CN108664894A - The human action radar image sorting technique of neural network is fought based on depth convolution - Google Patents
The human action radar image sorting technique of neural network is fought based on depth convolution Download PDFInfo
- Publication number
- CN108664894A CN108664894A CN201810317164.5A CN201810317164A CN108664894A CN 108664894 A CN108664894 A CN 108664894A CN 201810317164 A CN201810317164 A CN 201810317164A CN 108664894 A CN108664894 A CN 108664894A
- Authority
- CN
- China
- Prior art keywords
- radar
- data
- radar image
- image
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
Abstract
The present invention relates to a kind of human action radar image sorting techniques for fighting neural network based on depth convolution, including：Build data set；Radar image data enhancing is realized by DCGAN：Establish DCGAN networks, individually learn each radar spectrogram using network, new radar spectrogram is generated according to the feature that network is acquired, in the case where data volume is certain, expand training set sample, passes through network adjustment parameter so that failed regeneration image is minimum, EDS extended data set to greatest extent realizes data enhancing；Extract the envelope of upper, middle and lower three of radar image, as feature vector, upper lower envelope representative body four limbs echo radial velocity, the echo radial velocity of intermediate envelope representative body trunk, using three feature vectors as the input of support vector machine classifier, classified to radar image data using support vector machines.
Description
Technical field
The invention belongs to human action Activity recognition, Radar Targets'Detection, data enhancing, depth convolution to fight neural network
(DCGAN, Deep Convolutional Generative Adversarial Networks) and machine learning field, is related to
Feature extraction to radar image and using DCGAN carry out data enhancing go forward side by side pedestrian's body classification of motion the problem of.
Background technology
Human action Activity recognition [1] is one research hotspot of computer vision field in recent years, is widely used in people
The fields such as machine interaction, virtual reality and video monitoring, although the research of domestic and international human action Activity recognition obtained in recent years
Many progress, but the high complexity of human motion and more variabilities to identify that high efficiency is each from fully meeting with accuracy
The related request of a industry.The difficult point of human action Activity recognition essentially consists in the complexity in space and the otherness of time.It is empty
Between complexity include different visual angle, background, the action scene problem of illumination, human action in different directions, angle not
Same amplitude problem, the mutual occlusion issue between person to person, people and object；The time difference opposite sex includes that can not determine human action
Time point problem, judge action effect effective time and interval problem, the white space problem occurred in action.The two
Problem makes the ununified effective frame of human action Activity recognition research field, the relevant technologies and unified effective analysis point
Class method.Therefore the present invention uses Doppler radar as a sensor to detection human action behavior.
With the decline of the rising and use cost of Doppler radar [2] precision, detecting human body target using it becomes
New research hotspot.The echo that the signal of radar emission is radiated at moving target generation contains relatively rich general Le frequency letter
Breath, while the relative motion of parts of body will produce complicated micro-doppler frequency displacement (Micro-Doppler), micro-doppler frequency
It includes abundant movable information to move.Therefore there is prodigious development space using Doppler radar detection human action behavior.It utilizes
Doppler radar detection human body target can be widely applied to many aspects such as disaster assistance, security protection, national defense construction.It connects
The echo-signal that receipts device receives can be by Short Time Fourier Transform (ShortTime FourierTransform, STFT)
Obtained radar spectrogram, then suitable classification tool is selected to classify.
Support vector machines [3] (supportvectormachine, SVM) is a kind of classics formally delivered in nineteen ninety-five
The criteria for classification of two disaggregated models, support vector machines is returned originating from Logistic, and basic model is defined as on feature space
The maximum linear classifier in interval.Because the excellent performance of its showing in text classification of task is quickly become machine learning
Mainstream technology.The kernel function (kernel function) that statistical learning is brought is so that support vector machines has powerful life
Power, kernel function directly determines the final performance of support vector machines, but the selection of kernel function is always an open question, is compared
General kernel function includes：Linear kernel, polynomial kernel, Gaussian kernel etc..
Production fights network[4](GenerativeAdversarialNets, GAN) proposed just to have obtained wide from 2014
General concern, model are mainly made of a generator and an arbiter.Production fights generator and differentiation in network
Confronting with each other between device so that the data distribution infinite approach truthful data of output is distributed, and GAN is provided to numerous researchers
New training thinking, has greatly pushed the development of artificial intelligence.
[1] Li Ruifeng, Wang Liangliang, Wang Ke (2014) human action Activity recognition Review Study pattern-recognitions with it is artificial
Intelligence, 27 (1), 35-48.
[2]Chen,V.C.(2000).Analysis of radar micro-Doppler with time-
frequency transform.Statistical Signal and Array Processing,2000.Proceedings
ofthe TenthIEEE Workshop on(pp.463-466).IEEE.
[3]Ukil,A.(2002).Support vector machine.ComputerScience,1(4),1-28.
[4]Goodfellow,I.J.,Pouget-Abadie,J.,Mirza,M.,Xu,B.,Warde-Farley,D.,&
Ozair,S.,et al.(2014).Generative adversarial nets.International Conference on
NeuralInformation Processing Systems(Vol.3,pp.2672-2680).MIT Press.
The present invention, which reaches production confrontation network with the classification of radar human action, to be combined, for security protection, military affairs prison
Control, the real works such as fire-fighting and rescue, which have, greatly to help.
Invention content
The object of the present invention is to provide one kind being directed to the insufficient practical problem of radar data amount, proposes a kind of solid
The method is simultaneously applied to human action radar image sorting technique by data enhancement methods.The present invention using support vector machines as point
Class device classifies to seven kinds of human action behavior radar datas, it is contemplated that the situation of radar image data amount deficiency, the present invention
Data enhancing is carried out using production confrontation type network.Technical solution is as follows：
A kind of human action radar image sorting technique for being fought neural network based on depth convolution, is included the following steps：
(1) data set is built；The data that human body behavior act is acquired using optical motion capture device build data set, the number
It is calculated under elliposoidal manikin according to the radar return of collection.Radar return obtains Radar Spectrum by Short Time Fourier Transform
Figure；
(2) radar image data enhancing is realized by DCGAN：DCGAN networks are established, individually learn each using network
Radar spectrogram generates new radar spectrogram according to the feature that network is acquired, and in the case where data volume is certain, expands training set sample
This, passes through network adjustment parameter so that failed regeneration image is minimum, to greatest extent EDS extended data set, realizes data enhancing；
(3) envelope of upper, middle and lower three for extracting radar image, as feature vector, upper lower envelope representative body four limbs return
Wave radial velocity, the echo radial velocity of intermediate envelope representative body trunk, using three feature vectors as support vector machines point
The input of class device classifies to radar image data using support vector machines；
(4) radar image for generating DCGAN is added as enhancing data in training set, trained by support vector machines
To disaggregated model.
The characteristics of present invention can independently realize unsupervised learning according to production confrontation network proposes a kind of suitable for carrying
The data enhancement method of high radar image recognition effect improves the accuracy rate of Human bodys' response.The present invention is to be based on MOCAP
The Radar Doppler image of data set generation and be research object by the enhanced image of data, including the structure of data set with
Enhancing, foundation, model training and the test of production confrontation network.The advantages of present invention is according to radar system, for radar number
According to the insufficient objective condition of amount, it is proposed that a kind of novel data enhancement methods, and then make the human action based on radar image
Classification accuracy is improved.
Description of the drawings
Fig. 1 is DCGAN structural models.
Fig. 2 is generator structural model.
Fig. 3 is human body ellipsoidal model.
Fig. 4 is the radar spectral image of seven kinds of actions.
Specific implementation mode
To keep technical scheme of the present invention clearer, the specific embodiment of the invention is further described through below.
The present invention implements according to the following steps：
1. radar time-frequency image data set is built
The present invention uses the MOCAP data sets established by Carnegie Mellon University graph experiment room.The data set is according to people
Body spheroid action model gathered data, the model be originated from Boulic body gait models, Boulic nineteen ninety propose one
A whole world body gait model, the model model human body target echo, and human body can be divided into ten scattering positions, be respectively
Head, thoracic cavity, left large arm, right large arm, left forearm, right forearm, left thigh, right thigh, left leg and right leg.Different limbs
Movement has different curve movement equations, the adduction of echo shaping, that is, all different limb motion situations of human body.This ten dissipate
It penetrates the used shape of position modeling and respective relevant parameter value is specifically as shown in table 1.
Table 1：Human body scattering part ranks table
| Scatter position | Shape | Length symbol | Value/m | Radius designation | Value/m |
| Head | Sphere | -- | -- | Rhe | 0.20 |
| Trunk | Spheroid | Hto | 0.80 | Rto | 0.25 |
| Upper arm | Spheroid | Hua | 0.45 | Rua | 0.05 |
| Forearm | Spheroid | Hla | 0.45 | Rla | 0.04 |
| Thigh | Spheroid | Hul | 0.50 | Rul | 0.10 |
| Shank | Spheroid | Hll | 0.50 | Rll | 0.07 |
The RCS calculation formula of circle and spheroid are as follows：
σ=π R2 (1)
The shape and radius at ten scattering positions of human body are provided by table 1, are calculated respectively in connection with the RCS of ellipsoid and sphere public
Formula can calculate the radar scattering area of human body different parts.For the transmitting signal under single-frequency continuous wave radar system, form
For sin (2 π f0T), form of the body gait radar echo signal after I/Q quadrature demodulations is：
Wherein i=1 ..., 10 indicate that corresponding different scattering positions, k are indicated and returned in ten point scattering site model of human body
The relevant coefficient of intensity of wave, σiIndicate the RCS, τ at each scattering position of bodyi(t) indicate that the echo at each scattering position of body prolongs
Late, e (t) indicates the sum of ten scattering position radar returns of human body.
Human body ellipsoidal model is as shown in figure 3, entire manikin is made of multiple ellipsoids, each spheroid radar reflection
Wave-amplitude can be by being approximately that the RCS of ellipse is obtained.Data acquisition device is that the movement developed by ViconIndustries is caught
System is caught, which is made of 12 infrared ray MX40 video cameras, and each video camera can be with 120Hz frame speed recordings
Image, while the system represents human body various pieces with 41 mark points, it can be by partes corporis humani's point when collecting data
Movement is reduced to movement a little.The data set contains total 2605 tests movement including six kinds of movement scenarios.This six kinds
Movement scenarios are in interpersonal interactive, between man and nature interaction, sports, autogenic movement, movement respectively
Scene change and test activity.It is total to acquire 2605 groups of experimental datas, select wherein seven kinds of common actions in process of the present invention
For generating radar image, this seven kinds actions are respectively：It boxes, crawl, creep, jump, run, walk, stand.Then it utilizes
Short Time Fourier Transform (STFT, Short-timeFourierTransform) is handled from the reflected thunder of human body various pieces
Radar spectrogram is obtained up to echo.The appearance of Short Time Fourier Transform is intended to solve the time domain and frequency localization lance of signal
Shield, basic thought are：Fourier transformation, is carried out certain change by local time-domain information in order to obtain, is carried out in signal
The window function of a finite time length is multiplied by before Fourier transformation, can give tacit consent to stationary signal has in window function in limit
It is stable, window function moves on a timeline, is converted paragraph by paragraph to signal, finally obtains " part " of signal different moments
Frequency spectrum.
The very short window function η (t) of a time span is given, signal to be analyzed is s (t), then signal s's (t) is short
When Fourier transformation STFT be defined as：
The selection of window function η (t) has larger impact to the performance of Short Time Fourier Transform, and window function η (t) is shorter, in short-term
Fourier transformation temporal resolution is higher；On the contrary, window function η (t) time widths are longer, Short Time Fourier Transform frequency resolution
It is higher.
The characteristics of present invention is according to radar spectrogram obtains data set, for each being moved in classification task by " sliding window method "
Make to can get the data set that size is 500 pictures, the data set of each action is divided into two parts by the present invention, respectively
For 400 training sets and 100 test sets.
2. the radar image data enhancing based on DCGAN
Confrontation network is made of a discrimination model and generation model.Network structure is substantially as shown in Figure 1.Compared to volume
Code device (Auto Encoder) or deconvolution neural network, DCGAN can more preferably generate image faster.DCGAN is to convolution
The structure of neural network has carried out some changes, for example removes all pond layers, is sampled with warp lamination, removes and connect entirely
Layer is connect, network is made to become full convolutional layer structure.
The network structure of DCGAN generators as shown in Fig. 2, carry out image preprocessing first, and input picture and label are one by one
It is corresponding, therefore input data can be considered and be uniformly distributed, the label for obeying equally distributed input sample and input cascades, as one
A entirety inputs network, and data dimension is 100, and dimension is become 1024 data by the advanced row linear transformation of first convolutional layer,
It is non-using line rectification function (RectifiedLinear Unit, ReLU) progress nonlinear transformation acquisition first after normalization
The output of linear layer, then cascade the input as next layer with the label of input.Second convolution will be inputted by linear transformation
Data become 512 dimension datas, to carrying out linear R eLU transformation after the normalization of its block, then reshape are needed to obtain second
The output of a non-linear layer.Data and label are finally cascaded to the input as next layer.It is finally obtained by four convolutional layers
The image of one 64*64*3, finally passes through a warp lamination, and the effect of the layer network is that data are carried out with the reverse behaviour of convolution
Make, that is, by the output signal Jing Guo convolution, the input signal of convolution can be restored by deconvolution.Warp lamination does not do block
Normalization operation directly carries out non-linear Sigmoid transform, generates image.Obtained image in arbiter with true picture
It compares, arbiter returns to penalty values according to loss function, and constantly correction generates image so that generates image and becomes closer to very
Real image.The arbiter network structure of DCGAN is similar with generator, is made of five layers of convolutional layer.
The parameter setting of DCGAN is most important for the generation of final image, is iterations epoch first, with
Epoch is gradually increased since 1, is generated image and is become closer to true picture, but when epoch is excessive, output image with
The similarity of true picture can reduce, and experiment display, epoch values generate image and input picture when between 350 to 600
It is roughly the same, but still have parts of images failed regeneration, failure image is differed with the training set sample of input from range estimation
Too much, these image classification effects, which are added, can also decline, it is therefore desirable to remove these images, the generation image that will finally choose
Training sets are added according to seven kinds of classification of motion, expand the training set sample present invention epoch values be 350,400,450,500,
550 and 600 generate image respectively, and when wherein epoch values are 500, failed regeneration image is minimum, therefore according to this hair of experimental conditions
Bright that the value of epoch is set as 500, the amount of images of learning rate 0.0002, each iteration is 1, the height of input picture and
Wide is 120.Output image is dimensioned to 128.Since the image size of original data set is 120, as training set
The operation for being first uniformly adjusted scale is needed before input support vector machines.Design parameter is as shown in table 2.
Table 2：DCGAN parameter settings
3. the radar image classification based on support vector machines
(1) support vector machines
The grader of the present invention uses support vector machines (SVM).In machine learning, support vector machines is the mid-90
A kind of machine learning method based on Statistical Learning Theory to grow up improves study by seeking structuring least risk
Machine generalization ability realizes the minimum of empiric risk and fiducial range, to reach in the case where statistical sample amount is less, also
The purpose of good statistical law can be obtained.One group of training sample is given, each label is two classes, and support vector machines training is calculated
Method establishes a model, and it is a kind of or other classes to distribute new example, becomes non-probability binary linearity classification.Support to
Chance is measured by the point on the maps feature vectors to two dimensional surface in image, a most robust, generalization ability are found by algorithm
Strongest line of demarcation is separated by two class data.When two dimensional surface cannot meet the requirement of linear classification, support vector machines
Data can be mapped by kernel function to higher dimensional space, a suitable hyperplane is found in higher dimensional space to realize
To the linear classification of data.
More classification problems are realized in order to make the support vector machines of binary classifier, present invention employs decision tree structure,
Decision tree is a kind of tree structure, wherein each internal node indicates that the test on an attribute, each branch represent a survey
Examination output, each leaf node represent a kind of classification.So, more classification problems are just converted into multiple two classification problem, this hair
Bright human body behavior act needs to be divided into seven classes, that is, needs application decision tree construction, carries out five two classification problems.
The kernel function of support vector machines, which is chosen, to need to seek by testing, because training sample is usually independently to go out
Existing, they always occur in the form of the inner product of pairs of sample, and inner product is substituted by using appropriate kernel function, can be implicit
Nonlinear training data is mapped to higher dimensional space, without increase adjustable parameter number.The selection master of kernel function at present
If several common kernel functions, such as Polynomial kernel function, gaussian kernel function, linear kernel function and Radial basis kernel function.
Linear kernel function is found through experiments that, the classification accuracy of Radial basis kernel function is less than Polynomial kernel function, and gaussian kernel function
It can not classify Deng other kernel functions, therefore the present invention has chosen the best Polynomial kernel function of classifying quality.Polynomial kernel function
Mathematic(al) representation be：
K (x, xi)=(xxi+1)dD=1,2 ..., N (5)
Polynomial kernel function application is more flexible, xx in the present inventioniTwo vectorial inner products are represented for two variables.d
Representation vector dimension, dimension of the invention are 360.
(2) feature extraction
The behavior act of human body is a time-varying, non-stable random process, and by each different parts of body
Movement composition.The motion model of body different parts, the characteristics of motion are different, therefore the radar return of body gait signal is
Time-varying, complicated, spectrum component is abundant, so the present invention needs to be moved to obtain human body behavior with Short Time Fourier Transform
Make the spectrogram of echo.
Radar image used in the present invention is as shown in figure 4, human body behavior act signal obtains after doing Short Time Fourier Transform
It is arriving as a result, abscissa be the time, ordinate is radial velocity, and RED sector is backward energy the best part, i.e. human body body
Stem portion is with respect to the Doppler frequency that radar radial direction treads are come, and Doppler frequency changes caused by being moved except trunk,
The limbs such as those energy are slightly weak, have the curve of periodic undulations to be arm during body gait moves, leg is dry are micro- more caused by swinging
General Le information.Frequency spectrum complicated component in figure, but can be seen that substantially body gait movement period and body gait movement in due to
Doppler frequency caused by trunk body moves, to which the kinematic parameters such as body gait frequency, walking speed can be obtained.
The present invention extracts the envelope of radar image as feature, Fig. 4 lists seven kinds respectively according to radar image feature
The radar spectral image of action.Its horizontal axis indicates that time, the longitudinal axis indicate that radial velocity, color indicate echo strength.Every chart
Up to the behavior act different in 1 second human body.The partially red straight line of intermediate colors indicates the radar return of trunk, and both sides are then
The radar return of four limbs.When radar detection human body behavior act, ignore the radar because of displacement distance and environmental factor generation
RL return loss, it is certain that can be approximately considered total radar echo intensity that each time point human body is reflected back, therefore can be with
The radial velocity of the echo of human limb and trunk echo is showed by parameter setting, it is of the invention by parameter according to experiment
It is set as 0.28,0.5 and 0.77, three envelopes of upper, middle and lower of radar spectral image can be extracted by the parameter setting, up and down
Envelope represents human limb echo radial velocity, the echo radial velocity of intermediate envelope representative body trunk.By these three features
Input of the vector as support vector machine classifier.
4. training pattern and testing classification accuracy rate
The human body behavior act of the data set of the present invention includes altogether 7 classes, respectively boxes, crawls, creeps, jumps, runs
Step, walking, seven classes of standing.The present invention first acts 7 classes, and per class, action is acted comprising 400 images as training set per class
100 images are as test set.Using polynomial kernel as kernel function, it is trained to obtain model using support vector machines, most
The detection model on test set afterwards.The classification accuracy of the model has reached 80.5714%.
Using above-mentioned experiment as benchmark, per class, the training set of action increases by 400 DCGAN generation images, and test set is kept
Constant, kernel function is Polynomial kernel function, trains to obtain model by support vector machines, finally in same test collection testing classification
As a result.Its classification accuracy is 82.4286%.2 percentage points are improved compared to benchmark.
Claims (1)
1. a kind of human action radar image sorting technique for being fought neural network based on depth convolution, is included the following steps：
(1) data set is built；The data that human body behavior act is acquired using optical motion capture device build data set, the data set
Radar return be calculated under elliposoidal manikin.Radar return obtains radar spectrogram by Short Time Fourier Transform；
(2) radar image data enhancing is realized by DCGAN：DCGAN networks are established, individually learn each radar using network
Spectrogram generates new radar spectrogram according to the feature that network is acquired, and in the case where data volume is certain, expands training set sample,
Pass through network adjustment parameter so that failed regeneration image is minimum, to greatest extent EDS extended data set, realizes data enhancing；
(3) envelope of upper, middle and lower three for extracting radar image, as feature vector, upper lower envelope representative body four limbs echo diameter
To speed, the echo radial velocity of intermediate envelope representative body trunk, using three feature vectors as support vector machine classifier
Input, using support vector machines to radar image data classify；
(4) radar image for generating DCGAN is trained point as in enhancing data addition training set by support vector machines
Class model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810317164.5A CN108664894A (en) | 2018-04-10 | 2018-04-10 | The human action radar image sorting technique of neural network is fought based on depth convolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810317164.5A CN108664894A (en) | 2018-04-10 | 2018-04-10 | The human action radar image sorting technique of neural network is fought based on depth convolution |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108664894A true CN108664894A (en) | 2018-10-16 |
Family
ID=63783221
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810317164.5A Pending CN108664894A (en) | 2018-04-10 | 2018-04-10 | The human action radar image sorting technique of neural network is fought based on depth convolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108664894A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109331441A (en) * | 2018-11-19 | 2019-02-15 | 吉林师范大学 | A kind of callisthenics body rectificative training device and training method |
| CN109711538A (en) * | 2018-12-14 | 2019-05-03 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN109740729A (en) * | 2018-12-14 | 2019-05-10 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN109753150A (en) * | 2018-12-11 | 2019-05-14 | 北京字节跳动网络技术有限公司 | Figure action control method, device, storage medium and electronic equipment |
| CN110210371A (en) * | 2019-05-29 | 2019-09-06 | 华南理工大学 | A kind of aerial hand-written inertia sensing signal creating method based on depth confrontation study |
| CN110414426A (en) * | 2019-07-26 | 2019-11-05 | 西安电子科技大学 | A kind of pedestrian's Approach for Gait Classification based on PC-IRNN |
| CN110516552A (en) * | 2019-07-29 | 2019-11-29 | 南京航空航天大学 | A kind of multipolarization radar image classification method and system based on timing curve |
| CN110610207A (en) * | 2019-09-10 | 2019-12-24 | 重庆邮电大学 | Small sample SAR image ship classification method based on transfer learning |
| CN111008650A (en) * | 2019-11-13 | 2020-04-14 | 江苏大学 | Metallographic structure automatic rating method based on deep convolution countermeasure neural network |
| CN112241001A (en) * | 2020-10-10 | 2021-01-19 | 深圳大学 | Radar human body action recognition method and device, electronic equipment and storage medium |
| CN112433207A (en) * | 2020-11-06 | 2021-03-02 | 浙江理工大学 | Human body identity recognition method based on two-channel convolutional neural network |
| CN112529806A (en) * | 2020-12-15 | 2021-03-19 | 哈尔滨工程大学 | SAR image data enhancement method based on generation of countermeasure network information maximization |
| CN113191268A (en) * | 2021-04-30 | 2021-07-30 | 中山大学 | SAR target recognition countermeasure sample generation method based on deep coding network |
| CN113238222A (en) * | 2021-05-13 | 2021-08-10 | 天津大学 | Human body action recognition method based on envelope density characteristics |
| CN113296087A (en) * | 2021-05-25 | 2021-08-24 | 沈阳航空航天大学 | Frequency modulation continuous wave radar human body action identification method based on data enhancement |
| US11403349B2 (en) * | 2018-11-09 | 2022-08-02 | Accenture Global Solutions Limited | Dark web content analysis and identification |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106483514A (en) * | 2016-09-23 | 2017-03-08 | 电子科技大学 | A kind of airplane motion mode identification method based on EEMD and SVMs |
| CN107169435A (en) * | 2017-05-10 | 2017-09-15 | 天津大学 | A kind of convolutional neural networks human action sorting technique based on radar simulation image |
| CN107490795A (en) * | 2017-07-24 | 2017-12-19 | 长沙学院 | It is a kind of to realize that human motion state knows method for distinguishing by radar |
-
2018
- 2018-04-10 CN CN201810317164.5A patent/CN108664894A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106483514A (en) * | 2016-09-23 | 2017-03-08 | 电子科技大学 | A kind of airplane motion mode identification method based on EEMD and SVMs |
| CN107169435A (en) * | 2017-05-10 | 2017-09-15 | 天津大学 | A kind of convolutional neural networks human action sorting technique based on radar simulation image |
| CN107490795A (en) * | 2017-07-24 | 2017-12-19 | 长沙学院 | It is a kind of to realize that human motion state knows method for distinguishing by radar |
Non-Patent Citations (3)
| Title |
|---|
| ALEC RADFORD ET AL.: "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks", 《ARXIV:1511.06434V2》 * |
| CONG, LONGJIAN ET AL.: "Convolutional neural network using generated data for SAR ATR with limited samples", 《MIPPR2017:PATTERN RECOGNITION AND COMPUTER VISION》 * |
| YOUNGWOOK KIM AND HAO LING: "Human Activity Classification Based on Micro-Doppler Signatures Using a Support Vector Machine", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11403349B2 (en) * | 2018-11-09 | 2022-08-02 | Accenture Global Solutions Limited | Dark web content analysis and identification |
| US11768866B2 (en) * | 2018-11-09 | 2023-09-26 | Accenture Global Solutions Limited | Dark web content analysis and identification |
| US20220342941A1 (en) * | 2018-11-09 | 2022-10-27 | Accenture Global Solutions Limited | Dark web content analysis and identification |
| CN109331441A (en) * | 2018-11-19 | 2019-02-15 | 吉林师范大学 | A kind of callisthenics body rectificative training device and training method |
| CN109753150A (en) * | 2018-12-11 | 2019-05-14 | 北京字节跳动网络技术有限公司 | Figure action control method, device, storage medium and electronic equipment |
| CN109711538A (en) * | 2018-12-14 | 2019-05-03 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN109740729A (en) * | 2018-12-14 | 2019-05-10 | 北京中科寒武纪科技有限公司 | Operation method, device and Related product |
| CN110210371A (en) * | 2019-05-29 | 2019-09-06 | 华南理工大学 | A kind of aerial hand-written inertia sensing signal creating method based on depth confrontation study |
| CN110210371B (en) * | 2019-05-29 | 2021-01-19 | 华南理工大学 | In-air handwriting inertial sensing signal generation method based on deep confrontation learning |
| CN110414426A (en) * | 2019-07-26 | 2019-11-05 | 西安电子科技大学 | A kind of pedestrian's Approach for Gait Classification based on PC-IRNN |
| CN110414426B (en) * | 2019-07-26 | 2023-05-30 | 西安电子科技大学 | Pedestrian gait classification method based on PC-IRNN |
| CN110516552A (en) * | 2019-07-29 | 2019-11-29 | 南京航空航天大学 | A kind of multipolarization radar image classification method and system based on timing curve |
| CN110610207B (en) * | 2019-09-10 | 2022-11-25 | 重庆邮电大学 | Small sample SAR image ship classification method based on transfer learning |
| CN110610207A (en) * | 2019-09-10 | 2019-12-24 | 重庆邮电大学 | Small sample SAR image ship classification method based on transfer learning |
| CN111008650A (en) * | 2019-11-13 | 2020-04-14 | 江苏大学 | Metallographic structure automatic rating method based on deep convolution countermeasure neural network |
| CN111008650B (en) * | 2019-11-13 | 2024-03-19 | 江苏大学 | Metallographic structure automatic grading method based on deep convolution antagonistic neural network |
| CN112241001A (en) * | 2020-10-10 | 2021-01-19 | 深圳大学 | Radar human body action recognition method and device, electronic equipment and storage medium |
| CN112241001B (en) * | 2020-10-10 | 2023-06-23 | 深圳大学 | Radar human body action recognition method, radar human body action recognition device, electronic equipment and storage medium |
| CN112433207A (en) * | 2020-11-06 | 2021-03-02 | 浙江理工大学 | Human body identity recognition method based on two-channel convolutional neural network |
| CN112529806A (en) * | 2020-12-15 | 2021-03-19 | 哈尔滨工程大学 | SAR image data enhancement method based on generation of countermeasure network information maximization |
| CN113191268A (en) * | 2021-04-30 | 2021-07-30 | 中山大学 | SAR target recognition countermeasure sample generation method based on deep coding network |
| CN113191268B (en) * | 2021-04-30 | 2024-04-23 | 中山大学 | SAR target recognition countermeasure sample generation method based on depth coding network |
| CN113238222A (en) * | 2021-05-13 | 2021-08-10 | 天津大学 | Human body action recognition method based on envelope density characteristics |
| CN113296087A (en) * | 2021-05-25 | 2021-08-24 | 沈阳航空航天大学 | Frequency modulation continuous wave radar human body action identification method based on data enhancement |
| CN113296087B (en) * | 2021-05-25 | 2023-09-22 | 沈阳航空航天大学 | Frequency modulation continuous wave radar human body action recognition method based on data enhancement |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108664894A (en) | The human action radar image sorting technique of neural network is fought based on depth convolution | |
| CN107169435B (en) | Convolutional neural network human body action classification method based on radar simulation image | |
| CN108226892B (en) | Deep learning-based radar signal recovery method in complex noise environment | |
| Arif et al. | Automated body parts estimation and detection using salient maps and Gaussian matrix model | |
| CN110045348A (en) | A kind of human motion state classification method based on improvement convolutional neural networks | |
| CN105374026B (en) | A kind of detection method of marine infrared small target suitable for coast defence monitoring | |
| Belloni et al. | Explainability of deep SAR ATR through feature analysis | |
| CN108509910A (en) | Deep learning gesture identification method based on fmcw radar signal | |
| CN108898620A (en) | Method for tracking target based on multiple twin neural network and regional nerve network | |
| Li et al. | Sign language recognition based on computer vision | |
| CN107025420A (en) | The method and apparatus of Human bodys' response in video | |
| CN104484890A (en) | Video target tracking method based on compound sparse model | |
| CN112215296B (en) | Infrared image recognition method based on transfer learning and storage medium | |
| CN110728698A (en) | Multi-target tracking model based on composite cyclic neural network system | |
| CN110647788B (en) | Human daily behavior classification method based on micro-Doppler characteristics | |
| CN108537181A (en) | A kind of gait recognition method based on the study of big spacing depth measure | |
| CN106127161A (en) | Fast target detection method based on cascade multilayer detector | |
| Yang et al. | Enhancing PIR-based multi-person localization through combining deep learning with domain knowledge | |
| CN104301585A (en) | Method for detecting specific kind objective in movement scene in real time | |
| Bera et al. | Interactive crowd-behavior learning for surveillance and training | |
| Batool et al. | Telemonitoring of daily activities based on multi-sensors data fusion | |
| Chen et al. | Human activity recognition using temporal 3DCNN based on FMCW radar | |
| CN113887469A (en) | Method, system and storage medium for pedestrian fall detection | |
| Song et al. | Dense face network: A dense face detector based on global context and visual attention mechanism | |
| Huo et al. | 3DVSD: An end-to-end 3D convolutional object detection network for video smoke detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181016 |