present embodiment is a kind of method of data fusion, combines common acceleration
The advantages of sensor and gyroscope, the attitude angle obtained by integral operation in conjunction with the angle and gyroscope that linear acceleration is found out
Degree calculates the weight of the two according to statistical law, merges out accurate state value automatically.
(3) data characteristics that analysis Various types of data is showed, (t is more than 1 integer) a data per class data decimation t
Statistical nature, a shared t*6 data characteristics, as the conditional attribute for differentiating human body attitude.
(4) data statistical characteristics for extracting the data set under all kinds of postures, constitute sample set.By the sample in sample set point
For training sample and test sample.
In the present embodiment, as shown in Fig. 2 a- Fig. 2 d, which equally also comes from tumble and non-tumble action sequence,
Include from standing to falling, from walking to falling, from running to falling, from squatting down to falling multiple action sequences.(Fig. 2 a-
In Fig. 2 d, " X acceleration ", " Y acceleration " and " Z acceleration " refers respectively to the linear acceleration of X-direction, Y-direction and Z-direction；
" X angular acceleration ", " Y angular acceleration " and " Z angular acceleration " refers respectively to the angular acceleration of X-direction, Y-direction and Z-direction.)
As shown in Figure 3.One sample set comes from a continuous action sequence, include on foot, sit elevator, on
Building, downstairs, running, squat etc. multiple postures.Each posture extracts ten samples as training sample from an action sequence, from
Ten are randomly choosed in remaining sample in sample set is used as test sample.
(5) Decision-Tree Classifier Model is promoted according to above-mentioned training sample and test sample structure gradient.
Gradient employed in the present embodiment promotes Decision-Tree Classifier Model based on a kind of decision Tree algorithms of iteration, by more
Classification regression tree (CART) form.Boosted tree is that all decision trees carry out Shared Decision Making, and letter is lost when using square error
When number, the approximation method declined using steepest utilizes the negative gradient of loss function in the value of "current" model, as regression problem
The approximation of the middle residual error for promoting tree algorithm, is fitted a regression tree.Boosted tree is to utilize addition model and forward direction Distribution Algorithm
Realize the optimization process of study.
(6) data to be detected acquired are handled according to above-mentioned steps (1) to (3) based on the Decision-Tree Classifier Model
(data are different from the above-mentioned data for training and test decision tree classification model, which is freshly harvested to be detected afterwards
Data, also according to step (1) to (3) acquire and handle) real-time grading, identify tumble posture and other movement postures.
Kalman filter method employed in step (2) not only needs to be used in training pattern in step (5), to reality
The data acquired in equally first carry out the filtering, it can remove the shadow of noise and interference in sensing system
It rings, restores time of day data.
In this example, it is assumed that it is related that human motion, which is with the state at t-1 moment in the state of t moment, is below
To the modeling process of human motion system, the model created is as follows：
xt=Ftxt-1+Btut+wt (1)
Wherein：xtIt is one and represents human body kinematic system in the time of day vector of t moment, which includes speed
The information such as degree；
FtIt is the state transition matrix of t moment, the system mode that represent the t-1 moment is to the moment (i.e. t moment)
The influence of system state；
utIt is the control input vector of t moment, for example whether turning round；
BtIt is the input transition matrix of t moment, that is, has quantified to input the influence to state；
wtIt is the process noise of t moment, it will be assumed that the noise is obeyed the multivariate normal that mean value is 0 and is distributed, association side
Poor matrix is labeled as Qt。
The error of measured value is indicated with the following formula：
zt=Htxt+vt (2)
Wherein：ztThe human body kinematic system is represented in the observed value vector of t moment, which includes the result measured；
HtIt is the measures conversion matrix of t moment, this matrix indicates the relationship between state vector and observed value vector；
vtIt is the calculation matrix of t moment, similar procedure noise, it is assumed that the error obeys the multivariate normal point that mean value is 0
Cloth, covariance matrix are labeled as Rt。
Assuming that required state obeys a normal distribution, i.e. x~N (μ, Σ), Kalman filter is by an iteration
Process solve known Ft、ut、Bt、zt、xt-1、HtAnd wtAnd vtIn the case of obeying known distribution, x is solvedtThe problem of.It is each
Secondary iteration includes two steps, is prediction and observation respectively：Prediction steps state (can use a normal distribution description) mean value and
Covariance is worth update according to the observation according to the model modification of system in the mean value and covariance of observation of steps state.
In step (3), the t data statistics per class data is characterized in based on being presented to the data under different human body posture
The different fluctuation characteristics that go out and determination.
In the present embodiment, the data fluctuations degree under different human body posture is of different sizes, thus chooses 3 data systems
Feature is counted, i.e. t=3, this 3 data statistical characteristics are respectively variance, maximum value and minimum value.
In step (4) when building sample set, uses and take n consecutive numbers strong point per class data using the time as axis, take every class
T data statistics characteristic value of data, altogether n*t data feature values constitute a sample.Wherein, it is divided into d between specimen sample
A consecutive numbers strong point sampling time (first data point of i.e. first sample and first data point of second sample it
Between be separated by (d-1) a data point), i.e., be divided into d*0.1s between specimen sample.Wherein, the n, t, d are positive integer, and n is more than
d。
In this example, a consecutive numbers strong points 15 (i.e. n=15) are taken per class data using the time as axis, take respectively 15 continuously
3 data feature values of data point, altogether 18 data feature values constitute a sample.Wherein, 10 are divided between specimen sample (i.e.
D=10) the sampling time at a consecutive numbers strong point is 1s.
In the present embodiment, it is contemplated that under the sampling interval of 1s, the coincidence factor of data should not be excessively high between sample, no
It can then cause two sample data characteristic values about the same, sample is caused to repeat.Therefore in each sample consecutive numbers strong point
Number is selected as 15,15 consecutive numbers strong points and so that the coincidence factor of the initial data between sample is 1/3, and data volume is also unlikely to
It is difficult to reflect the motion characteristic of single posture very little.
The Decision-Tree Classifier Model built in above-mentioned steps (5) promotes Decision-Tree Classifier Model using gradient, builds gradient
Decision-Tree Classifier Model is promoted to include the following steps：
A1：Initialization gradient promotes Decision-Tree Classifier Model, and (the targeted model of the initialization is to select from the prior art
The gradient selected promotes Decision-Tree Classifier Model), estimation makes the constant value of loss function minimization, the gradient promote decision tree classification
Model be only there are one root node tree；
A2：The negative gradient of counting loss function "current" model value, using it as the estimation of residual error；
A3：Estimation returns leaf nodes region, with the approximation of regression criterion；
A4：Using the value in linear search estimation leaf node region, make loss function minimization；
A5：Update regression tree.
Wherein, loss function refers to the function of residual error between quantitative prediction value and actual value.Residual error=actual value-prediction
Value.
In the present embodiment, after the above-mentioned steps (6) of above-mentioned human posture recognition method identify human body attitude, also into
The output of row recognition result.Here, in the identification of tumble posture and the identification of other movement postures, tumble posture has highest excellent
First grade, once this is identified as a result, exporting the result at once.And for other postures, then use certain logic synthesis defeated
Go out recognition result, repeated and redundant is avoided to export.In the present embodiment, specially：
Using the recognition result of a sample as one small recognition unit (the first recognition result), it is with 4 continuous samples
60 data points are one and identify that big unit (the second recognition result) exports the recognition result in 4 seconds.
According to fixed recognition logic, using the recognition result after logic judgment in 4 seconds as final recognition result.
The recognition logic is：
Identify in subsection " tumble " occur as a result, directly the recognition result of output " tumble " is known as final when this 4
Other result；
When the result that this 4 identify subsection is unanimously and not " tumble ", that is, the result is exported as final identification knot
Fruit；
Identify that the result of subsection is inconsistent when this 4, the recognition result for selecting occurrence number most is as final identification
As a result；When appearance 2:When 2 the case where, select the result of the first two identification subsection as final recognition result.
It should be as a result, being convenient for subsequent this ensure that can accurately and timely be exported when the dangerous sexual act such as occurring falling
Supervision and emergency processing.For other postures, 4 seconds primary output frequencies can promote whole gesture recognition accuracy, rule
Keep away the interference of abnormal posture.
Decision-Tree Classifier Model is promoted for gradient, the method for using Feature Selection makes the performance of model protrude,
Those a small amount of features larger to the performance boost of model can be rapidly found, rejects and promotes little redundancy feature combination,
Decision-Tree Classifier Model is set to be simplified.In this example, Feature Selection is the screening to 18 characteristic values, before choosing successively
20%, preceding 40%, preceding 60%, preceding 80% feature formation test sample and training sample, calculate separately model accuracy, finding makes
The highest feature combination of accuracy of Decision-Tree Classifier Model, is combined as best eigenvalue.
In the build process that gradient promotes Decision-Tree Classifier Model, algorithm model is realized using the method for cross validation
Verification.The method of the cross validation is that whole sample sets that training sample set and test sample collection are formed are divided into z parts
(z>2, z be integer), take a copy of it as test sample collection every time, others are used as training sample set.By multiple authentication
Obtain the overall accuracy of model.This method is mutual exclusion between examining used test sample collection every time, can be ensured every
One data available is all authenticated by model.In the present embodiment, the method for selecting 5 foldings (i.e. z=5) cross validation, by sample
5 parts are equally divided into, verification calculating is carried out respectively as test sample and training sample.
The invention also provides a kind of human body attitude identifying system for realizing above-mentioned human posture recognition method, the system packets
Include data detector and controller.Wherein, the data detector is at least acquiring human body on space X, tri- directions Y, Z
Six class parameter of linear acceleration and angular acceleration, X, Y, Z-direction are the direction of setting, and the parameter of acquisition is sent to controller；
The parameter that the controller is acquired based on data detector executes above-mentioned human posture recognition method.In the present embodiment, Data Detection
Device is six-axle acceleration sensor, and the parameter of acquisition is sent to controller by way of bluetooth；Controller is host computer, should
Host computer is background terminal.
This hair can be understood and applied the above description of the embodiments is intended to facilitate those skilled in the art
It is bright.Person skilled in the art obviously easily can make various modifications to these embodiments, and described herein
General Principle is applied in other embodiment without having to go through creative labor.Therefore, the present invention is not limited to implementations here
Example, those skilled in the art's announcement according to the present invention, improvement and modification made without departing from the scope of the present invention all should be
Within protection scope of the present invention.
Claims (10)
1. a kind of human posture recognition method, it is characterised in that：Include the following steps：
(1) linear acceleration and angular acceleration six class parameter of the human body on space X, tri- directions Y, Z is at least acquired；The X, Y,
Z-direction is the direction of setting；
(2) data characteristics that all kinds of parameters are presented is analyzed, for per class t data statistical characteristics of data decimation, as sentencing
The conditional attribute of disconnected human body attitude；Wherein, t is the integer more than 1；
(3) data statistical characteristics for extracting the data set under all kinds of human body attitudes, constitute sample set, the sample set are divided into instruction
Practice sample and test sample；
(4) according to the training sample and the test sample, Decision-Tree Classifier Model is built；
(5) it is based on the Decision-Tree Classifier Model, real-time grading is carried out to the data acquired, identifies human body attitude；
The data that are acquired described in the step (5) is acquire and handle according to the step (1) to the step (2)
Data to be identified.
2. human posture recognition method according to claim 1, it is characterised in that：The human body attitude is tumble posture.
3. human posture recognition method according to claim 1, it is characterised in that：The X, Y, Z-direction are mutually perpendicular to；
Preferably, the X, Y, Z-direction are respectively the horizontal left and right directions and perpendicular using human body to be measured as the horizontal front-rear direction of reference
Histogram to；
Preferably, the step (1) acquires human parameters using six axle sensors；
It is further preferred that the function of the six axle sensors integrated tri-axial acceleration meter and three-axis gyroscope.
4. human posture recognition method according to claim 1, it is characterised in that：It, will be described before the step (2)
At least six class parameters of step (1) acquisition are filtered；
Preferably, described to be filtered using kalman filter method.
5. human posture recognition method according to claim 1, it is characterised in that：It is each anthropoid in the step (3)
Posture includes coming from the human body attitude of action sequence；
Preferably, the action sequence include from standing to tumble, and/or from walking to tumble, and/or from running to tumble,
And/or from squatting down tumble；
Preferably, the action sequence further include by walk, sit elevator, upstairs, downstairs, the continuous action sequence that constitutes of crouching；
It is further preferred that 10 samples are extracted from the action sequence as training sample for each human body attitude, from
Select 10 samples as test sample in remaining sample in the action sequence.
6. human posture recognition method according to claim 1, it is characterised in that：All kinds of people of extraction in the step (3)
The data statistical characteristics of data set under body posture, constituting sample set includes：It is that every class parameter chooses n continuously by axis of the time
Data point, take t data statistics characteristic value of every class parameter, n*t characteristic value constitutes a sample altogether；Specimen sample
Time interval is the sampling time at d consecutive numbers strong point；
Wherein, n, t, d are positive integer, and n ＞ d；
Preferably, the sampling time at consecutive numbers strong point is 0.1s.
7. human posture recognition method according to claim 1, it is characterised in that：The decision built in the step (4)
Tree classification model is that gradient promotes Decision-Tree Classifier Model；
Preferably, structure Decision-Tree Classifier Model includes the following steps in the step (4)：
(41) initialization gradient promotes Decision-Tree Classifier Model, and estimation makes the constant value of loss function minimization, after initialization
It is the tree only there are one root node that gradient, which promotes Decision-Tree Classifier Model,；
(42) value that gradient of the negative gradient of the loss function after the initialization promotes Decision-Tree Classifier Model is calculated, is made
For the estimation of residual error；
(43) estimation returns leaf nodes region, with the approximation of regression criterion；
(44) using the value for linearly searching element estimation leaf node region, make loss function minimization；
(45) regression tree is updated；
It is further preferred that the loss function refers to the function of residual error between quantitative prediction value and actual value；
It is further preferred that the residual error subtracts predicted value equal to actual value.
8. human posture recognition method according to claim 2, it is characterised in that：The human posture recognition method is in institute
Stating step (5) further includes later：Export the human body attitude of the step (5) identification；
Preferably, it includes when falling, to export tumble signal to recognize human body attitude；
Preferably, the human posture recognition method further includes after the step (5)：In the human body attitude recognized, people
Body posture is the output priority highest fallen；Human body attitude other than tumble is exported according to the logic synthesis for excluding redundancy；
It is further preferred that the logic for excluding redundancy is：
Using the recognition result of a sample as the first recognition result, using the recognition result of l continuous sample as the second identification
As a result, the recognition result based on l continuous sample, exports final recognition result in accordance with the following methods：
When l the first recognition results include the human body attitude fallen, output human body attitude is to fall to tie as final identification
Fruit；Otherwise,
When the l the first recognition results are consistent, output human body attitude is that the consistent recognition result is tied as final identification
Fruit；
When the l the first recognition results are inconsistent, the most recognition result of output occurrence number；It is tied when there is the identification of m kinds
Fruit, and when the number that occurs of each recognition result is equal, the recognition result that occurs at first is exported as final recognition result.
9. human posture recognition method according to claim 1, it is characterised in that：Decision is built in the step (5)
When tree classification model, using the method reduced decision tree disaggregated model of Feature Selection, described in the method validation using cross validation
Decision-Tree Classifier Model.
10. a kind of human body attitude identifying system for realizing the human posture recognition method described in any one of claim 1-9,
It is characterized in that：Including data detector and controller；The data detector is at least acquisition human body in space X, Y, Z tri-
Six class parameter of linear acceleration and angular acceleration on direction；The X, Y, Z-direction are the direction of setting, and the parameter of acquisition are sent out
Give the controller；
The controller parameter that detector acquires based on the data, executes the human posture recognition method；
Preferably, the parameter of acquisition is sent to the controller by the data detector by way of bluetooth；
Preferably, the data detector is six-axle acceleration sensor；
Preferably, the controller is host computer；
It is further preferred that the host computer is background terminal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810269770.4A CN108509897A (en) | 2018-03-29 | 2018-03-29 | A kind of human posture recognition method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810269770.4A CN108509897A (en) | 2018-03-29 | 2018-03-29 | A kind of human posture recognition method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108509897A true CN108509897A (en) | 2018-09-07 |
Family
ID=63379159
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810269770.4A Pending CN108509897A (en) | 2018-03-29 | 2018-03-29 | A kind of human posture recognition method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108509897A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109118523A (en) * | 2018-09-20 | 2019-01-01 | 电子科技大学 | A kind of tracking image target method based on YOLO |
| CN109508698A (en) * | 2018-12-19 | 2019-03-22 | 中山大学 | A kind of Human bodys' response method based on binary tree |
| CN109584521A (en) * | 2018-10-15 | 2019-04-05 | 西安理工大学 | A kind of incorrect sitting-pose monitoring method based on Notch sensor |
| CN110008998A (en) * | 2018-11-27 | 2019-07-12 | 美律电子(深圳)有限公司 | Label data generating system and method |
| CN110032932A (en) * | 2019-03-07 | 2019-07-19 | 哈尔滨理工大学 | A kind of human posture recognition method based on video processing and decision tree given threshold |
| CN110275161A (en) * | 2019-06-28 | 2019-09-24 | 台州睿联科技有限公司 | A kind of wireless human body gesture recognition method applied to Intelligent bathroom |
| CN110476879A (en) * | 2019-08-26 | 2019-11-22 | 重庆邮电大学 | Milk cow behavior discriminant classification method and device based on multi-tag chain type ecological environment |
| CN110674683A (en) * | 2019-08-15 | 2020-01-10 | 深圳供电局有限公司 | Robot hand motion recognition method and system |
| CN110988861A (en) * | 2019-10-31 | 2020-04-10 | 复旦大学 | Human body posture recognition system based on millimeter wave radar |
| CN111401507A (en) * | 2020-03-12 | 2020-07-10 | 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 |
| CN111401435A (en) * | 2020-03-13 | 2020-07-10 | 安徽工业大学 | Human body motion mode identification method based on motion bracelet |
| CN111767888A (en) * | 2020-07-08 | 2020-10-13 | 北京澎思科技有限公司 | Object state detection method, computer device, storage medium, and electronic device |
| CN114831627A (en) * | 2022-03-17 | 2022-08-02 | 吉林大学 | Lower limb prosthesis movement identification method based on three decision trees |
| CN115601505A (en) * | 2022-11-07 | 2023-01-13 | 广州趣丸网络科技有限公司（Cn） | Human body three-dimensional posture restoration method and device, electronic equipment and storage medium |
| CN115861871A (en) * | 2022-11-10 | 2023-03-28 | 深圳蓄能发电有限公司 | Multiple verification detection device, method and medium for human body posture detection of field personnel |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102772211A (en) * | 2012-08-08 | 2012-11-14 | 中山大学 | Human movement state detection system and detection method |
| US20130023798A1 (en) * | 2011-07-20 | 2013-01-24 | Intel-Ge Care Innovations Llc | Method for body-worn sensor based prospective evaluation of falls risk in community-dwelling elderly adults |
| CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
| CN107220617A (en) * | 2017-05-25 | 2017-09-29 | 哈尔滨工业大学 | Human body attitude identifying system and method |
| US20170352240A1 (en) * | 2016-06-03 | 2017-12-07 | John Carlton-Foss | Method and system for motion analysis and fall prevention |
-
2018
- 2018-03-29 CN CN201810269770.4A patent/CN108509897A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130023798A1 (en) * | 2011-07-20 | 2013-01-24 | Intel-Ge Care Innovations Llc | Method for body-worn sensor based prospective evaluation of falls risk in community-dwelling elderly adults |
| CN102772211A (en) * | 2012-08-08 | 2012-11-14 | 中山大学 | Human movement state detection system and detection method |
| CN103211599A (en) * | 2013-05-13 | 2013-07-24 | 桂林电子科技大学 | Method and device for monitoring tumble |
| US20170352240A1 (en) * | 2016-06-03 | 2017-12-07 | John Carlton-Foss | Method and system for motion analysis and fall prevention |
| CN107220617A (en) * | 2017-05-25 | 2017-09-29 | 哈尔滨工业大学 | Human body attitude identifying system and method |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109118523A (en) * | 2018-09-20 | 2019-01-01 | 电子科技大学 | A kind of tracking image target method based on YOLO |
| CN109118523B (en) * | 2018-09-20 | 2022-04-22 | 电子科技大学 | Image target tracking method based on YOLO |
| CN109584521A (en) * | 2018-10-15 | 2019-04-05 | 西安理工大学 | A kind of incorrect sitting-pose monitoring method based on Notch sensor |
| CN110008998A (en) * | 2018-11-27 | 2019-07-12 | 美律电子(深圳)有限公司 | Label data generating system and method |
| CN110008998B (en) * | 2018-11-27 | 2021-07-13 | 美律电子(深圳)有限公司 | Label data generating system and method |
| CN109508698B (en) * | 2018-12-19 | 2023-01-10 | 中山大学 | Human behavior recognition method based on binary tree |
| CN109508698A (en) * | 2018-12-19 | 2019-03-22 | 中山大学 | A kind of Human bodys' response method based on binary tree |
| CN110032932A (en) * | 2019-03-07 | 2019-07-19 | 哈尔滨理工大学 | A kind of human posture recognition method based on video processing and decision tree given threshold |
| CN110032932B (en) * | 2019-03-07 | 2021-09-21 | 哈尔滨理工大学 | Human body posture identification method based on video processing and decision tree set threshold |
| CN110275161A (en) * | 2019-06-28 | 2019-09-24 | 台州睿联科技有限公司 | A kind of wireless human body gesture recognition method applied to Intelligent bathroom |
| CN110674683A (en) * | 2019-08-15 | 2020-01-10 | 深圳供电局有限公司 | Robot hand motion recognition method and system |
| CN110674683B (en) * | 2019-08-15 | 2022-07-22 | 深圳供电局有限公司 | Robot hand motion recognition method and system |
| CN110476879A (en) * | 2019-08-26 | 2019-11-22 | 重庆邮电大学 | Milk cow behavior discriminant classification method and device based on multi-tag chain type ecological environment |
| CN110988861B (en) * | 2019-10-31 | 2022-09-16 | 复旦大学 | Human body posture recognition system based on millimeter wave radar |
| CN110988861A (en) * | 2019-10-31 | 2020-04-10 | 复旦大学 | Human body posture recognition system based on millimeter wave radar |
| 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 |
| CN111401507A (en) * | 2020-03-12 | 2020-07-10 | 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 |
| CN111401435A (en) * | 2020-03-13 | 2020-07-10 | 安徽工业大学 | Human body motion mode identification method based on motion bracelet |
| CN111401435B (en) * | 2020-03-13 | 2023-04-07 | 安徽工业大学 | Human body motion mode identification method based on motion bracelet |
| CN111767888A (en) * | 2020-07-08 | 2020-10-13 | 北京澎思科技有限公司 | Object state detection method, computer device, storage medium, and electronic device |
| CN114831627A (en) * | 2022-03-17 | 2022-08-02 | 吉林大学 | Lower limb prosthesis movement identification method based on three decision trees |
| CN115601505A (en) * | 2022-11-07 | 2023-01-13 | 广州趣丸网络科技有限公司（Cn） | Human body three-dimensional posture restoration method and device, electronic equipment and storage medium |
| CN115601505B (en) * | 2022-11-07 | 2023-03-14 | 广州趣丸网络科技有限公司 | Human body three-dimensional posture restoration method and device, electronic equipment and storage medium |
| CN115861871A (en) * | 2022-11-10 | 2023-03-28 | 深圳蓄能发电有限公司 | Multiple verification detection device, method and medium for human body posture detection of field personnel |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108509897A (en) | A kind of human posture recognition method and system | |
| CN109522793B (en) | Method for detecting and identifying abnormal behaviors of multiple persons based on machine vision | |
| CN109460702B (en) | Passenger abnormal behavior identification method based on human body skeleton sequence | |
| CN110425005B (en) | Safety monitoring and early warning method for man-machine interaction behavior of belt transport personnel under mine | |
| CN103955699B (en) | A kind of real-time fall events detection method based on monitor video | |
| CN101558996B (en) | Gait recognition method based on orthogonal projection three-dimensional reconstruction of human motion structure | |
| CN109522853B (en) | Face datection and searching method towards monitor video | |
| CN105279483A (en) | Fall-down behavior real-time detection method based on depth image | |
| CN101867699B (en) | Real-time tracking method of nonspecific target based on partitioning | |
| CN108446733A (en) | A kind of human body behavior monitoring and intelligent identification Method based on multi-sensor data | |
| CN104063719A (en) | Method and device for pedestrian detection based on depth convolutional network | |
| CN102302370A (en) | Method and device for detecting tumbling | |
| CN110490080A (en) | A kind of human body tumble method of discrimination based on image | |
| CN103345626A (en) | Intelligent wheelchair static gesture identification method | |
| CN109325546A (en) | A kind of combination footwork feature at time footprint recognition method | |
| WO2023035093A1 (en) | Inertial sensor-based human body behaviour recognition method | |
| CN108537181A (en) | A kind of gait recognition method based on the study of big spacing depth measure | |
| CN104586402B (en) | A kind of feature extracting method of physical activity | |
| CN109902564A (en) | A kind of accident detection method based on the sparse autoencoder network of structural similarity | |
| CN105139029A (en) | Activity recognition method and activity recognition device for persons serving sentences | |
| CN107358248B (en) | Method for improving falling detection system precision | |
| CN103699874A (en) | Crowd abnormal behavior identification method based on SURF (Speed-Up Robust Feature) stream and LLE (Locally Linear Embedding) sparse representation | |
| CN108717548B (en) | Behavior recognition model updating method and system for dynamic increase of sensors | |
| CN104200218A (en) | Cross-view-angle action identification method and system based on time sequence information | |
| Ning et al. | Fall detection algorithm based on gradient boosting decision tree |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180907 |
|
| RJ01 | Rejection of invention patent application after publication |