CN110348288A - A kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING - Google Patents
A kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING Download PDFInfo
- Publication number
- CN110348288A CN110348288A CN201910445702.3A CN201910445702A CN110348288A CN 110348288 A CN110348288 A CN 110348288A CN 201910445702 A CN201910445702 A CN 201910445702A CN 110348288 A CN110348288 A CN 110348288A
- Authority
- CN
- China
- Prior art keywords
- gesture
- signal
- spectrogram
- feature
- spectrum
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- 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
-
- 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
- G06V40/28—Recognition of hand or arm movements, e.g. recognition of deaf sign language
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Social Psychology (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The present invention provides a kind of gesture identification methods based on 77GHz MMW RADAR SIGNAL USING, intermediate-freuqncy signal including obtaining different gesture motions by radar first, and innovatively its low frequency coefficient is pre-processed using a kind of improved wavelet threshold function, solve the problems, such as that the short distance gesture as caused by antenna coupling phenomenon is unrecognized, secondly, to pretreated intermediate-freuqncy signal extraction time-apart from spectrogram, when m- normal-moveout spectrum figure and when m- angular spectrum figure, innovatively three kinds of feature spectrograms are spliced to obtain Multivariate characteristics figure, and it is input to convolutional neural networks and is trained, it optimizes tional identification algorithm information and expresses incomplete problem, it is also beneficial to the simplification of network structure simultaneously, and finally obtain preferable recognition effect.
Description
Technical field
The present invention relates to Radar Signal Processing identification technology fields, in particular to a kind of to be believed based on 77GHz millimetre-wave radar
Number gesture identification method.
Background technique
Since 21st century, the development advanced by leaps and bounds with computer technology, human-computer interaction technology is had become now
One of great subject technology.Currently, common man-machine interaction method is using mouse and keyboard as mechanical input equipment, so
And these methods have no idea to realize the information interaction of simple, efficient height free people and computer.In computer and letter
In number process field development process, Gesture Recognition with it lively, vivid, intuitive, efficient expression characteristic possess it is more next
More application scenarios, such as smart home system, sign language real-time teaching, gesture manipulation game system etc..Along with
The rapid development of human-computer interaction technology, Gesture Recognition have become the research hotspot of domestic and international each scholar at present.
Traditional Gesture Recognition overwhelming majority is all based on video and image, such as the body-sensing of Microsoft is set
Standby Kinect, is exactly formed with the scene figure of depth using 3D body-sensing camera and pumped FIR laser technology, by assessing depth image
Pixel-level, carry out picture depth identification, each position or the gesture of human body are captured in conjunction with skeleton tracer technique
Movement.However, traditional gesture identification method based on image or video has some limitations.Firstly, it is traditional based on
The recognition correct rate of the Gesture Recognition of image is highly prone to the influence of the factors such as illumination, weather and working environment.Its
Secondary, traditional gesture identification method based on image or video is also easy the influence being blocked, such as wall, bookcase, when
When gesture executor is located at after wall or somewhere is partially or completely the position blocked in room, this method will be complete
It is unworkable.In addition, there is also the risks that privacy of user is revealed for traditional gesture identification method based on image or video.A
People's information highly sensitive epoch, privacy leakage problem caused by this method may generate very serious consequence.Finally,
Traditional gesture identification method based on image or video is opposite with the requirement of energy consumption to computing resource relatively high, general feelings
Independent extraneous power supply system is required under condition, this is very restricted its application scenarios and scale all.
For tradition is based on the gesture identification method of video or image, the gesture identification side based on radar signal
Method usually has non-contact, is not illuminated by the light, the features such as weather and working environment influence, can effectively solve the problem that illumination deficiency etc.
Condition influences the problem of recognition correct rate.Radar also has the function of that propagation is blocked in certain penetrating simultaneously, can effectively keep away
Exempt from the influence of the shelters such as wall, bookcase, so that gesture executor realizes gesture in the case where being blocked completely or partial occlusion
Control is possibly realized with interacting.And the gesture identification method based on radar signal can be eliminated effectively due to shooting video
Or privacy of user leakage problem caused by image, this advantage have also ensured the safety of user while protecting privacy of user
Property.In addition, radar sensor can be integrated in mostly on the relatively low high quality chip of energy consumption, identification can be substantially reduced
Cost simultaneously reduces computational complexity, greatly increases the application scenarios and scale of Gesture Recognition.77GHz millimetre-wave radar with
Its light weight, feature small in size, have been more and more widely used.Meanwhile 77GHz millimetre-wave radar is easy for reach
To higher spatial resolution, so that ranging, angle measurement, the precision that tests the speed are higher.In conjunction with current 77GHz millimetre-wave radar in intelligence
The application of energy driving, smart home etc., the Gesture Recognition based on 77GHz millimetre-wave radar possess widely
Application prospect.
It is had the following problems in the existing gesture identification method based on radar signal.Firstly, for due to transmitting antenna
For the biggish low frequency signal of energy generated with receiving antenna coupling, current method is intercepted from distance domain, is lost
Fall short distance information.In fact, when gesture is closer apart from radar, this way can incite somebody to action for dynamic gesture motion
Useful hand signal is got rid of together, is unscientific.Secondly, not filled to radar angular information using less at present
Divide the useful information that can be provided using radar, and lacks when constructing the input data set of convolutional neural networks to characteristic spectrum
The information of figure merges, and increases the design difficulty of the convolutional neural networks in later period.
Summary of the invention
The object of the present invention is to provide a kind of removal since the energy of transmitting antenna and receiving antenna coupling generation is biggish
The method of low frequency signal, and a kind of construction method of radar hand signal Multivariate characteristics map is provided.With traditional gesture identification
Technology is compared, and on the one hand can be identified to short distance hand signal, be further increased the applied field of Gesture Recognition
Scape;On the other hand make hand signal feature representation more complete, it is lower to solve the problems, such as that previous gesture information amount describes,
And be conducive to simplify the design of later period convolutional neural networks, it is easy to implement the precise classification of various gestures.
A kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING of the present invention, comprising the following steps:
Step 1: design N number of gesture motion, by taking N=4 as an example, design makes hook, radially waves, rotates clockwise and the inverse time
Needle rotates four kinds of gesture motions, and carries out the data acquisition of corresponding gesture under microwave dark room environment by different volunteers, always
4*ClassNum group data are obtained;
Step 2: signal S will be emittedT(t) and receive signal SR(t) by frequency mixer, mixed frequency signal S is obtainedMIX(t),
SMIX(t) by low-pass filter, intermediate-freuqncy signal x (t) is parsed, and extracts corresponding receiving antenna from intermediate-freuqncy signal and corresponds to
Signal SNI\Q(t);
Step 3: there is one in gesture intermediate-freuqncy signal x (t) due to the coupling phenomenon of transmitting antenna and receiving antenna
The very big low frequency signal of energy, using wavelet thresholding methods to above-mentioned intermediate-freuqncy signal SNI\Q(t) it is pre-processed, selects one kind
Improved threshold function table, and after only being processed to low frequency coefficient, reconstruct obtains new intermediate-freuqncy signal
Step 4: to intermediate-freuqncy signalHandled, estimate gesture motion when m- distance spectrum figure, when it is m-
Speed spectrogram and when m- angular spectrum figure, and numerical value normalized is carried out respectively to three kinds of spectrograms;
Step 5: after being normalized in step 4 when m- distance spectrum figure, when m- normal-moveout spectrum figure and when m- angular spectrum
Figure is spliced, and Multivariate characteristics spectrogram A is constructed;
Step 6: carrying out step 2-step 6 behaviour respectively to collected 4*ClassNum group gesture echo data
Make, obtains the original hand signal Multivariate characteristics spectrogram collection of 4*ClassNum group
Step 7: the Multivariate characteristics spectrogram collection that step 6 is obtainedEach of sample standard deviation carry out gray processing at
Reason, obtains original Multivariate characteristics atlas B;
Step 8: mean value is carried out to all samples in original Multivariate characteristics atlas B and carries out dimension normalization,
Obtain Multivariate characteristics atlasAnd it labels to each width characteristic pattern;
Step 9: by Multivariate characteristics atlasIt is divided into training set S according to a certain percentagetrain, verifying collection SvalAnd test
Collect Stest, such as training set accounts for 70%, verifying collection accounts for 20%, test set and accounts for 10%;
Step 10: by Strain、SvalInput data set of the label corresponding with its together as convolutional neural networks
Cinput, and network weight is initialized, wherein StrainFor being trained to net coefficients, SvalIt is laggard in training a period of time
Row network verification, and be adjusted by weight of the error to network;
Step 11: by input data set CinputA convolution pondization operation is carried out, the scale of convolution kernel is set
Kernel_size1, convolution step-length kernel_stride1 and pond size pool_size1, pond step-length pool_
Stride1 obtains feature atlas feature1;
Step 12: extracting further feature to feature atlas feature1 further progress convolution pond, convolution is set
Size kernel_size2, the convolution step-length kernel_stride2 and pond size pool_size2, pond step-length of core
Pool_stride2 obtains feature atlas feature2;
Step 13: carrying out a convolution pondization operation again to feature atlas feature2, deeper time feature is extracted,
Size kernel_size3, the convolution step-length kernel_stride3 and pond size pool_size3, pond of convolution kernel are set
Change step-length pool_stride3, obtains feature atlas feature3;
Step 14: feature3 is passed sequentially through full articulamentum fc4, fc5, fc6, and fc4, fc5, fc6 is respectively set
Size be size4, size5, size6, feature atlas is converted into the column vector v1 of 1 × size6;
Step 15: output is different gesture classifications, by multiple by column vector v1 by softmax classifier
Iteration, network accuracy rate and loss function tend towards stability, and obtain trained convolutional neural networks model Netmodel.
Step 16: by test data set StestIt is loaded into Netmodel, obtains gesture classification result y.
The step 3 the following steps are included:
3.1 choose suitable wavelet basis function, to gesture intermediate-freuqncy signal SNI\Q(t) N layers of wavelet transformation are carried out, are obtained close
Like coefficient A (i) and detail coefficients D (1, i), D (2, i) ..., D (N, i);
3.2 couples of low frequency coefficient A (i) do threshold process, i.e.,
Ai'=mAi+(1-m)sgn(Ai)(|Ai|-n)
3.3 utilizingD (1, i), D (2, i) ..., D (N, i) signal is reconstructed, obtain pretreated gesture
Intermediate-freuqncy signal x'(t).
The step 4 the following steps are included:
4.1 extraction times-are apart from spectrogram: making FFT in fast time-domain first, then weighting asks flat in slow time-domain
, it finally carries out interframe and accumulates m- distance spectrum figure, size FFTNum1*FrameNum when can be obtained;
4.2 extraction times-speed spectrogram: carrying out Two-dimensional FFT first as unit of frame, obtains distance-Doppler figure, connects
Detection range-Dopplergram maximum value, and extract the maximum value and be expert at, using the velocity information of the row as present frame
The speed of signal, finally m- normal-moveout spectrum figure, size FFTNum2*FrameNum when interframe is accumulated to obtain;
4.3 extraction times-angle spectrogram: first as unit of frame, the correspondence frame signal of N number of receiving antenna is done respectively
Two-dimensional FFT obtains N range from-Dopplergram.Is extracted by maximum value respectively, and maximum value is arranged from-Dopplergram for N range
At the one-dimensional vector of 1 × N, FFT is done to the one-dimensional vector, the corresponding velocity information of current frame signal can be obtained, finally in frame
Between carry out accumulation m- angle figure, size FFTNum3*FrameNum when can be obtained;
4.4 numerical value normalizeds: first have to each feature spectrogram carry out deviation standardization, specifically, with when
For the m- D apart from spectrogram, numerical value scaling is carried out as follows to it:
By above-mentioned zoom operations, when m- distance spectrum figure in each value all fall in [0,1] range.
Detailed description of the invention:
Fig. 1 is flow chart of the invention.
Fig. 2 is the specific gesture motion that identification is designed in the present invention.
Fig. 3 is each gesture motion wavelet low frequency threshold process analogous diagram.
Fig. 4 is the when m- distance spectrum figure analogous diagram of each gesture motion.
Fig. 5 is the when m- normal-moveout spectrum figure analogous diagram of each gesture motion.
Fig. 6 is the when m- angular spectrum figure analogous diagram of each gesture motion.
Fig. 7 is the Multivariate characteristics figure (result after normalization) of each gesture motion.
Fig. 8 is the network architecture of convolutional neural networks.
Fig. 9 is table 1.
Specific embodiment
The present invention is further illustrated in the following with reference to the drawings and specific embodiments.
By taking four kinds of gestures as an example:
Step 1: hook is made in design, four kinds of gesture motions are waved, rotate clockwise and rotated counterclockwise to radial direction, such as Fig. 2 institute
Show, the relevant parameter of 77GHz millimetre-wave radar is configured first.In this patent, setting sample frequency is 2000kHz, frame
Period is 55ms, acquires 100 frame data every time, has 128 chirp signals in every frame, each chirp signal has 64 samplings
Point.Antenna uses one transmitter and four receivers, i.e. a transmitting antenna, four receiving antennas, and is volunteered under microwave dark room environment by three
Person carries out corresponding gesture data acquisition;
Step 2: signal S will be emittedT(t) with receive signal SR(t) it is mixed by frequency mixer, obtains mixed frequency signal
SMIX(t), and by SMIX(t) it by low-pass filter, obtains intermediate-freuqncy signal x (t), and extracts from intermediate-freuqncy signal and connect accordingly
Receive the corresponding signal SN of antennaI\Q(t), the specific steps are as follows:
Step: the expression of 2-1 radar emission signal are as follows:
Wherein, ATFor the amplitude for emitting signal, fcFor carrier frequency, T is sawtooth period, and B is signal bandwidth, fT(t)
For the frequency for emitting signal in T time;
Step 2-2 radar return signal is after emitting signal delay Δ t as a result, its result is as follows:
Wherein, ARFor the amplitude for receiving signal, fRIt (t) is the frequency of the inscribed collection of letters number of T time, Δ f is frequency shift (FS);
Step 2-3: signal S will be emittedT(t) and receive signal SR(t) by frequency mixer, mixed frequency signal S is obtainedMIX(t)
Expression formula is as follows:
SMIX(t)=ST(t)SR(t)
Step 2-4: by mixed frequency signal SMIX(t) by low-frequency filter, it is as follows to obtain intermediate-freuqncy signal x (t) expression formula:
Step 2-5: use one transmitter and four receivers, four receiving antennas share 8 road I the channel Q, by intermediate-freuqncy signal x (t) every 8
A point is assigned to a circuit-switched data, can extract the corresponding intermediate-freuqncy signal SN of the road antenna that hauntsI\Q(t);
Step 3: there is an energy in gesture intermediate-freuqncy signal due to the coupling phenomenon of transmitting antenna and receiving antenna
Very big low frequency signal, using wavelet thresholding methods to above-mentioned intermediate-freuqncy signal SNI\Q(t) it is pre-processed, selects a kind of improvement
Threshold function table, and after only processing to low frequency coefficient, reconstruct obtains new intermediate-freuqncy signalSpecific step is as follows:
Step 3-1: select Sym6 as wavelet basis function, to gesture intermediate-freuqncy signalCarry out three layers of wavelet decomposition
Afterwards, approximation coefficient A (i), detail coefficients D (1, i), D (2, i), D (3, i) are obtained,
Step 3-2: threshold process is carried out to its low frequency coefficient A (i), i.e.,
Ai'=mAi+(1-m)sgn(Ai)(|Ai|-n)
Step 3-3: it usesD (1, i), D (2, i), D (3, i) are reconstructed to arrive pretreated signal x'
(t), each gesture motion wavelet low frequency threshold process simulation result is as shown in Figure 3.
Step 4: to intermediate-freuqncy signalHandled, estimate gesture motion when m- distance spectrum figure, when it is m-
Speed spectrogram and when m- angular spectrum figure, and numerical value normalized is carried out respectively to three kinds of spectrograms.Specific step is as follows:
4.1 extraction times-are apart from spectrogram: making FFT in fast time-domain first, then weighting asks flat in slow time-domain
, it finally carries out interframe and accumulates m- distance spectrum figure when can be obtained, size FFTNum1*FrameNum, each gesture is moved
M- distance spectrum figure simulation result is as shown in Figure 4 when making;
4.2 extraction times-speed spectrogram: carrying out Two-dimensional FFT first as unit of frame, obtains distance-Doppler figure, connects
Detection range-Dopplergram maximum value, and extract the maximum value and be expert at, using the velocity information of the row as present frame
The speed of signal, the last m- normal-moveout spectrum figure when interframe is accumulated to obtain, size FFTNum2*FrameNum, respectively
M- normal-moveout spectrum figure simulation result is as shown in Figure 5 when gesture motion;
4.3 extraction times-angle spectrogram: first as unit of frame, the correspondence frame signal of N number of receiving antenna is done respectively
Two-dimensional FFT obtains N range from-Dopplergram.Is extracted by maximum value respectively, and maximum value is arranged from-Dopplergram for N range
At the one-dimensional vector of 1 × N, FFT is done to the one-dimensional vector, the corresponding velocity information of current frame signal can be obtained, finally in frame
Between carry out accumulation m- angle figure when can be obtained, size FFTNum3*FrameNum, when each gesture motion m- angle
Spectrogram simulation result is as shown in Figure 6;
4.4 numerical value normalizeds: first have to each feature spectrogram carry out deviation standardization, specifically, with when
For the m- D apart from spectrogram, numerical value scaling is carried out as follows to it:
By above-mentioned zoom operations, when m- distance spectrum figure in each value all fall in [0,1] range.
Step 5: after being normalized in step 4 when m- distance spectrum figure, when m- normal-moveout spectrum figure and when m- angular spectrum
Figure is spliced by column, and construction its size of Multivariate characteristics spectrogram A is (FFTNum1+FFTNum2+FFTNum3) *
FrameNum;
Step 6: carrying out step 2-step 6 behaviour respectively to collected 4*ClassNum group gesture echo data
Make, obtains the original hand signal Multivariate characteristics spectrogram collection of 4*ClassNum group
Step 7: the Multivariate characteristics spectrogram collection that step 6 is obtainedEach of sample standard deviation carry out gray processing at
Reason, obtains original Multivariate characteristics atlas B;
Step 8: mean value is carried out to all samples in original Multivariate characteristics atlas B and carries out dimension normalization,
Obtain Multivariate characteristics atlasEach gesture motion Multivariate characteristics figure result is as shown in fig. 7, and paste each width characteristic pattern
Label, steps are as follows for collective:
8.1, according to following formula, seek the mean value of all samples in Multivariate characteristics atlas B
Here it is worth noting that averaged to all samples in diversification data set B here, rather than only needle
It averages in class to certain one kind gesture;
Any sample in 8.2 couples of Multivariate characteristics atlas B carries out mean value cancellation, i.e.,
8.3 pairs of Multivariate characteristics atlasCarry out dimension normalization: the dimension normalization by characteristic pattern each in data set is
Hight × Width specifically carries out down-sampling when primitive character figure size is greater than Hight × Width, otherwise carries out
Sampling, finally obtains normalized Multivariate characteristics atlas
Step 9: by Multivariate characteristics atlasIt is divided into training set S according to a certain percentagetrain, verifying collection SvalAnd test
Collect Stest, such as training set accounts for 70%, verifying collection accounts for 20%, test set and accounts for 10%;
Step 10: by Strain、SvalInput data set of the label corresponding with its together as convolutional neural networks
Cinput, and network weight is initialized, wherein StrainFor being trained to net coefficients, SvalIt is laggard in training a period of time
Row network verification, and be adjusted by weight of the error back propagation to network;
Step 11: by input data set CinputCarry out the operation of convolution pondization, be arranged the scale 11 of convolution kernel ×
11, convolution step-length 4 and pond size 3 × 3, pond step-length 2, obtain feature atlas feature1;
Step 12: extracting further feature to feature atlas feature1 further progress convolution pond, convolution is set
Size 5 × 5, convolution step-length 1 and pond size 3 × 3, the pond step-length 2 of core, obtain feature atlas feature2;
Step 13: carrying out a convolution pondization operation again to feature atlas feature2, deeper time feature is extracted,
Size 3 × 3, convolution step-length 1 and pond size 3 × 3, the pond step-length 2 of convolution kernel are set, feature atlas is obtained
feature3;
Step 14: feature3 is passed sequentially through full articulamentum fc4, fc5, fc6, and fc4, fc5, fc6 is respectively set
Size be 4096,2048,1000, feature atlas is converted into 1 × 1000 column vector v1;
Step 15: output is different gesture classifications, by multiple by column vector v1 by softmax classifier
Iteration, network accuracy rate and loss function tend towards stability, and obtain trained convolutional neural networks model Netmodel.
Step 16: by test data set StestIt is loaded into Netmodel, obtains gesture classification result y, obtain gesture
Data classification result such as table 1.
Claims (3)
1. a kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING, which comprises the following steps:
Step 1: designing N number of gesture motion, and the data of corresponding gesture are carried out under microwave dark room environment by different volunteers
Acquisition, is always obtained N*ClassNum group data;
Step 2: being parsed to radar initial data, obtain intermediate-freuqncy signal x (t), and extract from intermediate-freuqncy signal and connect accordingly
Receive the corresponding intermediate-freuqncy signal SN of antennaI\Q(t);
Step 3: there is an energy in gesture intermediate-freuqncy signal x (t) due to the coupling phenomenon of transmitting antenna and receiving antenna
Very big low frequency signal, using wavelet thresholding methods to above-mentioned intermediate-freuqncy signal SNI\Q(t) it is pre-processed, is selected a kind of improved
Threshold function table, and after only being processed to low frequency coefficient, reconstruct obtains new intermediate-freuqncy signal
Step 4: to intermediate-freuqncy signalHandled, estimate gesture motion when m- distance spectrum figure, when m- normal-moveout spectrum
Figure and when m- angular spectrum figure, and numerical value normalized is carried out respectively to three kinds of spectrograms;
Step 5: after being normalized in step 4 when m- distance spectrum figure, when m- normal-moveout spectrum figure and when m- angular spectrum figure carry out
Splicing constructs Multivariate characteristics spectrogram A;
Step 6: carrying out the operation of step 2-step 6 respectively to collected N*ClassNum group gesture echo data, obtain
The original hand signal Multivariate characteristics spectrogram collection of N*ClassNum group
Step 7: the Multivariate characteristics spectrogram collection that step 6 is obtainedEach of sample standard deviation carry out gray processing processing, obtain
Original Multivariate characteristics atlas B;
Step 8: carrying out mean value to all samples in original Multivariate characteristics atlas B and carrying out dimension normalization, obtain more
Memberization feature atlasAnd it labels to each width characteristic pattern;
Step 9: by Multivariate characteristics atlasIt is divided into training set S according to a certain percentagetrain, verifying collection SvalAnd test set
Stest, such as training set accounts for 70%, verifying collection accounts for 20%, test set and accounts for 10%;
Step 10: by Strain、SvalInput data set C of the label corresponding with its together as convolutional neural networksinput, and just
Beginningization network weight, wherein StrainFor being trained to net coefficients, SvalNetwork verification is carried out after training a period of time,
And it is adjusted by weight of the error back propagation to network;
Step 11: by input data set CinputA convolution pondization operation is carried out, the scale kernel_ of convolution kernel is set
Size1, convolution step-length kernel_stride1 and pond size pool_size1, pond step-length pool_stride1, obtain
Feature atlas feature1;
Step 12: extracting further feature to feature atlas feature1 further progress convolution pond, the ruler of convolution kernel being arranged
Very little kernel_size2, convolution step-length kernel_stride2 and pond size pool_size2, pond step-length pool_
Stride2 obtains feature atlas feature2;
Step 13: carrying out a convolution pondization operation again to feature atlas feature2, deeper time feature, setting volume are extracted
Size kernel_size3, the convolution step-length kernel_stride3 and pond size pool_size3, pond step-length of product core
Pool_stride3 obtains feature atlas feature3;
Step 14: feature3 is passed sequentially through full articulamentum fc4, fc5, fc6, and the ruler of fc4, fc5, fc6 is respectively set
Very little size is size4, size5, size6, and feature atlas is converted into the column vector v1 of 1 × size6;
Step 15: output is different gesture classifications by column vector v1 by softmax classifier, by successive ignition,
Network accuracy rate and loss function tend towards stability, and obtain trained convolutional neural networks model Netmodel.
Step 16: by test data set StestIt is loaded into Netmodel, obtains gesture classification result y.
2. a kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING according to claim 1, feature exist
In: the preprocess method in the step 3 specifically includes:
Step 3-1 is directed to the biggish low frequency signal of energy due to transmitting antenna and receiving antenna coupling generation, changes with one kind
Into wavelet threshold function be removed, the threshold function table is as shown in Equation 1:
Wherein
Wj,k(m, n)=m ωj,k+(1-m)sgn(ωj,k)(|ωj,k|-n)
As can be seen that improved threshold function table is continuous at λ and-λ, while also the value of adjustable α and β adapts to this method
The demand of a variety of noise scenarios;
Step 3-2 is directed to the characteristics of interference signal, to gesture intermediate-freuqncy signalAfter carrying out N layers of wavelet decomposition, approximation is obtained
Coefficient A (i), detail coefficients D (1, i), D (2, i) ..., D (N, i), only to its low frequency coefficient A (i) carry out threshold process, i.e.,
Ai'=mAi+(1-m)sgn(Ai)(|Ai|-n)
And then it uses againD (1, i), D (2, i) ..., D (N, i) is reconstructed to get to pretreated signal x'(t).
3. a kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING according to claim 1, feature exist
In: the building of Multivariate characteristics spectrogram A in the step 5 specifically includes:
Step 5-1: extraction time-is apart from spectrogram: making FFT in fast time-domain first, then weighting asks flat in slow time-domain
, it finally carries out interframe and accumulates m- distance spectrum figure, size FFTNum1*FrameNum when can be obtained;
Step 5-2: extraction time-speed spectrogram: carrying out Two-dimensional FFT first as unit of frame, obtains distance-Doppler figure, connects
Detection range-Dopplergram maximum value, and extract the maximum value and be expert at, believed using the velocity information of the row as present frame
Number speed, finally m- normal-moveout spectrum figure, size FFTNum2*FrameNum when interframe is accumulated to obtain;
Step 5-3: extraction time-angle spectrogram: first as unit of frame, the correspondence frame signal of N number of receiving antenna is done respectively
Two-dimensional FFT obtains N range from-Dopplergram.Is extracted by maximum value respectively, and maximum value is arranged from-Dopplergram for N range
At the one-dimensional vector of 1 × N, FFT is done to the one-dimensional vector, the corresponding velocity information of current frame signal can be obtained, finally in frame
Between carry out accumulation m- angle figure, size FFTNum3*FrameNum when can be obtained;
Step 5-4: building Multivariate characteristics spectrogram A: three kinds of feature spectrogram columns having the same that the above method extracts, three
After kind spectrogram is normalized respectively, the feature spectrogram after normalization, which is carried out splicing with row, can be obtained Multivariate characteristics spectrum
Figure, size are (FFTNum1+FFTNum2+FFTNum3) * FrameNum.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910445702.3A CN110348288B (en) | 2019-05-27 | 2019-05-27 | Gesture recognition method based on 77GHz millimeter wave radar signal |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910445702.3A CN110348288B (en) | 2019-05-27 | 2019-05-27 | Gesture recognition method based on 77GHz millimeter wave radar signal |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110348288A true CN110348288A (en) | 2019-10-18 |
CN110348288B CN110348288B (en) | 2023-04-07 |
Family
ID=68174720
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910445702.3A Active CN110348288B (en) | 2019-05-27 | 2019-05-27 | Gesture recognition method based on 77GHz millimeter wave radar signal |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110348288B (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110687816A (en) * | 2019-10-31 | 2020-01-14 | 复旦大学 | Intelligent household control system and method based on millimeter wave radar |
CN110988863A (en) * | 2019-12-20 | 2020-04-10 | 北京工业大学 | Novel millimeter wave radar gesture signal processing method |
CN111461037A (en) * | 2020-04-07 | 2020-07-28 | 电子科技大学 | End-to-end gesture recognition method based on FMCW radar |
CN111929674A (en) * | 2020-07-10 | 2020-11-13 | 西安电子科技大学 | Intelligent amplitude comparison angle measurement method based on neural network, storage medium and equipment |
CN112148128A (en) * | 2020-10-16 | 2020-12-29 | 哈尔滨工业大学 | Real-time gesture recognition method and device and man-machine interaction system |
CN112183279A (en) * | 2020-09-21 | 2021-01-05 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN112198966A (en) * | 2020-12-08 | 2021-01-08 | 中南大学 | Stroke identification method and system based on FMCW radar system |
CN112446348A (en) * | 2020-12-08 | 2021-03-05 | 电子科技大学 | Behavior identification method based on characteristic spectrum flow |
CN113408328A (en) * | 2020-03-16 | 2021-09-17 | 哈尔滨工业大学(威海) | Gesture segmentation and recognition algorithm based on millimeter wave radar |
CN113406610A (en) * | 2021-06-16 | 2021-09-17 | 深圳大学 | Target detection method, device, equipment and storage medium |
CN113567950A (en) * | 2021-09-24 | 2021-10-29 | 巍泰技术(武汉)有限公司 | Millimeter wave radar distance and velocity spectrum estimation method and system |
CN114114223A (en) * | 2021-12-01 | 2022-03-01 | 长春理工大学 | Millimeter wave sitting posture detection intelligent desk lamp based on convolutional neural network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106055089A (en) * | 2016-04-27 | 2016-10-26 | 深圳市前海万象智慧科技有限公司 | Control system for gesture recognition based on man-machine interaction equipment and control method for same |
CN106295684A (en) * | 2016-08-02 | 2017-01-04 | 清华大学 | A kind of the most continuous based on micro-Doppler feature/discontinuous gesture recognition methods |
CN108509910A (en) * | 2018-04-02 | 2018-09-07 | 重庆邮电大学 | Deep learning gesture identification method based on fmcw radar signal |
CN109271838A (en) * | 2018-07-19 | 2019-01-25 | 重庆邮电大学 | A kind of three parameter attributes fusion gesture identification method based on fmcw radar |
-
2019
- 2019-05-27 CN CN201910445702.3A patent/CN110348288B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106055089A (en) * | 2016-04-27 | 2016-10-26 | 深圳市前海万象智慧科技有限公司 | Control system for gesture recognition based on man-machine interaction equipment and control method for same |
CN106295684A (en) * | 2016-08-02 | 2017-01-04 | 清华大学 | A kind of the most continuous based on micro-Doppler feature/discontinuous gesture recognition methods |
CN108509910A (en) * | 2018-04-02 | 2018-09-07 | 重庆邮电大学 | Deep learning gesture identification method based on fmcw radar signal |
CN109271838A (en) * | 2018-07-19 | 2019-01-25 | 重庆邮电大学 | A kind of three parameter attributes fusion gesture identification method based on fmcw radar |
Non-Patent Citations (2)
Title |
---|
ZHANFENG ZHAO: "Feature Extraction and Identification of Pipeline Intrusion Based on Phase-Sensitive Optical Time Domain Reflectometer" * |
张佳俊: "基于5GHz频段雷达的手势识别***研究" * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110687816A (en) * | 2019-10-31 | 2020-01-14 | 复旦大学 | Intelligent household control system and method based on millimeter wave radar |
CN110988863A (en) * | 2019-12-20 | 2020-04-10 | 北京工业大学 | Novel millimeter wave radar gesture signal processing method |
CN113408328B (en) * | 2020-03-16 | 2023-06-23 | 哈尔滨工业大学(威海) | Gesture segmentation and recognition algorithm based on millimeter wave radar |
CN113408328A (en) * | 2020-03-16 | 2021-09-17 | 哈尔滨工业大学(威海) | Gesture segmentation and recognition algorithm based on millimeter wave radar |
CN111461037A (en) * | 2020-04-07 | 2020-07-28 | 电子科技大学 | End-to-end gesture recognition method based on FMCW radar |
CN111929674A (en) * | 2020-07-10 | 2020-11-13 | 西安电子科技大学 | Intelligent amplitude comparison angle measurement method based on neural network, storage medium and equipment |
CN112183279B (en) * | 2020-09-21 | 2022-06-10 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN112183279A (en) * | 2020-09-21 | 2021-01-05 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN112148128A (en) * | 2020-10-16 | 2020-12-29 | 哈尔滨工业大学 | Real-time gesture recognition method and device and man-machine interaction system |
CN112148128B (en) * | 2020-10-16 | 2022-11-25 | 哈尔滨工业大学 | Real-time gesture recognition method and device and man-machine interaction system |
CN112198966B (en) * | 2020-12-08 | 2021-03-16 | 中南大学 | Stroke identification method and system based on FMCW radar system |
CN112446348B (en) * | 2020-12-08 | 2022-05-31 | 电子科技大学 | Behavior identification method based on characteristic spectrum flow |
CN112446348A (en) * | 2020-12-08 | 2021-03-05 | 电子科技大学 | Behavior identification method based on characteristic spectrum flow |
CN112198966A (en) * | 2020-12-08 | 2021-01-08 | 中南大学 | Stroke identification method and system based on FMCW radar system |
CN113406610A (en) * | 2021-06-16 | 2021-09-17 | 深圳大学 | Target detection method, device, equipment and storage medium |
CN113406610B (en) * | 2021-06-16 | 2023-06-23 | 深圳大学 | Target detection method, device, equipment and storage medium |
CN113567950A (en) * | 2021-09-24 | 2021-10-29 | 巍泰技术(武汉)有限公司 | Millimeter wave radar distance and velocity spectrum estimation method and system |
CN113567950B (en) * | 2021-09-24 | 2021-12-17 | 巍泰技术(武汉)有限公司 | Millimeter wave radar distance and velocity spectrum estimation method and system |
CN114114223A (en) * | 2021-12-01 | 2022-03-01 | 长春理工大学 | Millimeter wave sitting posture detection intelligent desk lamp based on convolutional neural network |
Also Published As
Publication number | Publication date |
---|---|
CN110348288B (en) | 2023-04-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110348288A (en) | A kind of gesture identification method based on 77GHz MMW RADAR SIGNAL USING | |
CN108226892B (en) | Deep learning-based radar signal recovery method in complex noise environment | |
CN107122738B (en) | Radio signal identification method based on deep learning model and implementation system thereof | |
US20210003697A1 (en) | Method and apparatus for end-to-end sar image recognition, and storage medium | |
CN111626961B (en) | Radar image clutter suppression method and system based on generation countermeasure network | |
CN109188414A (en) | A kind of gesture motion detection method based on millimetre-wave radar | |
CN111461037B (en) | End-to-end gesture recognition method based on FMCW radar | |
CN106355151A (en) | Recognition method, based on deep belief network, of three-dimensional SAR images | |
CN113837131B (en) | Multi-scale feature fusion gesture recognition method based on FMCW millimeter wave radar | |
Ni et al. | Human identification based on natural gait micro‐Doppler signatures using deep transfer learning | |
Wu et al. | Dynamic hand gesture recognition using FMCW radar sensor for driving assistance | |
Zhao et al. | Cubelearn: End-to-end learning for human motion recognition from raw mmwave radar signals | |
CN103473559A (en) | SAR image change detection method based on NSCT domain synthetic kernels | |
CN113408328A (en) | Gesture segmentation and recognition algorithm based on millimeter wave radar | |
Wang et al. | Multi-classification of UWB signal propagation channels based on one-dimensional wavelet packet analysis and CNN | |
CN116824629A (en) | High-robustness gesture recognition method based on millimeter wave radar | |
CN114428234A (en) | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention | |
Qu et al. | Human activity recognition based on WRGAN-GP-synthesized micro-doppler spectrograms | |
CN115343704A (en) | Gesture recognition method of FMCW millimeter wave radar based on multi-task learning | |
CN116109925A (en) | Multi-mode remote sensing image classification method based on heterogeneous feature learning network | |
CN115240040A (en) | Method and device for enhancing human behavior characteristics of through-wall radar | |
CN113901931A (en) | Knowledge distillation model-based behavior recognition method for infrared and visible light videos | |
Sun et al. | Efficient parameter inference for gravitational wave signals in the presence of transient noises using temporal and time-spectral fusion normalizing flow | |
CN112241001A (en) | Radar human body action recognition method and device, electronic equipment and storage medium | |
CN116524537A (en) | Human body posture recognition method based on CNN and LSTM combination |
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 |