CN111710129B - Real-time pre-collision falling detection method for old people - Google Patents
Real-time pre-collision falling detection method for old people Download PDFInfo
- Publication number
- CN111710129B CN111710129B CN202010537951.8A CN202010537951A CN111710129B CN 111710129 B CN111710129 B CN 111710129B CN 202010537951 A CN202010537951 A CN 202010537951A CN 111710129 B CN111710129 B CN 111710129B
- Authority
- CN
- China
- Prior art keywords
- angular velocity
- triaxial
- data set
- acceleration
- model
- 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.)
- Expired - Fee Related
Links
Images
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0407—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis
- G08B21/043—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons based on behaviour analysis detecting an emergency event, e.g. a fall
-
- G—PHYSICS
- 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
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/04—Alarms for ensuring the safety of persons responsive to non-activity, e.g. of elderly persons
- G08B21/0438—Sensor means for detecting
- G08B21/0446—Sensor means for detecting worn on the body to detect changes of posture, e.g. a fall, inclination, acceleration, gait
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B29/00—Checking or monitoring of signalling or alarm systems; Prevention or correction of operating errors, e.g. preventing unauthorised operation
- G08B29/18—Prevention or correction of operating errors
- G08B29/185—Signal analysis techniques for reducing or preventing false alarms or for enhancing the reliability of the system
Landscapes
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Emergency Management (AREA)
- Business, Economics & Management (AREA)
- Gerontology & Geriatric Medicine (AREA)
- General Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Security & Cryptography (AREA)
- Psychiatry (AREA)
- Psychology (AREA)
- Social Psychology (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a real-time before-collision falling detection method for old people, which comprises the steps of wearing an inertial sensor for the old people in real time, collecting a human body activity data set in the human body activity process in real time, filtering through a filter, screening the filtered data through a threshold value method to obtain approximate falling and falling action data, inputting the data into a before-collision falling detection network model as a training sample after marking, and adjusting model parameters to obtain an optimized network model; filtering and filtering the data of the accelerometer and the gyroscope collected in real time, and inputting the filtered and filtered data into an optimized before-collision falling detection network model to obtain a real-time before-collision falling detection result; the invention solves the problems that the machine learning algorithm influences the classification effect on the deviation of a large amount of non-falling data, and false reports and missing reports are generated.
Description
Technical Field
The invention relates to the technical field of behavior detection of old people, in particular to a real-time falling-down detection method before collision for the old people.
Background
With age, the lower limbs of the elderly become weaker, the physical coordination is poor, and the risk of falling is increased. In real life, most falls are caused by the unbalancing of the human body caused by unexpected slips or stumbling. Fall detection helps to discover and take protective measures in a timely manner. Most of the existing fall detection detects the occurrence of a fall event after the fall occurs. However, in order to prevent or reduce the injury to the human body caused by falling, it is necessary to perform fall detection before the fall collision occurs, that is, to perform fall detection before collision in real time, and to take corresponding fall protection measures in time to reduce the physical injury caused by falling.
The existing fall detection method before collision mainly comprises a threshold-based method and a machine learning-based method. The threshold-based method uses inertial sensor data to construct artificial features, and judges whether a fall is about to occur in real time according to a set threshold. The method based on machine learning, such as artificial neural network and support vector machine, achieves the purpose of detecting falling in advance by training a machine learning model by using a sample data set.
The paper "Evaluation of Inertial Sensor-Based Pre-Impact surface Detection Using Public data set" proposes a threshold-Based Pre-Impact tumble Detection method, which first performs noise reduction on data acquired by an Inertial Sensor, and judges the occurrence of a tumble event by two threshold methods, namely a Vertical Angle (VA) and a Triangular Feature (TF). Both algorithms can detect falls 100% accurately on the public dataset SisFall, but the specificity of both methods is 78.3% and 83.9%, respectively, with more false positives.
The chinese patent CN106874847A "artificial neural network-based fall prediction method and fall airbag protection device" provides an artificial neural network-based fall prediction method and fall airbag protection device, which collects multiple groups of human motion sample data in advance to train the artificial neural network, thereby calculating each dimension data classification result for the input multidimensional inertial sensor data, and performing weighted summation to perform fall prediction.
The current fall detection algorithm before collision is mainly realized by a threshold value method and a simple machine learning algorithm, wherein the threshold value method can obtain higher precision on experimental data, but has poor specificity on a public data set and more false alarm conditions. Although the machine learning algorithm has better universality compared with a threshold value method, the method directly trains sample data, and due to unbalanced samples, part of information difference between falling and non-falling can be lost, so that the classification effect can be influenced on the bias of a large amount of non-falling data, and false reports are generated.
Disclosure of Invention
Aiming at the defects in the prior art, the method for detecting the falling of the old people before the collision in real time solves the problems that the machine learning algorithm can influence the classification effect on the deviation of a large amount of non-falling data, and false alarm and missing alarm are generated.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that: a real-time pre-collision falling detection method for old people comprises the following steps:
s1, wearing an inertial sensor for the old, and acquiring data of an accelerometer and a gyroscope in the human body activity process in real time to obtain an original triaxial acceleration data set and a triaxial angular velocity data set;
s2, filtering the original triaxial acceleration data set and the triaxial angular velocity data set through a fourth-order IIR Butterworth low-pass filter to obtain a standard triaxial acceleration data set and a triaxial angular velocity data set;
s3, filtering the standard triaxial acceleration data set and the triaxial angular velocity data set through a threshold method to obtain a triaxial acceleration data set and a triaxial angular velocity data set which are suspected to fall;
s4, labeling the triaxial acceleration data set and the triaxial angular velocity data set suspected to fall, using the labeled triaxial acceleration data set and the labeled triaxial angular velocity data set as training samples, inputting the labeled training samples into the fall detection network model before collision, and adjusting model parameters to obtain an optimized fall detection network model before collision;
and S5, filtering the accelerometer and gyroscope data acquired in real time through the step S2 and the step S3, and inputting the data into the optimized before-collision falling detection network model to obtain a real-time before-collision falling detection result.
Further, the step S3 includes the following sub-steps:
s31, establishing a threshold judgment model of a threshold method, wherein the threshold judgment model comprises: a body offset angle model and an acceleration projection area model;
s32, inputting the standard triaxial acceleration data and triaxial angular velocity data into a threshold judgment model;
s33, calculating total acceleration vector ACC through a body deviation angle modelSVMTotal vector of angular velocity ωSVMAngle Deg to left and right from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωSVMAngle Deg to left and right from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalWhether the current triaxial acceleration data and the current triaxial angular velocity data exceed respective thresholds or not is judged, if yes, the current triaxial acceleration data and the current triaxial angular velocity data are reserved, and the step S35 is skipped, and if not, the step S34 is skipped;
s34, calculating total acceleration vector ACC through an acceleration projection area modelSVMTotal vector of angular velocity ωSVMAnd the human inclination vector PSVMAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωSVMAnd the human inclination vector PSVMIf the current triaxial acceleration data and the current triaxial angular velocity data exceed the respective threshold values, keeping the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35, otherwise, discarding the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35;
s35, judging whether all data in the standard triaxial acceleration data set and the triaxial angular velocity data set are input into the threshold judgment model, if so, jumping to the step S36, and if not, jumping to the step S32;
and S36, forming a triaxial acceleration data set and a triaxial angular velocity data set of the suspected fall from all the reserved triaxial acceleration data and triaxial angular velocity data.
The beneficial effects of the above further scheme are: the body offset angle model and the acceleration projection area model calculate the angle change of the human action, in the falling action, the larger angle change is one of the most obvious characteristics, and the algorithm adopting the angle has higher specificity. The body deviation angle model is considered as a vertical axis and is unchanged, the deviation angle of a person is changed, the acceleration projection area model is considered as a man-made center, the inclination angle change of the person is calculated through three-axis acceleration, the three-axis acceleration and the three-axis acceleration are combined to consider more conditions, the missing report is reduced, for example, when the body of the person is bent such as sitting, the falling detection capacity of the body deviation angle model is stronger, and when the body of the person is not bent such as walking and running, the falling detection capacity of the acceleration projection area model is stronger.
Further, the body offset angle model in step S31 is:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
Further, in step S31, the acceleration projection area model is:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the Z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
Further, the fall detection network model before collision in step S4 includes: an input layer, a hidden layer and an output layer; the input layer, the hidden layer and the output layer are connected in sequence.
Further, the relational expressions of the input layer, the hidden layer and the output layer are as follows:
Hβ=T*
where H is the output of the hidden layer node, β is the output weight of the hidden layer, T*Is a predictive tag.
Further, the process of calculating the output weight β of the hidden layer in step S4 is as follows:
a1, establishing a weight optimization model:
rl=|log2pl|,l=1,…,m
wherein L is the number of hidden layer nodes, C is a hyper-parameter, m is the total number of classification categories, rlIs a weight of class I,. epsilonlIs an error vector of class I, betaiIs the ith weight in the output weight beta, i.e. the output weight of the ith hidden node, i is more than or equal to 1 and less than or equal to L, G (a)i,bi,xj) Is input xjOutput of the corresponding i-th hidden node, aiAs input weights, biBias for the ith hidden node, xjFor the jth input training sample, xj∈RdD is the dimension of the input training sample, tjIs the jth sample xjCorresponding true value label vector, tj∈RmM is the dimension of the label vector,. epsilonjFor the jth input training sampleError vector of the book, plThe proportion of the l-th sample in all training samples is U-th.
A2, converting the weight optimization model into a Lagrangian function:
a3, obtaining the output weight beta of the hidden layer according to the Lagrange function:
wherein T' is a real label.
The beneficial effects of the above further scheme are: each class is assigned with a weight, so that the problem of unbalanced data samples is solved, the weight of falling samples which occupy less data samples can be relatively increased, and the unbalanced influence is reduced. Also, the training time can be significantly reduced compared to the method of assigning a weight to each sample.
In conclusion, the beneficial effects of the invention are as follows:
(1) and a threshold method is adopted to quickly filter a large amount of daily activity sample data, retain the sample data of real falling and approximate falling, and reduce the subsequent model training time.
(2) The problem that real falling samples and approximate falling samples are unbalanced is considered, the weight is set for each sample according to the proportion of each sample, all falling samples can be guaranteed to be detected, meanwhile, the false alarm rate is greatly reduced, and the overall identification precision is improved.
(3) The output weight beta of the hidden layer is determined at one time through an equation set without iterative adjustment, the operation amount of the falling detection network model before collision is reduced, the generalization performance of the model is improved, and the model operation is efficient and quick.
Drawings
Fig. 1 is a flowchart of a method for detecting falling before collision of an elderly person in real time.
Fig. 2 is a schematic structural diagram of a fall detection network model before collision.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
As shown in fig. 1, a method for detecting falling before collision of an elderly person in real time includes the following steps:
s1, wearing an inertial sensor for the old, and acquiring data of an accelerometer and a gyroscope in the human body activity process in real time to obtain an original triaxial acceleration data set and a triaxial angular velocity data set;
s2, filtering the original triaxial acceleration data set and the triaxial angular velocity data set through a fourth-order IIR Butterworth low-pass filter to obtain a standard triaxial acceleration data set and a triaxial angular velocity data set;
in the embodiment, a fourth-order IIR Butterworth low-pass filter with the cut-off frequency of 5hz is adopted for preprocessing, so that noise caused by more noise data and unknown abnormal conditions existing in the acquired inertial sensor data due to the wearing stability and the uncertainty of the activity of the wearer is eliminated.
S3, filtering the standard triaxial acceleration data set and the triaxial angular velocity data set through a threshold method to obtain a triaxial acceleration data set and a triaxial angular velocity data set which are suspected to fall;
because the standard triaxial acceleration data set and the triaxial angular velocity data set contain a large amount of daily living Activity (ADL) data and a small amount of falling data, the standard triaxial acceleration data set and the triaxial angular velocity data set are filtered by a threshold method to obtain real falling data and approximate falling data (such as going up and down stairs and the like), namely the triaxial acceleration data set and the triaxial angular velocity data set which are suspected to fall, the accuracy of the detection of the fall detection network model before collision is improved.
The step S3 includes the following sub-steps:
s31, establishing a threshold judgment model of a threshold method, wherein the threshold judgment model comprises: a body offset angle model and an acceleration projection area model;
the body offset angle model is:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
The acceleration projection area model is as follows:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
S32, inputting the standard triaxial acceleration data and triaxial angular velocity data into a threshold judgment model;
s33, calculating total acceleration vector ACC through a body deviation angle modelSVMTotal vector of angular velocity ωSVMAngle Deg to left and right from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωSVMAngle Deg to left and right from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalWhether the current triaxial acceleration data and the current triaxial angular velocity data exceed respective thresholds or not is judged, if yes, the current triaxial acceleration data and the current triaxial angular velocity data are reserved, and the step S35 is skipped, and if not, the step S34 is skipped;
s34, calculating total acceleration vector ACC through an acceleration projection area modelSVMTotal vector of angular velocity ωSVMAnd the human inclination vector PSVMAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωS■VMAnd the human inclination vector PSVMIf the current triaxial acceleration data and the current triaxial angular velocity data exceed the respective threshold values, keeping the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35, otherwise, discarding the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35;
s35, judging whether all data in the standard triaxial acceleration data set and the triaxial angular velocity data set are input into the threshold judgment model, if so, jumping to the step S36, and if not, jumping to the step S32;
and S36, forming a triaxial acceleration data set and a triaxial angular velocity data set of the suspected fall from all the reserved triaxial acceleration data and triaxial angular velocity data.
The body offset angle model and the acceleration projection area model calculate the angle change of the human action, in the falling action, the larger angle change is one of the most obvious characteristics, and the algorithm adopting the angle has higher specificity. The body deviation angle model is considered as a vertical axis and is unchanged, the deviation angle of a person is changed, the acceleration projection area model is considered as a man-made center, the inclination angle change of the person is calculated through three-axis acceleration, the three-axis acceleration and the three-axis acceleration are combined to consider more conditions, the missing report is reduced, for example, when the body of the person is bent such as sitting, the falling detection capacity of the body deviation angle model is stronger, and when the body of the person is not bent such as walking and running, the falling detection capacity of the acceleration projection area model is stronger.
S4, labeling the triaxial acceleration data set and the triaxial angular velocity data set suspected to fall, using the labeled triaxial acceleration data set and the labeled triaxial angular velocity data set as training samples, inputting the labeled training samples into the fall detection network model before collision, and adjusting model parameters to obtain an optimized fall detection network model before collision;
as shown in fig. 2, the fall detection network model before collision includes: an input layer, a hidden layer and an output layer; the input layer, the hidden layer and the output layer are connected in sequence.
The relational expressions of the input layer, the hidden layer and the output layer are as follows:
Hβ=T* (6)
where H is the output of the hidden layer node, β is the output weight of the hidden layer, T*Is a predictive tag.
The process of calculating the output weight β of the hidden layer in step S4 is as follows:
a1, establishing a weight optimization model:
rl=|log2pl|,l=1,…,m (12)
wherein L is the number of hidden layer nodes, C is a hyper-parameter, m is the total number of classification categories, rlIs a weight of class I,. epsilonlIs an error vector of class I, betaiIs the ith weight in the output weight beta, i.e. the output weight of the ith hidden node, i is more than or equal to 1 and less than or equal to L, G (a)i,bi,xj) Is input xjThe output of the corresponding i-th hidden node, G (a)i,bi,xj)=g(aixj+bi) G (—) is an activation function in the hidden layer, aiAs input weights, biBias for the ith hidden node, xjFor the jth input training sample, xj∈RdD is the dimension of the input training sample, tjIs the jth sample xjCorresponding true value label vector, tj∈RmM is the dimension of the label vector,. epsilonjError vector, p, for the jth input training sample1The proportion of the l-th sample in all training samples is U-th.
A2, converting the weight optimization model into a Lagrangian function:
a3, obtaining the output weight beta of the hidden layer according to the Lagrange function:
wherein T' is a real label.
Each class is assigned with a weight, so that the problem of unbalanced data samples is solved, the weight of falling samples which occupy less data samples can be relatively increased, and the unbalanced influence is reduced. Also, the training time can be significantly reduced compared to the method of assigning a weight to each sample.
And S5, filtering the accelerometer and gyroscope data acquired in real time through the step S2 and the step S3, and inputting the data into the optimized before-collision falling detection network model to obtain a real-time before-collision falling detection result.
In conclusion, the beneficial effects of the invention are as follows:
(1) and a threshold method is adopted to quickly filter a large amount of daily activity sample data, retain the sample data of real falling and approximate falling, and reduce the subsequent model training time.
(2) The problem that real falling samples and approximate falling samples are unbalanced is considered, the weight is set for each sample according to the proportion of each sample, all falling samples can be guaranteed to be detected, meanwhile, the false alarm rate is greatly reduced, and the overall identification precision is improved.
(3) The output weight beta of the hidden layer is determined at one time through an equation set without iterative adjustment, the operation amount of the falling detection network model before collision is reduced, the generalization performance of the model is improved, and the model operation is efficient and quick.
Claims (7)
1. A real-time pre-collision falling detection method for old people is characterized by comprising the following steps:
s1, wearing an inertial sensor for the old, and acquiring data of an accelerometer and a gyroscope in the human body activity process in real time to obtain an original triaxial acceleration data set and a triaxial angular velocity data set;
s2, filtering the original triaxial acceleration data set and the triaxial angular velocity data set through a fourth-order IIR Butterworth low-pass filter to obtain a standard triaxial acceleration data set and a triaxial angular velocity data set;
s3, filtering the standard triaxial acceleration data set and the triaxial angular velocity data set through a threshold method to obtain a triaxial acceleration data set and a triaxial angular velocity data set which are suspected to fall;
s4, labeling the triaxial acceleration data set and the triaxial angular velocity data set suspected to fall, using the labeled triaxial acceleration data set and the labeled triaxial angular velocity data set as training samples, inputting the labeled training samples into the fall detection network model before collision, and adjusting model parameters to obtain an optimized fall detection network model before collision;
and S5, filtering the accelerometer and gyroscope data acquired in real time through the step S2 and the step S3, and inputting the data into the optimized before-collision falling detection network model to obtain a real-time before-collision falling detection result.
2. The method for detecting falling before collision of the elderly people in real time as claimed in claim 1, wherein the step S3 comprises the following sub-steps:
s31, establishing a threshold judgment model of a threshold method, wherein the threshold judgment model comprises: a body offset angle model and an acceleration projection area model;
s32, inputting the standard triaxial acceleration data and triaxial angular velocity data into a threshold judgment model;
s33, calculating total acceleration vector ACC through a body deviation angle modelSVMTotal vector of angular velocity ωSVMAngle Deg to left and right from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωSVMTo the left and rightAngle Deg from vertical axissaggitalAnd an angle Deg to deviate fore and aft from the vertical axisFrontalWhether the current triaxial acceleration data and the current triaxial angular velocity data exceed respective thresholds or not is judged, if yes, the current triaxial acceleration data and the current triaxial angular velocity data are reserved, and the step S35 is skipped, and if not, the step S34 is skipped;
s34, calculating total acceleration vector ACC through an acceleration projection area modelSVMTotal vector of angular velocity ωSVMAnd the human inclination vector PSVMAnd determining the total acceleration vector ACCSVMTotal vector of angular velocity ωSVMAnd the human inclination vector PSVMIf the current triaxial acceleration data and the current triaxial angular velocity data exceed the respective threshold values, keeping the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35, otherwise, discarding the current triaxial acceleration data and the current triaxial angular velocity data, and jumping to the step S35;
s35, judging whether all data in the standard triaxial acceleration data set and the triaxial angular velocity data set are input into the threshold judgment model, if so, jumping to the step S36, and if not, jumping to the step S32;
and S36, forming a triaxial acceleration data set and a triaxial angular velocity data set of the suspected fall from all the reserved triaxial acceleration data and triaxial angular velocity data.
3. The method for detecting falling before collision of the elderly people in real time as claimed in claim 2, wherein the body offset angle model in the step S31 is:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
4. The method for detecting falling before collision of the elderly people in real time as claimed in claim 2, wherein the accelerated projection area model in the step S31 is:
wherein, ACCxAcceleration in the x-axis, ACCyAcceleration in the y-axis, ACCzAcceleration in the z-axis, ωpitchIs the pitch angular velocity, omegarollIs the roll angle angular velocity.
5. The method for detecting falling before collision in real time for the elderly as claimed in claim 1, wherein the falling before collision detection network model in step S4 comprises: an input layer, a hidden layer and an output layer; the input layer, the hidden layer and the output layer are connected in sequence.
6. The method for detecting falling before collision of the old in real time according to claim 5, wherein the relational expressions of the input layer, the hidden layer and the output layer are as follows:
Hβ=T*
where H is the output of the hidden layer node, β is the output weight of the hidden layer, T*Is a predictive tag.
7. The method for detecting falling before collision of the elderly people in real time as claimed in claim 6, wherein the step S4 is to calculate the output weight β of the hidden layer by:
a1, establishing a weight optimization model:
rl=|log2pl|,l=1,…,m
wherein L is the number of hidden layer nodes, C is a hyper-parameter, m is the total number of classification categories, rlIs a weight of class I,. epsilonlIs an error vector of class I, betaiIs the ith weight in the output weight beta, i.e. the output weight of the ith hidden node, i is more than or equal to 1 and less than or equal to L, G (a)i,bi,xj) Is input xjOutput of the corresponding i-th hidden node, aiAs input weights, biBias for the ith hidden node, xjFor the jth input training sample, xj∈RdD is the dimension of the input training sample, tjIs the jth sample xjCorresponding true value label vector, tj∈RmM is the dimension of the label vector,. epsilonjTraining for jth inputError vector of sample, plThe proportion of the class I samples in all training samples is U, and the number is a union;
a2, converting the weight optimization model into a Lagrangian function:
a3, obtaining the output weight beta of the hidden layer according to the Lagrange function:
wherein T' is a real label.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010537951.8A CN111710129B (en) | 2020-06-12 | 2020-06-12 | Real-time pre-collision falling detection method for old people |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010537951.8A CN111710129B (en) | 2020-06-12 | 2020-06-12 | Real-time pre-collision falling detection method for old people |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111710129A CN111710129A (en) | 2020-09-25 |
CN111710129B true CN111710129B (en) | 2021-06-01 |
Family
ID=72539863
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010537951.8A Expired - Fee Related CN111710129B (en) | 2020-06-12 | 2020-06-12 | Real-time pre-collision falling detection method for old people |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111710129B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114386459A (en) * | 2021-12-31 | 2022-04-22 | 华南理工大学 | Real-time falling prediction method based on multi-input convolutional network |
CN116631155B (en) * | 2023-06-01 | 2024-07-05 | 深圳市震有智联科技有限公司 | Old people falling identification method and automatic help calling system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101950464A (en) * | 2010-09-17 | 2011-01-19 | 中国科学院深圳先进技术研究院 | Method and system for fall monitoring and warning |
KR20170004265A (en) * | 2015-07-01 | 2017-01-11 | 김주철 | Apparatus and method for detecting falldown |
CN106875630A (en) * | 2017-03-13 | 2017-06-20 | 中国科学院计算技术研究所 | A kind of wearable fall detection method and system based on hierarchical classification |
CN107153871A (en) * | 2017-05-09 | 2017-09-12 | 浙江农林大学 | Fall detection method based on convolutional neural networks and mobile phone sensor data |
CN110047247A (en) * | 2019-05-21 | 2019-07-23 | 武汉理工大学 | A kind of smart home device accurately identifying Falls in Old People |
-
2020
- 2020-06-12 CN CN202010537951.8A patent/CN111710129B/en not_active Expired - Fee Related
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101950464A (en) * | 2010-09-17 | 2011-01-19 | 中国科学院深圳先进技术研究院 | Method and system for fall monitoring and warning |
KR20170004265A (en) * | 2015-07-01 | 2017-01-11 | 김주철 | Apparatus and method for detecting falldown |
CN106875630A (en) * | 2017-03-13 | 2017-06-20 | 中国科学院计算技术研究所 | A kind of wearable fall detection method and system based on hierarchical classification |
CN107153871A (en) * | 2017-05-09 | 2017-09-12 | 浙江农林大学 | Fall detection method based on convolutional neural networks and mobile phone sensor data |
CN110047247A (en) * | 2019-05-21 | 2019-07-23 | 武汉理工大学 | A kind of smart home device accurately identifying Falls in Old People |
Non-Patent Citations (2)
Title |
---|
A Fall Detection Study Based on Neural Network Algorithm Using AHRS;Qingbin Zhang 等;《Proceeding of the IEEE International Conference on Information and Automation》;20131231;第773-779页 * |
老年人跌倒检测***的研究;雒静;《中国优秀硕士学位论文全文数据库 信息科学辑》;20121115;I140-182 * |
Also Published As
Publication number | Publication date |
---|---|
CN111710129A (en) | 2020-09-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Saleh et al. | Driving behavior classification based on sensor data fusion using LSTM recurrent neural networks | |
Tong et al. | HMM-based human fall detection and prediction method using tri-axial accelerometer | |
CN111710129B (en) | Real-time pre-collision falling detection method for old people | |
CN108509897A (en) | A kind of human posture recognition method and system | |
Bourke et al. | Fall detection algorithms for real-world falls harvested from lumbar sensors in the elderly population: A machine learning approach | |
CN104586398A (en) | Old man falling detecting method and system based on multi-sensor fusion | |
CN106875630B (en) | A kind of wearable fall detection method and system based on hierarchical classification | |
CN110287825A (en) | It is a kind of that motion detection method is fallen down based on crucial skeleton point trajectory analysis | |
CN103533888B (en) | For the apparatus and method of the body dimension of mammal of classifying | |
Zhao et al. | Recognition of human fall events based on single tri-axial gyroscope | |
CN106650300B (en) | Old man monitoring system and method based on extreme learning machine | |
Xie et al. | Human activity recognition method based on inertial sensor and barometer | |
EP3459453B1 (en) | Information processing device, information processing method, and information processing program | |
JP4830765B2 (en) | Activity measurement system | |
WO2023035093A1 (en) | Inertial sensor-based human body behaviour recognition method | |
CN106503643A (en) | Tumble detection method for human body | |
CN112115827A (en) | Falling behavior identification method based on human body posture dynamic characteristics | |
CN111931568A (en) | Human body falling detection method and system based on enhanced learning | |
CN103632133A (en) | Human gesture recognition method | |
CN107688828A (en) | A kind of bus degree of crowding estimating and measuring method based on mobile phone sensor | |
CN107007285B (en) | Fall detection method based on pressure and acceleration transducer | |
Goh et al. | Multilayer perceptron neural network classification for human vertical ground reaction forces | |
De et al. | Fall detection approach based on combined two-channel body activity classification for innovative indoor environment | |
CN114913585A (en) | Household old man falling detection method integrating facial expressions | |
Jo et al. | Safety air bag system for motorcycle using parallel neural networks |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210601 |
|
CF01 | Termination of patent right due to non-payment of annual fee |