CN110111360A - A kind of through-wall radar human action characterizing method based on self-organized mapping network - Google Patents
A kind of through-wall radar human action characterizing method based on self-organized mapping network Download PDFInfo
- Publication number
- CN110111360A CN110111360A CN201910316233.5A CN201910316233A CN110111360A CN 110111360 A CN110111360 A CN 110111360A CN 201910316233 A CN201910316233 A CN 201910316233A CN 110111360 A CN110111360 A CN 110111360A
- Authority
- CN
- China
- Prior art keywords
- aen
- indicate
- data
- range profile
- network
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10044—Radar image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
Disclosed herein is a kind of through-wall radar human action characterizing method based on self-organized mapping network, is related to through-wall radar human action characterization technique field, acts characterization problems especially for human body target hidden after the nontransparent barrier such as wall.Initial range picture is pre-processed first, then to reduce data dimension and extracts the feature of each monocycle Range Profile using the AEN with three layers of full articulamentum.Finally, trained AEN is connected using SOM network, to establish the mapping of word to index, the timing monocycle Range Profile data of human motion through walls are converted into the integer sequence comprising movement semantic information.It is an advantage of the invention that the Stepped frequency radar that can accurately estimate the human body specific action type through walls after the nontransparent barrier such as wall, and need to only be received using a hair one, data processing amount are small.
Description
Technical field
The present invention relates to through-wall radar human action characterization technique fields, after the nontransparent barrier such as wall
Hidden human body target acts characterization problems.
Background technique
Through-wall radar human action be characterized in fall down detection, security monitoring, counterterrorism operations and hostage rescue etc. fields have
Significant application value.Domestic and international many research institutions have carried out the research of through-wall radar human action characterization, it was also proposed that a system
The characteristic manner of column human action.Tsinghua University G.Li professor team is acquired human action using more base radars, obtains
Corresponding micro-Doppler feature, and propose the personal identification based on depth convolutional neural networks and Approach for Gait Classification
(Z.Chen,G.Li,F.Fioranelli,et al.Personnel recognition and gait classification
based on multistatic micro-doppler signatures using deep convolutional neural
networks[J].IEEE Geoscience and Remote Sensing Letters,2018:1-5).U.S.'s wiener Nova
University M.Amin professor team has studied the old man based on radar and falls down detection method and technology, by being become using Fourier in short-term
It changes and falls down feature with the linear and bilinearity Time-Frequency Analysis Methods such as wavelet transformation, acquisition and complete to fall down the identification of movement
(M.Amin,Y.Zhang,F.Ahmad,et al.Radar signal processing for elderly fall
detection:The future for in-home monitoring[J].IEEE Signal Processing
Magazine,2016,33(2):71-80).The researcher in portion, Missouri, USA Missouri Economic Commission for Europe, Colombia makes
Indicate the Doppler Feature of various mankind's activities with cepstrum coefficient (MFCC), such as walk, bend over, fall, using SVM and
Two different classifiers such as kNN, according to the MFCC feature of extraction detect automatically tumble movement (L.Liu, M.Popescu,
M.Skubic,et al.Automatic fall detection based on Doppler radar motion
signature[C].International Conference on Pervasive Computing Technologies for
Healthcare and Workshops,Dublin,Ireland,2011,222-225)。
The studies above uses signal transformation method, by the way that original radar signal is transformed into various unlike signal domains, such as
Human action echo through walls is decomposed into several relatively simple representation in components by Doppler domain, wavelet field and cepstrum domain.However this
A little components may have lost some essential characteristics for being difficult to be detected, therefore the inherence that cannot represent human action well is special
Sign.
Summary of the invention
The present invention is to solve through-wall radar human action characterization problems, is proposed a kind of based on coding Self-organizing Maps net certainly
The new characterizing method of the human action through walls of network.Firstly, making it meet deep learning as data pre-process initial range
Training dataset condition.Then using tool, there are three autocoder network (the auto encoder of full articulamentum
Network, AEN) it reduces Range Profile data dimension and extracts Range Profile feature, finally, using Self-organizing Maps (self
Organized mapping, SOM) network connects trained AEN, to establish hidden human action transient posture
The timing Range Profile data of human action through walls are converted to the integer sequence comprising movement semantic information by index mapping.
Technical solution of the present invention is as follows: a kind of through-wall radar human action characterizing method based on self-organized mapping network,
The following steps are included:
Step 1: radar return data processing
Step 1.1, radar echo signal modeling
Using one hair one receipts through-wall radar detection be hidden in a thickness of wall after single human body target movement, one hair
One stepped frequency waveforms for receiving through-wall radar transmitting are s (t), and expression formula is
Wherein, K indicates that the frequency point sum of stepped frequency signal, k indicate that the frequency point number of stepped frequency signal, T indicate each frequency point
Duration, f0Indicate initial frequency, Δ f indicates frequency stepping, and function rect () is
Then, t moment is apart from radar RtarPlace's target reflection echo is expressed as
Wherein, p indicates p-th of scattering point on target surface in P strong scattering, σpIndicate p-th of strong scattering point
Scattering strength, τpIndicate that the trip time delay between p-th of strong scattering point and radar, ψ (t) indicate clutter and noise；
By sr(t) it is mixed with s (t), after filtering out second order and order harmonic frequency, acquired results is carried out with section T
Sampling, obtains following discrete vector
sr=[sr,0,sr,1,…,sr,K-1]T (4)
Wherein
ψkIndicate the clutter and noise when kth time sampling；
By the s in formula (4)rBy Hamming window to inhibit distance side lobe level, f is filled into from 0Hz with interval delta f0Hz it
Afterwards, N point fast adverse Fourier transform is carried out to it, the Range Profile for obtaining a slow time cycle (is referred to as the monocycle below
Range Profile), it is expressed as
sRP=[sRP(0),sRP(1),…,sRP(N-1)]T (6)
Wherein
Indicate the corresponding Range Profile data of n-th of distance unit, n indicates that index number of the distance to unit, value range are
[0, N-1], K0It indicates from 0Hz to f00 filler of Hz, N indicate total points of fast adverse Fourier transform.
Step 1.2, on the basis of the monocycle Range Profile that step 1.1 obtains, using moving-target detect (moving
Target indicator, MTI) algorithm, normalized and thresholding method improve the signal-to-noise ratio of Range Profile data
(signal-noise ratio, SNR), threshold value thresholding can be calculated by following formula:
Wherein,Indicate noise maximum decibel value,Indicate the maximum decibel value of all sample datas,
Indicate the minimum decibel value of all sample datas；
Step 1.3, monocycle Range Profile distance truncation: to reduce calculation amount and data being made to meet actual needs, by single-revolution
Phase Range Profile is intercepted, and interception points should be calculated as
Wherein, NdIndicate the Range Profile points of interception, Rtar,maxIndicate the farthest generation distance of hidden human body target movement,
Rradar,maxIndicate that the farthest detectable one way distance of through-wall radar, N indicate total points of fast adverse Fourier transform；
Step 2: the Range Profile data characteristics preextraction based on AEN；
Since hidden human action changes over time, then the human body attitude at each slow moment time corresponds to a single-revolution
Phase Range Profile, so entire human action corresponds to multicycle Range Profile.Further, since each monocycle Range Profile wraps
Containing hundreds of even more distance unit, so that monocycle Range Profile data processing dimension with higher, it is therefore desirable to right
Pretreated each monocycle Range Profile data carry out dimension-reduction treatment.In order to excavate human body in monocycle Range Profile as much as possible
Posture information and important information is avoided to lose, the present invention carries out posture feature preextraction to monocycle Range Profile using AEN；
Step 2.1, training AEN network:
AEN is constructed using the full articulamentum of multilayer, including numeral network portion is conciliate in coding sub-network part.Two parts exist
Mirror symmetry in structure, and share the common layer for being known as characteristic layer.When inputting monocycle Range Profile action data, coding
Weight parameter effect in data and network is generated coding characteristic and from feature at output by sub-network, and then the coding characteristic is again
Weight parameter effect in decoded sub-network and network attempts to restore original monocycle Range Profile action data.Therefore pass through tune
Weight parameter in whole AEN keeps the error between the output of AEN and input minimum, can obtain monocycle Range Profile in characteristic layer
Effective low-dimensional character representation of action data；
If m-th of slow time cycle Range Profile is expressed asIt is entered into
AEN, then the output of first of hidden layer is expressed as in AEN
Wherein,WithOutputting and inputting for first of hidden layer is respectively indicated, and
W[l-1,l]Indicate the connection weight matrix between l-1 and first of hidden layer；b[l-1,l]Indicate being biased towards in l hidden layer
Amount；F indicates activation primitive, and selects ReLU function as activation primitive, is expressed as
Therefore, the training process of AEN is equivalent to solve following optimization problem
Wherein, W and b respectively indicates all weights in AEN and all bias vectors；Function dist (x, y) indicates x and y
The distance between, usually using mean square error (mean squared error, MSE) function, it is expressed asAEN(sRP
[m], W, b) it indicates to input s under the conditions of W and bRPThe estimation of [m]；
After the completion of step 2.2, AEN network training, only retain the coding sub-network part of AEN, by obtaining corresponding AEN
Dimensionality reduction and the feature extraction of monocycle Range Profile action data are realized in characteristic layer output；
Step 3: the timing characterization of the human action type based on SOM network
After extracting feature by the AEN of pre-training, low-dimensional character representation form is converted by monocycle Range Profile.Therefore,
The time series that the corresponding multicycle Range Profile of hidden human action has translated into a series of corresponding low-dimensional feature vectors indicates
Form.Due to these feature vectors be numerically it is continuous, where character pair space be finite dimension continuous space.But
It is that hidden human action posture of the same movement in the adjacent slow time cycle has similitude, and hidden human action type
Be it is discrete, therefore, continuous feature space is mapped to another new finite dimension discrete space using SOM network by the present invention,
Make the unified character representation of similar posture feature, to be further simplified the characterization of human action through walls, highlights movement at any time
Between state transfer characteristic.
SOM network is a kind of unsupervised learning method using competition mechanism study input data topological structure.SOM network
It is made of single input layer and single competition layer, is fully connected between two layers by largely power connection, the dimension of each power connection
Number is equal to the dimension of SOM network inputs.
Competition layer is made of the multiple neurons for being arranged in array format on two-dimensional surface, these neurons are mutual
Competition, i.e., only one neuron is entered vector activation in SOM network competition layer, remaining neuron is all suppressed.Have
The input of similar features can activate the neuron being closer in competition layer, and the input with different characteristic can activate competition layer
Middle apart from farther away neuron, therefore, SOM network is substantially a kind of clustering method, i.e., special to the monocycle Range Profile of extraction
Sign vector is clustered, and with cluster centre neuron coordinate in a two-dimensional plane be indicated.
Step 3.1, training SOM network:
If including M sample by the data set that the range distribution feature vector extracted from trained AEN forms；
Data: training dataset XM, the sample size M of training data concentration；
Input: size L=Q × Q of competition layer, the number of iterations I learn attenuation rate η；
Output: the weight W of SOM network；
S1. random initializtion Wj(i), wherein j=1 ..., L, i=0；
S2. learning rate α (i)=1, i=0 is initialized；
S3. neighbouring scale σ is initializedΛ(i)=Q/2, i=0；
S4. it for i-th iteration, is handled as follows:
S5. from XMIn take out a data sample X (i) at random；
S6. it solves
S7. renewal learning rate α (i)=α (0) e-η(i-1)；
S8. neighborhood function Λ (j, j are updated*(i),i)；
S9. error is calculated
S10. weight is updated
S11. to step S5~S10 loop iteration I times.
Wherein, j*(i) it is known as the best match unit (Best Matching Unit, BMU) during i-th iteration,
The weight vectors of connection and the input data of i-th iteration are most like；In order to further decrease the weight vector and input data of BMU
Between error, the error transfer factor current weight that right value update process is indicated according to S9；Weight and farther nerve due to BMU
The weight difference of member is increasing, only updates the weight vectors of the neuron near BMU within the scope of adjustable distance, this table in S8
It is shown as neighborhood function Λ (j, j*(i),i)；In general, Λ (j, j*(i), i) it is designated as Gaussian function, it indicates are as follows:
Wherein
Then, the step according to S10 updates weight vectors, and final weight is obtained after I iteration；
Step 3.2, training SOM network after, by step 2 AEN extract Range Profile feature vector be transferred to competition
In layer on the coordinate of BMU, Range Profile maps feature vectors are gone out two similar movements by trained SOM network can be by
Two sentences with similar semantic are translated into, and are shown as two similar space three-dimensional tracks.
The beneficial effects of the present invention are:
The invention proposes a kind of based on from the human action characterizing method through walls for encoding self-organized mapping network, is similar to
Each monocycle Range Profile is considered as a word by natural language, and the human action Range Profile data that more time frames are hidden regard
Simplify for a sentence so that the echo-signal of human motion through walls to be construed to the simple integer sequence with movement semantic
And the validity for keeping human action through walls to indicate.
Detailed description of the invention
Fig. 1 is that a hair one receives the human action schematic diagram of a scenario after through-wall radar detection wall；
Fig. 2 is to characterize block schematic illustration based on the human action through walls from coding self-organized mapping network；
Fig. 3 is autoencoder network structural schematic diagram；
Fig. 4 is self-organized mapping network structural schematic diagram；
Fig. 5 is experiment scene schematic diagram in specific embodiment；
Fig. 6 is the Range Profile example of four seed types movement in specific embodiment, (a) boxing；(b) the not lift of swinging arm
Leg movement；(c) it picks up；(d) two arms are gone up and down in parallel；
Fig. 7 is the initial range picture of four seed types movement and AEN estimation output Range Profile and feature in specific embodiment
The example of extraction；(a)~(d) is boxing, the movement of lift leg, pickup and two arms of swinging arm do not go up and down four kinds in parallel and move
The initial range picture of work；(e)~(h) is that the AEN after training exports Range Profile to the estimation of (a)~(d)；(i)~(l) is corresponding
In the 16 dimension Range Profile feature vectors of (a)~(d) learnt；
The test sample of final sequence expression and corresponding initial range picture that Fig. 8 acts for four seed types, (a)~
It (d) is boxing, the movement of lift leg, pickup and two arms of swinging arm do not go up and down the initial range pictures of four kinds of movements in parallel；
(e)~(h) is the corresponding transfer integer sequence of visual (a)~(d) in three dimensions.
Specific embodiment
A specific embodiment of the invention is provided below according to a specific experimental data:
Specific embodiment using one kind voluntarily develop one hair one receive through-wall radar, by tranmitting frequency be 1.6GHz extremely
2.2GHz stepped frequency signal carries out human action detection through walls.Data acquisition is carried out in the Scientific Research Building A of University of Electronic Science and Technology, is built
Vertical data set.It is as shown in Figure 5 to test scene.It invites 4 volunteers to carry out data acquisition to stand in the range of visibility of radar
Apart from 1.5 position meter Chu of wall, following four movement is carried out: (a) boxing；(b) the not lift leg movement of swinging arm；(c) it bends over
It picks up；(d) two arms are gone up and down in parallel.Fig. 6 gives the Range Profile example of four seed types movement.Each experimenter will be each
Movement repeats 54 times, and the duration acted every time is different, obtains corresponding Range Profile data after pre-processing to radar return.
Step 1: radar return data processing
Step 1.1, radar echo signal modeling
Assuming that being hidden in using one receipts through-wall radar detection of a hair with a thickness of dwSingle human body mesh after the wall of=70cm
Target movement, a hair one receive the stepped frequency signal frequency point sum K=301 of through-wall radar transmitting, and the duration T of each frequency point=
100 μ s, initial frequency f0=1.6GHz, frequency stepping Δ f=2MHz, the expression formula of the stepped frequency signal waveform s (t) of transmitting are
Wherein, function rect () is
Then t moment is apart from radar RtarPlace's target reflection echo can be expressed as
By sr(t) it is mixed with s (t), after filtering out second order and order harmonic frequency, with section T=10-4S is to gained knot
Fruit is sampled, and following discrete vector is obtained
sr=[sr,0,sr,1,…,sr,300]T (19)
Wherein
ψkIndicate the clutter and noise when kth time sampling.
By the s in formula (18)r(t) it by Hamming window with suppressed sidelobes level, and is filled out with interval delta f=2MHz from 0Hz
It is charged to f0=1.6GHz, filler K0=f0/ Δ f=1.6GHz/2MHz=800 carries out N=8192 point fast Fourier to it
Inverse transformation obtains the monocycle Range Profile of slow time dimension.
Step 1.2, on the basis of the Range Profile that step 1.1 obtains, using MTI algorithm, normalized and thresholding side
Method improves the signal-to-noise ratio of Range Profile data.In collected sample database, The setting of threshold value thresholding are as follows:
In addition, 11 differences are added in original radar return in order to which the movement characterization for generating frame has robustness
The additive white Gaussian noise of rank carries out data enhancing, and signal-to-noise ratio is expanded to 18dB, section 1dB from 8dB.Therefore, data
Collection contains (1+11) × 54 × 4 × 4=10368 sample.
Step 1.3, monocycle Range Profile distance truncation: since each test target executes each move at the distance of 1.5m
Make type, execute 8192 Inverse Fast Fourier Transforms, to reduce calculation amount and data being made to meet actual needs, by each example
The Range Profile period be truncated to 512 in pixel.Interception points and the corresponding relationship of actual range are
Step 2: the Range Profile data characteristics preextraction based on AEN
Collected data set is divided into one and includes 8448 exemplary trained Sub Data Set XMIt include 1920 with one
A exemplary test Sub Data Set.
An AEN is constructed to reduce pretreated Range Profile dimension, and carries out feature extraction.Using four fork cross validations
Method select suitable hyper parameter in AEN.The Adadelta algorithm conduct that learning rate is 1.0 is chosen in training for AEN
Optimizer chooses MSE function as initial range as the loss between data and AEN estimated result.Finally choose best super ginseng
Number are as follows: coding dense layer is 3 layers, 256,128,64 neurons of every layer of correspondence；Characteristic layer is 1 layer, 16 nerves of every layer of correspondence
Member.
Fig. 7 shows the example of the low-dimensional Range Profile of the transfer of four sports category.In Fig. 7, each Range Profile by
It is converted into fixed and short length vector, similar feature also corresponds to similar Range Profile.Therefore, AEN can reduce original
The size of beginning Range Profile, without significantly losing information.
Step 3: the timing characterization of the human action type based on SOM network
Competition layer in SOM network is designed and sized to 5 × 5, because bigger size can bring huge calculating to consume
With mode abruption problem.On the other hand, lesser size makes SOM network have lower mode identificating ability.Using from AEN
The feature corresponding with training data of middle extraction utilizes Tensorflow in two NVIDIA GeForce as input
A figure is constructed on GTX1080Ti GPU, each figure has 11GB memory, to enhance training for the SOM network constructed
Journey.
Data: training Sub Data Set XM, the sample size M=8448 of training data concentration；
Input: size L=5 × 5 of competition layer, the number of iterations I=500 learn attenuation rate η=1；
Output: the weight W of SOM network；
S1. random initializtion Wj(i), wherein j=1 ..., L, i=0；
S2. learning rate α (i)=1, i=0 is initialized；
S3. neighbouring scale σ is initializedΛ(i)=5/2, i=0；
S4. for i-th iteration:
S5. from XMIn take out a data sample X (i) at random；
S6. it solves
S7. renewal learning rate α (i)=α (0) e-η(i-1)；
S8. neighborhood function Λ (j, j are updated*(i),i)；
S9. error is calculated
S10. weight is updated
S11. to step S5~S10 loop iteration I times.
During the training period, maximum epoch is set as 500, and initial learning rate is set as 1.0.By 500 training, SOM net
Network has understood the similitude between any two Range Profile, and for each initial range as data establish index.Therefore,
When a people executes specific type of sports, corresponding continuous Range Profile data can be converted into a series of integers, this meaning
Taste movement " semanteme ".
In order to verify validity, proposed frame is had evaluated in test data set.Fig. 8 shows four kinds of movements through walls
The final sequence of type indicates and its corresponding test initial range picture.In fig. 8, in figure 5 it can be seen that (a) each type of sports
Any original time series Range Profile can be converted into the integer sequence extracted by the AEN-SOM frame proposed；(b) different
Type of sports correspond to different integer sequences.Therefore, the human motion representational framework through walls proposed is effective.
Radar parameter in 1 specific embodiment of table
| Parameter | Numerical value |
| Frequency range | 1.6GHz~2.2GHz |
| Frequency point number | 301 |
| The frequency point duration | 100μs |
| Duration monocycle | 30.1ms |
Claims (1)
1. a kind of through-wall radar human action characterizing method based on self-organized mapping network, comprising the following steps:
Step 1: radar return data processing
Step 1.1, radar echo signal modeling
Using one hair one receipts through-wall radar detection be hidden in a thickness of wall after single human body target movement, one hair one receive
The stepped frequency waveforms of through-wall radar transmitting are s (t), and expression formula is
Wherein, K indicates that the frequency point sum of stepped frequency signal, k indicate that the frequency point number of stepped frequency signal, T indicate holding for each frequency point
Continuous time, f0Indicate initial frequency, Δ f indicates frequency stepping, and function rect () is
Then, t moment is apart from radar RtarPlace's target reflection echo is expressed as
Wherein, p indicates p-th of scattering point on target surface in P strong scattering, σpIndicate that the scattering of p-th of strong scattering point is strong
Degree, τpIndicate that the trip time delay between p-th of strong scattering point and radar, ψ (t) indicate clutter and noise；
By sr(t) it is mixed with s (t), after filtering out second order and order harmonic frequency, acquired results is sampled with section T,
Obtain following discrete vector
sr=[sr,0,sr,1,…,sr,K-1]T (4)
Wherein
ψkIndicate the clutter and noise when kth time sampling；
By the s in formula (4)rBy Hamming window to inhibit distance side lobe level, f is filled into from 0Hz with interval delta f0After Hz,
N point fast adverse Fourier transform is carried out to it, the Range Profile for obtaining a slow time cycle (is referred to as monocycle distance below
Picture), it is expressed as
sRP=[sRP(0),sRP(1),…,sRP(N-1)]T (6)
Wherein
Indicate the corresponding Range Profile data of n-th of distance unit, n indicates that index number of the distance to unit, value range are [0, N-
1], K0It indicates from 0Hz to f00 filler of Hz, N indicate total points of fast adverse Fourier transform.
Step 1.2, on the basis of the monocycle Range Profile that step 1.1 obtains, using moving-target detection algorithm, normalized
The signal-to-noise ratio of Range Profile data is improved with thresholding method, threshold value thresholding can be calculated by following formula:
Wherein,Indicate noise maximum decibel value,Indicate the maximum decibel value of all sample datas,It indicates
The minimum decibel value of all sample datas；
Step 1.3, the monocycle Range Profile distance truncation: for reduce calculation amount and make data meet actual needs, by the monocycle away from
From as being intercepted, interception points should be calculated as
Wherein, NdIndicate the Range Profile points of interception, Rtar,maxIndicate the farthest generation distance of hidden human body target movement,
Rradar,maxIndicate that the farthest detectable one way distance of through-wall radar, N indicate total points of fast adverse Fourier transform；
Step 2: the Range Profile data characteristics preextraction based on AEN；
Step 2.1, training AEN network:
AEN is constructed using the full articulamentum of multilayer, including numeral network portion is conciliate in coding sub-network part.Two parts are in structure
Upper mirror symmetry, and share the common layer for being known as characteristic layer.When inputting monocycle Range Profile action data, subnet is encoded
Weight parameter effect in data and network is generated coding characteristic and from feature at output by network, and then the coding characteristic is again through solving
Weight parameter effect in numeral network and network attempts to restore original monocycle Range Profile action data.Therefore by adjusting
Weight parameter in AEN keeps the error between the output of AEN and input minimum, and it is dynamic can to obtain monocycle Range Profile in characteristic layer
Make effective low-dimensional character representation of data；
If m-th of slow time cycle Range Profile is expressed asAEN is entered into,
The output of first of hidden layer is expressed as in so AEN
Wherein,WithOutputting and inputting for first of hidden layer is respectively indicated, and
W[l-1,l]Indicate the connection weight matrix between l-1 and first of hidden layer；b[l-1,l]Indicate being biased towards in l hidden layer
Amount；F indicates activation primitive, and selects ReLU function as activation primitive, is expressed as
Therefore, the training process of AEN is equivalent to solve following optimization problem
Wherein, W and b respectively indicates all weights in AEN and all bias vectors；Function dist (x, y) is indicated between x and y
Distance be expressed as usually using mean square error functionAEN(sRP[m], W, b) it indicates to input under the conditions of W and b
sRPThe estimation of [m]；
After the completion of step 2.2, AEN network training, only retain the coding sub-network part of AEN, by obtaining corresponding AEN feature
Dimensionality reduction and the feature extraction of monocycle Range Profile action data are realized in layer output；
Step 3: the timing characterization of the human action type based on SOM network
Step 3.1, training SOM network:
If including M sample by the data set that the range distribution feature vector extracted from trained AEN forms；
Data: training dataset XM, the sample size M of training data concentration；
Input: size L=Q × Q of competition layer, the number of iterations I learn attenuation rate η；
Output: the weight W of SOM network；
S1. random initializtion Wj(i), wherein j=1 ..., L, i=0；
S2. learning rate α (i)=1, i=0 is initialized；
S3. neighbouring scale σ is initializedΛ(i)=Q/2, i=0；
S4. it for i-th iteration, is handled as follows:
S5. from XMIn take out a data sample X (i) at random；
S6. it solves
S7. renewal learning rate α (i)=α (0) e-η(i-1)；
S8. neighborhood function Λ (j, j are updated*(i),i)；
S9. error is calculated
S10. weight is updated
S11. to step S5~S10 loop iteration I times.
Wherein, j*(i) be known as the best match unit during i-th iteration, the weight vectors of connection and i-th iteration it is defeated
It is most like to enter data；In order to further decrease the error between the weight vector of BMU and input data, right value update process is according to S9
The error transfer factor current weight of expression；Since the weight of BMU and the weight difference of farther neuron are increasing, only update
The weight vectors of neuron near BMU within the scope of adjustable distance, this is expressed as neighborhood function Λ (j, j in S8*(i),i)；
In general, Λ (j, j*(i), i) it is designated as Gaussian function, it indicates are as follows:
Wherein
Then, the step according to S10 updates weight vectors, and final weight is obtained after I iteration；
Step 3.2, training SOM network after, by step 2 AEN extract Range Profile feature vector be transferred in competition layer
On the coordinate of BMU, Range Profile maps feature vectors, which are gone out two similar movements, by trained SOM network to be translated
At two sentences with similar semantic, and it is shown as two similar space three-dimensional tracks.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910316233.5A CN110111360B (en) | 2019-04-19 | 2019-04-19 | Through-wall radar human body action characterization method based on self-organizing mapping network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910316233.5A CN110111360B (en) | 2019-04-19 | 2019-04-19 | Through-wall radar human body action characterization method based on self-organizing mapping network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110111360A true CN110111360A (en) | 2019-08-09 |
| CN110111360B CN110111360B (en) | 2022-05-03 |
Family
ID=67485775
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910316233.5A Active CN110111360B (en) | 2019-04-19 | 2019-04-19 | Through-wall radar human body action characterization method based on self-organizing mapping network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110111360B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111796272A (en) * | 2020-06-08 | 2020-10-20 | 桂林电子科技大学 | Real-time gesture recognition method and computer equipment for through-wall radar human body image sequence |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090262005A1 (en) * | 2005-04-14 | 2009-10-22 | L-3 Communications Cyterra Corporation | Moving-entity detection |
| US20100026550A1 (en) * | 2007-07-17 | 2010-02-04 | Rosenbury Erwin T | Handheld Instrument Capable of Measuring Heartbeat and Breathing Motion at a Distance |
| CN106127110A (en) * | 2016-06-15 | 2016-11-16 | 中国人民解放军第四军医大学 | A kind of human body fine granularity motion recognition method based on UWB radar with optimum SVM |
| CN107132512A (en) * | 2017-03-22 | 2017-09-05 | 中国人民解放军第四军医大学 | UWB radar human motion micro-Doppler feature extracting method based on multichannel HHT |
| CN107219522A (en) * | 2017-05-08 | 2017-09-29 | 电子科技大学 | A kind of united through-wall radar object localization method of ellipse-hyperbolic |
| CN107657243A (en) * | 2017-10-11 | 2018-02-02 | 电子科技大学 | Neutral net Radar range profile's target identification method based on genetic algorithm optimization |
| CN108776336A (en) * | 2018-06-11 | 2018-11-09 | 电子科技大学 | A kind of adaptive through-wall radar static human body object localization method based on EMD |
-
2019
- 2019-04-19 CN CN201910316233.5A patent/CN110111360B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090262005A1 (en) * | 2005-04-14 | 2009-10-22 | L-3 Communications Cyterra Corporation | Moving-entity detection |
| US20100026550A1 (en) * | 2007-07-17 | 2010-02-04 | Rosenbury Erwin T | Handheld Instrument Capable of Measuring Heartbeat and Breathing Motion at a Distance |
| CN106127110A (en) * | 2016-06-15 | 2016-11-16 | 中国人民解放军第四军医大学 | A kind of human body fine granularity motion recognition method based on UWB radar with optimum SVM |
| CN107132512A (en) * | 2017-03-22 | 2017-09-05 | 中国人民解放军第四军医大学 | UWB radar human motion micro-Doppler feature extracting method based on multichannel HHT |
| CN107219522A (en) * | 2017-05-08 | 2017-09-29 | 电子科技大学 | A kind of united through-wall radar object localization method of ellipse-hyperbolic |
| CN107657243A (en) * | 2017-10-11 | 2018-02-02 | 电子科技大学 | Neutral net Radar range profile's target identification method based on genetic algorithm optimization |
| CN108776336A (en) * | 2018-06-11 | 2018-11-09 | 电子科技大学 | A kind of adaptive through-wall radar static human body object localization method based on EMD |
Non-Patent Citations (5)
| Title |
|---|
| MINGMIN ZHAO等: "Through-Wall Human Pose Estimation Using Radio Signals", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| XIAOLIN LIANG等: "Through-wall human being detection using UWB impulse radar", 《EURASIP JOURNAL ON WIRELESS COMMUNICATIONS AND NETWORKING》 * |
| 付庆霞等: "基于小波与EMD级联的穿墙雷达动目标检测", 《雷达科学与技术》 * |
| 李松林等: "基于改进Camshift的穿墙雷达运动人体目标成像跟踪算法", 《计算机应用》 * |
| 贾勇等: "穿墙雷达多视角建筑布局成像", 《电子与信息学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111796272A (en) * | 2020-06-08 | 2020-10-20 | 桂林电子科技大学 | Real-time gesture recognition method and computer equipment for through-wall radar human body image sequence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110111360B (en) | 2022-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109522857B (en) | People number estimation method based on generation type confrontation network model | |
| CN109993280B (en) | Underwater sound source positioning method based on deep learning | |
| CN108226892B (en) | Deep learning-based radar signal recovery method in complex noise environment | |
| CN109389058B (en) | Sea clutter and noise signal classification method and system | |
| CN109683161B (en) | Inverse synthetic aperture radar imaging method based on depth ADMM network | |
| CN104459668A (en) | Radar target recognition method based on deep learning network | |
| CN104820993B (en) | It is a kind of to combine particle filter and track the underwater weak signal target tracking for putting preceding detection | |
| Liu et al. | Deep learning and recognition of radar jamming based on CNN | |
| CN104899567A (en) | Small weak moving target tracking method based on sparse representation | |
| CN105738888B (en) | Bicharacteristic offshore floating small target detecting method based on ocean clutter cancellation | |
| CN112861813B (en) | Method for identifying human behavior behind wall based on complex value convolution neural network | |
| CN110109080A (en) | Method for detecting weak signals based on IA-SVM model | |
| CN103714331A (en) | Facial expression feature extraction method based on point distribution model | |
| Qu et al. | Human activity recognition based on WRGAN-GP-synthesized micro-Doppler spectrograms | |
| CN110933633A (en) | Onboard environment indoor positioning method based on CSI fingerprint feature migration | |
| CN110223342B (en) | Space target size estimation method based on deep neural network | |
| Zhang et al. | Multi-source information fused generative adversarial network model and data assimilation based history matching for reservoir with complex geologies | |
| CN109086667A (en) | Similar active recognition methods based on intelligent terminal | |
| CN116269259A (en) | Human body breathing and heartbeat detection method based on improved BP neural network | |
| CN110275147A (en) | Human behavior micro-Doppler classification and identification method based on migration depth neural network | |
| Shao et al. | Deep learning methods for personnel recognition based on micro-Doppler features | |
| CN114428234A (en) | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention | |
| CN110111360A (en) | A kind of through-wall radar human action characterizing method based on self-organized mapping network | |
| Wang et al. | Through-wall human motion representation via autoencoder-self organized mapping network | |
| Li et al. | Sea/land clutter recognition for over-the-horizon radar via deep CNN |
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 |