CN108520199A - Based on radar image and the human action opener recognition methods for generating confrontation model - Google Patents
Based on radar image and the human action opener recognition methods for generating confrontation model Download PDFInfo
- Publication number
- CN108520199A CN108520199A CN201810177104.8A CN201810177104A CN108520199A CN 108520199 A CN108520199 A CN 108520199A CN 201810177104 A CN201810177104 A CN 201810177104A CN 108520199 A CN108520199 A CN 108520199A
- Authority
- CN
- China
- Prior art keywords
- radar
- arbiter
- opener
- indicate
- human
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
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
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
-
- 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
Abstract
The present invention relates to Radar Technology fields and human action to identify field, to propose based on radar image and the human action opener recognition methods for generating confrontation model, it directly distinguishes input picture to be known or unknown action classification and export its classification information, to realize that the end-to-end opener to human action identifies.Thus, the technical solution adopted by the present invention is, based on radar image and the human action opener recognition methods for generating confrontation model, it can reflect the characteristic of human body fine motion using the micro-doppler image of radar, simultaneously using the arbiter generated in confrontation model as opener identifier, it directly distinguishes input picture to be known or unknown action classification and export its classification information, to realize that the end-to-end opener to human action identifies.Present invention is mainly applied to Radar Technology fields and human action to identify occasion.
Description
Technical field
The present invention relates to Radar Technology fields and human action to identify field, more particularly to a kind of based on generation confrontation model
Opener action identification method.
Background technology
In nearest decades, human action identification causes extensive concern in various fields.Due to action recognition
Demand in amusement, medical monitoring, security protection, emergency relief and other application field is growing, is considered as a kind of tool
There is the project of broad prospect of application.Human action identification depends on visual sensor data and in computer vision in the past
Many achievements are achieved under help.Later, the research in this field is promoted once again by depth transducer.Such as biography of " Kinect "
Sensor provides a kind of straightforward procedure obtaining depth information for researcher.However these sensors are all highly prone to illumination, hide
The influence of the environmental factors such as gear, weather, and the above situation is difficult to avoid that in practical applications, therefore these sensors are not
It is not strong with the robustness in application scenarios.
Radar since it can ignore the influence of the environmental factors such as weather, and can all weather operations the characteristics of, become people
A kind of emerging sensors in body action recognition." micro-Doppler effect " that radar receives echo is referred in target Continuous
Micro-doppler frequency displacement caused by certain fine motions (such as hand, foot, four limbs) in motion process.This category feature can be reflected in radar signal
On spectrogram after visualization.The current existing correlative study based on micro-doppler image, such as using artificial from radar image
Feature (such as trunk frequency, signal total bandwidth, period) is extracted, support vector machines (Support Vector are then used
Machine, SVM), the methods of k arest neighbors (k-Nearest Neighbor, kNN) carries out radar image based on these features
Classification.Compared to the method for needing manual extraction feature, convolutional neural networks (Convolutional Neural Network,
CNN) there is good non-linear mapping capability and can independently extract the hidden feature in image, therefore be widely applied.
Closed set data, i.e. test set data and training set data source phase currently are all based on to human action identification problem
Together and include same classification.However in actual environment, human action is ever-changing, marks to action structure data set and one by one
Note is clearly to be difficult to complete.Even if can define certain one kind action in fixed scene, same action is showed by different people
There is also great differences.Therefore, the action recognition in actual environment should be counted as opener identification problem, and data set includes
Such issues that known class and unknown class, solution, needs to provide a kind of mould that can automatically distinguish unknown class and known class
Type.
Have some opener recognition methods at present to be suggested, W.Scheirer et al. proposes a kind of " 1-vs-Set " study
Machine becomes one of the pionerring research in opener identification field.They propose a kind of CAP methods (Compact Abating later
Probability, CAP) and be combined itself and the extreme value theorem (Extreme Value Theory, EVT) in statistics, from
And a kind of Weibull calibration support vector machines (Weibull-calibrated SVM, W-SVM) is formd, experiments have shown that this method
It is promoted on recognition effect according to than 1-vs-Set Machine.Abhijit Bendale and Terrance Boult will most
Nearly class mean algorithm (Nearest Class Mean type algorithms, NCM) is extended to nearest value of the non-outlier algorithm
(Nearest Non-Outlier, NNO), this method can balance the relationship of recognition accuracy and opener degree.Utilize depth
The opener recognition methods of study also becomes a kind of emerging direction, and Abhijit Bendale and Terrance Boult propose one
Kind novel layer structure --- " OpenMax " layer, but this method is needed by pre-training model, therefore it is in other identification missions
Availability is not strong.
To sum up, existing opener method will rely on the reasonable selection to probability threshold value, therefore these methods are in other tasks
Middle shortage robustness.In addition, it is contemplated that the deficiency of the sensor and the advantage of radar, using radar micro-doppler image into
The problem of opener identification of row human action, is urgently to be resolved hurrily.
Invention content
In order to overcome the deficiencies of the prior art, the present invention is directed to propose being moved with the human body for generating confrontation model based on radar image
Make opener recognition methods, directly distinguishes input picture and be known or unknown action classification and export its classification information, to realize
End-to-end opener identification to human action.For this purpose, the technical solution adopted by the present invention is, fought based on radar image and generation
The human action opener recognition methods of model, the characteristic of human body fine motion can be reflected using the micro-doppler image of radar, simultaneously
Using the arbiter generated in confrontation model as opener identifier, it is known or unknown action class directly to distinguish input picture
Not and its classification information is exported, to realize that the end-to-end opener to human action identifies.
It is as follows：
Step 1：It sends and receives human body echo-signal using ULTRA-WIDEBAND RADAR module, after gathered data, data are carried out
Echo-signal is carried out the operations such as Short Time Fourier Transform, noise cancellation, determines useful signal section by pretreatment；
Step 2：Noise jamming is further eliminated using the method for given threshold, it is only aobvious in radar micro-doppler image
Show that echo strength is more than the point of threshold value；
Step 3：The data of acquisition are carried out to demarcate and determine training set, verification collection and test set；
Step 4：Utilize intensive connection network DenseNet (Densely Connected Convolutional
Networks) structure, which is established, generates confrontation model GAN (Generative Adversarial Nerworks), in sentencing for model
Other device output end will be originally generated the probability for fighting network output probability and being mapped as on of all categories using softmax functions；
Step 5：The generation confrontation model in step 4 is trained with the training set data determined in step 3, is instructed
It takes arbiter weight to test test set data after white silk, verifies model opener recognition effect.
Further, ultra-wide band radar used in the step 1 is 440 radar modules of PulsON, radar work frequency
Rate is 3.1GHz to 4.8GHz, using two directional aerials reception human body echo-signals when data acquire, data environment indoors
Lower acquisition, acquires seven kinds of typical human actions, and selected seven kinds of human actions are respectively：Walking, boxing, ground creep,
Move under water, stand, forward direction halt jump, running, each action by every subject in triplicate, each acquisition time be 7 seconds.
The Short Time Fourier Transform is non-stationary process to be regarded as a series of superposition of short-term stationarity signals, short-time characteristic
It is and Fourier transform to be done to signal in window by the way that adding window is realized on time dimension, the time varying spectrum of signal is obtained, in short-term in Fu
The formula of leaf transformation：
In formula, τ is time window length, and ω is angular frequency, and t is the time, and j is imaginary number, and e is natural constant, and g (t) is window letter
Number, f (t) is collected human body echo-signal, Gf() is the time varying spectrum after transformation.
The method that the noise cancellation uses mean value background cancel, i.e., entire echo-signal subtract echo strength average value row
Vector；
The method in the determining useful signal section when being by signal m- range image determine the area for having human motion
Between, then reasonable set carries out the time starting point and end point of time-frequency conversion.
Specifically, the intensity threshold that the step 2 setting is not shown will rely on the mode of artificial selection, using segmented
The method of threshold value filters out noise.
Specifically, the calibration in the step 3 to generation radar micro-doppler image, is marked with digital " 0 " to " 6 " successively
Note walks, boxes, ground creeps, moves under water, standing, forward direction is halted jump, this seven kinds actions of running, then according to 4:2:1 ratio
It will be by Step 1: the image generated after two be divided into training set, verification collection and test set.
Specifically, the intensive connection network in the step 4 includes " link block " and " transition zone " two parts, specifically：
Each connecting block structure is made of two convolutional layers and a connection operation layer, and connecting block structure connects before this layer
Each layer input of the feature as this layer, each convolutional layer is followed by one crowd of normalization operation BN (Batch
Normalization) and one is corrected linear unit R eLU (Rectified Linear Units) or linearly repairing with leakage
The expression formula of positive unit Leaky ReLU, ReLU and Leaky ReLU indicates as follows respectively：
In formula, p is the input of unit；
What transition zone indicated is the part between two connecting block structures, in the generator for generating confrontation model, transition
Layer is made of a convolutional layer and a warp lamination；In arbiter, transition zone is by a convolutional layer and a mean value pond
Layer composition.
Generation confrontation model described in step 4 is made of a generator and an arbiter, and generator is from potential sky
Between middle stochastical sampling need the authentic specimen that imitation exercise is concentrated as possible as input, output result, the input of arbiter is then
For authentic specimen or generate network output, the purpose is to which the output of generator is distinguished as far as possible from authentic specimen,
And generator will then cheat arbiter as much as possible, two networks confront with each other, constantly adjust each layer network weight, final purpose
It is to make arbiter that can not judge whether the output result of generator is true, the object function V (D, G) that production fights network is indicated
It is as follows：
Wherein, G indicates that generator, D indicate that arbiter, x indicate that input sample, z indicate the stochastic variable of input, min
() indicates to minimize operation, and max () indicates that maximum operation, log () are that denary logarithm is taken to operate, E ()
It indicating it is expected, Pdata (x) indicates to obey the data distribution of authentic specimen, and Pz (z) indicates to obey the data distribution of random distribution,
In addition, the output par, c of arbiter uses softmax functions, essence is exactly by the arbitrary real vector compression of a K dimension
At another K tie up real vector, wherein vector in each element value between (0,1), softmax functional forms
It is as follows：
In formula, zjIndicate j-th of element, zkIndicate that k-th of element, e are natural constant, σ (z)jIndicate j-th element
Softmax values；
In this way, the output of arbiter can be understood as probability of the input picture on each action classification, probability soprano
The as classification of arbiter judgement input picture.
Network training process described in step 5 uses adaptability moments estimation Adam (Adaptive Moment
Estimation) optimize network weight, also use gradient punishment strategy, i.e., penalty term is added in object functionWherein λ=10,α beStochastic variable between to 1,Indicate life
It growing up to be a useful person the dummy copy of generation, x indicates true sample,Expression asks gradient, E () to indicate it is expected, then object function V (D, G) is：
Common counter " F-measure-Openness curves " assessment models effect in the opener identification of use, F-
Measure is defined as follows：
Wherein,
TP indicates that by model prediction be positive positive sample, and TN indicates that by model prediction be the negative sample born, and FP is indicated by mould
Type is predicted as positive negative sample, and FN indicates that by model prediction be the positive sample born.
The features of the present invention and advantageous effect are：
The present invention utilizes radar micro-doppler image recognition human action, can avoid other sensors easily by environment shadow
Loud deficiency, and it is strong to the capturing ability of fine motion；The present invention is solved with the characteristic for generating confrontation model to unknown action
End-to-end identification identifies problem to the opener of human action, algorithm complexity is low, has certain application value.
Description of the drawings：
Fig. 1 is the human action opener recognition methods block diagram based on radar micro-doppler image；
Fig. 2 is the when m- range image of radar echo signal；
Fig. 3 is the micro-doppler example images of various actions；
Fig. 4 makes a living into generator structural schematic diagram in confrontation model；
Fig. 5 makes a living into arbiter structural schematic diagram in confrontation model；
Fig. 6 is experimental result of the present invention and other methods comparative result figure.
Specific implementation mode
The present invention is to solve the above-mentioned problems, it is proposed that a kind of human action based on radar image with generation confrontation model
Opener recognition methods, the present invention can reflect that human body is micro- when carrying out human action identification, using the micro-doppler image of radar
Dynamic characteristic, while using the arbiter generated in confrontation model as opener identifier, it is directly to distinguish input picture
Know or unknown action classification and export its classification information, to realize that the end-to-end opener to human action identifies.
To achieve the goals above, a kind of human action opener recognition methods based on radar image with generation confrontation model
Include the following steps：
Step 1：It sends and receives human body echo-signal using ULTRA-WIDEBAND RADAR module, after gathered data, data are carried out
Echo-signal is carried out the operations such as Short Time Fourier Transform, noise cancellation, determines useful signal section by pretreatment.
Step 2：Noise jamming is further eliminated using the method for given threshold, it is only aobvious in radar micro-doppler image
Show that echo strength is more than the point of threshold value.
Step 3：The data of acquisition are carried out to demarcate and determine training set, verification collection and test set.
Step 4：Using intensive connection network (Densely Connected Convolutional Networks,
DenseNet) structure, which is established, generates confrontation model (Generative Adversarial Nerworks, GAN), in sentencing for model
Other device output end will be originally generated the probability for fighting network output probability and being mapped as on of all categories using softmax functions.
Step 5：The generation confrontation model in step 4 is trained with the training set data determined in step 3, is instructed
It takes arbiter weight to test test set data after white silk, verifies model opener recognition effect.
Specifically, ultra-wide band radar used in the step 1 is 440 radar modules of PulsON, radar operating frequency
For 3.1GHz to 4.8GHz, human body echo-signal is received using two directional aerials when data acquire.Data are indoors under environment
Acquisition.Seven kinds of typical human actions are acquired in experiment, selected seven kinds of human actions are respectively：Walking, boxing, ground
Creep, move under water, standing, forward direction halt jump, running.In triplicate by every subject, each acquisition time is 7 seconds for each action
Left and right.
The Short Time Fourier Transform is non-stationary process to be regarded as a series of superposition of short-term stationarity signals, short-time characteristic
It is and Fourier transform to be done to signal in window by the way that adding window is realized on time dimension, obtain the time varying spectrum of signal.In short-term in Fu
The formula of leaf transformation can be write：
In formula, τ is time window length, and ω is angular frequency, and t is the time, and j is imaginary number, and e is natural constant, and g (t) is window letter
Number, f (t) is collected human body echo-signal, Gf() is the time varying spectrum after transformation.
The method that the noise cancellation uses mean value background cancel, i.e., entire echo-signal subtract echo strength average value row
Vector.
The method in the determining useful signal section when being by signal m- range image determine the area for having human motion
Between, then reasonable set carries out the time starting point and end point of time-frequency conversion.
Specifically, the intensity threshold that the step 2 setting is not shown will rely on the mode of artificial selection.It is acquired in data
During, since certain human motions are processes from far near, echo strength is in the trend become larger, according to
It will cause short distance echo-signal noise filtering is insufficient or remote echo-signal noise filtering is excessive for unified threshold value
And human body echo-signal detailed information is caused to lose.Therefore the present invention considers the influence of distance, using the side of segmented threshold value
Method filters out noise.
Specifically, the calibration in the step 3 to generation radar micro-doppler image, is marked with digital " 0 " to " 6 " successively
Note walks, boxes, ground creeps, moves under water, standing, forward direction is halted jump, this seven kinds actions of running, then according to 4:2:1 ratio
It will be by Step 1: the image generated after two be divided into training set, verification collection and test set.
Specifically, the intensive connection network in the step 4 includes " link block " and " transition zone " two parts, is divided below
Two component parts are not introduced.
Each connecting block structure is made of two convolutional layers and a connection operation layer, and connecting block structure can connect this layer
Input of the feature of each layer before as this layer.Each convolutional layer is followed by one crowd of normalization operation (Batch
Normalization, BN) and one correct linear unit (Rectified Linear Units, ReLU) or the line with leakage
Property amending unit (Leaky ReLU).The expression formula of ReLU and Leaky ReLU indicates as follows respectively：
In formula, p is the input of unit.
What transition zone indicated is the part between two connecting block structures.In the generator for generating confrontation model, transition
Layer is made of a convolutional layer and a warp lamination；In arbiter, transition zone is by a convolutional layer and a mean value pond
Layer composition.
Generation confrontation model described in step 4 is by a generator (Generator) and an arbiter
(Discriminator) it forms.Generator stochastical sampling from latent space (latent space) is tied as input, output
Fruit needs the authentic specimen that imitation exercise is concentrated as possible.The input of arbiter is then the output of authentic specimen or generation network,
Purpose is to distinguish the output of generator as far as possible from authentic specimen.And generator will then cheat differentiation as much as possible
Device.Two networks confront with each other, constantly adjust each layer network weight, and final purpose is to make arbiter that can not judge the defeated of generator
Whether true go out result.The object function V (D, G) of production confrontation network can indicate as follows：
Wherein, G indicates that generator, D indicate that arbiter, x indicate that input sample, z indicate the stochastic variable of input, min
() indicates to minimize operation, and max () indicates that maximum operation, log () are that denary logarithm is taken to operate, E ()
It indicates it is expected, Pdata (x) indicates to obey the data distribution of authentic specimen, and Pz (z) indicates to obey the data distribution of random distribution.
In addition, the output par, c of arbiter uses softmax functions, essence is exactly by the arbitrary real vector compression of a K dimension
The real vector that (mapping) is tieed up at another K, wherein each element value in vector is between (0,1).Softmax letters
Number form formula is as follows：
In formula, zjIndicate j-th of element, zkIndicate that k-th of element, e are natural constant, σ (z)jIndicate j-th element
Softmax values.
In this way, the output of arbiter can be understood as probability of the input picture on each action classification, probability soprano
The as classification of arbiter judgement input picture.
Network training process described in step 5 uses adaptability moments estimation (Adaptive Moment
Estimation, Adam) this first-order optimization method optimizes network weight, and Adam methods are in update it may be noted that step sizes
Selection and dynamically adjust the learning rate of each weight.In addition the problem of gradient disappears in order to prevent, the present invention additionally uses
Gradient punishment strategy, i.e., be added penalty term in object functionWherein λ=10,α beStochastic variable between to 1,Indicate that the dummy copy that generator generates, x indicate true sample,Expression asks gradient, E () to indicate it is expected.Then object function V (D, G) is：.Then object function V (D, G) of the invention is：
Common counter " F-measure-Openness curves " assessment models effect during the present invention is identified using opener,
F-measure is defined as follows：
Wherein,
TP indicates that by model prediction be positive positive sample, and TN indicates that by model prediction be the negative sample born, and FP is indicated by mould
Type is predicted as positive negative sample, and FN indicates that by model prediction be the positive sample born.
A kind of human action opener recognition methods based on radar image with generation confrontation model proposed by the present invention, is divided into
Five steps, as shown in Figure 1.The present invention is further explained with reference to the accompanying drawings and examples.
The first step, human body movement data acquisition.
The present invention is sent using 440 modules of ULTRA-WIDEBAND RADAR PulsON and receives human body echo-signal, radar operating frequency
For 3.1GHz to 4.8GHz, sample frequency 16GHz, pulse recurrence frequency (Pulse Recurrence Frequency, PRF)
For 368Hz, it is 0.2 second that coherent pulse, which accumulates (Coherent Pulse Interval, CPI), and two height are used when data acquire
The directional aerial that degree is 1.2m or so receives human body echo-signal.
In experiment, four subjects have done seven kinds of typical human actions in the direction of visual lines of radar, selected seven kinds
Human action is respectively：Walking, boxing, ground creep, move under water, standing, forward direction halt jump, running.Each action is by every quilt
In triplicate, each acquisition time is 7 seconds or so to examination person.After obtaining raw radar data, make its it is corresponding when m- distance
Image (as shown in Figure 2), being determined by the image has the section of human motion, and then reasonable set carries out the time of time-frequency conversion
Starting point and end point.The method for using mean value background cancel later, regards echo-signal as two-dimensional matrix, calculates separately each
The mean value of row data, is denoted as mi, wherein i indicate the i-th column data, then obtain column vector M '=[m of echo strength average value1,
m2,,mn]T.It is extended to the Mean Matrix of n × n：
The corresponding element of former data matrix and Mean Matrix is subtracted each other to get the signal matrix to after offseting.
Short Time Fourier Transform (Short-time Fourier Transform, STFT) is done to signal matrix later.It is short
When Fourier transformation be non-stationary process to be regarded as a series of superposition of short-term stationarity signals, it is real to pass through the adding window on time dimension
It is existing, and Fourier transformation is done to signal in window, obtain the time varying spectrum of signal.The formula of Short Time Fourier Transform can be write：
In formula, τ is time window length, and ω is angular frequency, and t is the time, and j is imaginary number, and e is natural constant, and g (t) is window letter
Number, f (t) is collected human body echo-signal, Gf() is the time varying spectrum after transformation.The Fourier in short-term that the present invention uses
The time window length of transformation is 0.1 second, and two time window Duplication are 0.9, and Fourier transformation points are 1024 in each window.
Second step, setting need intensity threshold to be shown.
Noise jamming is further eliminated using the method for given threshold, only display echo is strong in radar micro-doppler image
Point of the degree more than threshold value.
The intensity threshold that setting is not shown is needed when the present invention generates radar micro-doppler image will rely on artificial selection
Mode.The range that subject moves in data acquisition is in the range of apart from 1.2 meters to 5.4 meters of radar, due to certain people
Body movement is process from far near, therefore echo strength is in necessarily the trend become larger.Then can according to unified threshold value
Cause short distance echo-signal noise filtering insufficient, or remote echo-signal noise filtering excessively causes human body echo to believe
The problem of number detailed information is lost.
Therefore the present invention considers the influence of distance, and noise is filtered out using the method for segmented threshold value.Assuming that certain segment signal
Maximum of intensity is Max, then minimum strength to be shown is as shown in table 1 in each distance range.
Table 1
| Distance | Minimal intensity value |
| 1.2 meters~2 meters | Max-90 |
| 2 meters~3.2 meters | Max-80 |
| 3.2 meters~4.5 meters | Max-70 |
| 4.5 meters~5.4 meters | Max-60 |
Third walks, and builds data set.
By the operation of first two steps, 700 images can be obtained for each action, each action radar micro-doppler figure
The schematic diagram of picture is as shown in Figure 3.It is demarcated to generating radar micro-doppler image, is represented go with digital " 0 " to " 6 " respectively
It walks, box, ground is creeped, move under water, stand, forward direction is halted jump, this seven kinds actions of running, then according to 4:2:1 ratio will be every
The image of kind action is divided into training set, verification collection and test set.In this way, for each action can obtain 400 training set,
200 verification collection and 100 test sets.
Problem is identified for opener, it is also necessary to define known class and unknown class, training set is not included in test set
In classification become " unknown class ", be denoted as U；Not only the classification in test set but also in training set was known as " known class ", was denoted as K.This
The effect of model when invention demonstrates data set opener degree difference.
4th step, structure generate confrontation model.
The generation confrontation model that the present invention uses is by a generator (Generator) and an arbiter
(Discriminator) it forms.Generator from latent space (latent space) one variable z of stochastical sampling as defeated
Enter, output result needs the authentic specimen that imitation exercise is concentrated as possible.The input of arbiter is then authentic specimen or generation net
The output of network, the purpose is to distinguish the output of generator as far as possible from authentic specimen.And generator then will be as far as possible
Cheat arbiter in ground.Two networks confront with each other, constantly adjust each layer network weight, and final purpose is to make arbiter that can not judge
Whether the output result of generator is true.The object function V (D, G) of production confrontation model can indicate as follows：
Wherein, G indicates that generator, D indicate that arbiter, x indicate that input sample, z indicate the stochastic variable of input, min
() indicates to minimize operation, and max () indicates that maximum operation, log () are that denary logarithm is taken to operate, E ()
It indicates it is expected, Pdata(x) data distribution of obedience authentic specimen, P are indicatedz(z) data distribution of obedience random distribution is indicated.
Since arbiter itself is two graders, it can be determined that input picture is "true" or "false", therefore in opener
In identification problem, arbiter, which can also be realized, judges input picture for " known class " or the function of " unknown class ".In addition to this, it is
It can allow and arbiter while realize that the function of grader, the present invention output it part and be changed to softmax functions, output is made to become
Probability of the input picture in each classification.Its essence is exactly the arbitrary real vector compression (mapping) by a K dimension at another
The real vector of a K dimensions, wherein each element value in vector is between (0,1).Softmax functional forms are as follows：
In formula, zjIndicate j-th of element, zkIndicate that k-th of element, e are natural constant, σ (z)jIndicate j-th element
Softmax values.
In this way, the output of arbiter can be understood as probability of the input picture on each action classification, probability soprano
The as classification of arbiter judgement input picture.
It includes " link block " (Dense that the structure of generator and arbiter, which all uses intensive connection network, intensive connection network,
Block) and " transition zone " (Transition Layer) two parts, two component parts are introduced separately below.
Each connecting block structure is made of two convolutional layers and a connection operation layer, and connecting block structure can connect this layer
Input of the feature of each layer before as this layer.Each convolutional layer is followed by one crowd of normalization operation (Batch
Normalization, BN) and one correct linear unit (Rectified Linear Units, ReLU) or the line with leakage
Property amending unit (Leaky ReLU).The expression formula of ReLU and Leaky ReLU indicates as follows respectively：
In formula, p is the input of unit.
What transition zone indicated is the part between two connecting block structures.In the generator for generating confrontation model, transition
Layer is made of a convolutional layer and a warp lamination；In arbiter, transition zone is by a convolutional layer and a mean value pond
Layer composition.Fig. 4, Fig. 5 indicate the structure of generator and arbiter respectively.Table 2 lists the design parameter in model, in table " n ×
N deconv " indicate that the warp lamination that convolution kernel size is n × n, " n × n conv " indicate the convolution that convolution kernel size is n × n
Layer, " Padding " indicate that filler pixels number around picture, " Pooling " indicate the operation of mean value pondization.
Each layer design parameter of 2 production confrontation model of table
5th step, model training and test.
The present invention uses in the training process in order to generate gradient disappearance problem common in confrontation model training process
Gradient punishment strategy, i.e., be added penalty term in object functionWherein λ=10,α isStochastic variable between to 1,Indicate that the dummy copy that generator generates, x indicate true sample,Expression asks gradient, E () to indicate it is expected.Then object function V (D, G) is：.
Then object function V (D, G) of the invention is：
Optimizer in network training process uses first-order optimization method-adaptability moments estimation (Adaptive
Moment Estimation, Adam) adjust network weight, Adam methods in update it may be noted that the selection of step sizes simultaneously
Dynamically adjust the learning rate of each weight.
The present invention is using common " F-measure-Openness curves " in opener identification come assessment models effect, F-
Measure is defined as follows：
Wherein,
TP indicates that by model prediction be positive positive sample, and TN indicates that by model prediction be the negative sample born, and FP is indicated by mould
Type is predicted as positive negative sample, and FN indicates that by model prediction be the positive sample born.F-measure value ranges between (0,1),
Value is bigger, and to represent opener recognizer effect better.
Openness is used for indicating the degree of opener in opener identification problem, is defined as：
Wherein, NTAIndicate the classification number in training set, NTGIndicate the classification number to be identified, NTEIndicate the class in test set
Shuo not.
The experimental results showed that the present invention commonly uses opener recognizer compared with other, performance can improve about ten percentage points, in fact
Test that the results are shown in Figure 6.
The above embodiment is a preferred embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, it is other it is any without departing from the spirit and principles of the present invention made by changes, modifications, substitutions, combinations, simplifications,
Equivalent substitute mode is should be, is included within the scope of the present invention.
Claims (9)
1. a kind of human action opener recognition methods based on radar image with generation confrontation model, characterized in that utilize radar
Micro-doppler image can reflect the characteristic of human body fine motion, while using the arbiter generated in confrontation model as opener knowledge
Other device directly distinguishes input picture and is known or unknown action classification and exports its classification information, to realize to human action
End-to-end opener identification.
2. special as described in claim 1 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is to be as follows：
Step 1：It sends and receives human body echo-signal using ULTRA-WIDEBAND RADAR module, after gathered data, data are located in advance
Echo-signal is carried out Short Time Fourier Transform, noise cancellation operation, determines useful signal section by reason；
Step 2：Noise jamming is further eliminated using the method for given threshold, is only shown back in radar micro-doppler image
Intensity of wave is more than the point of threshold value；
Step 3：The data of acquisition are carried out to demarcate and determine training set, verification collection and test set；
Step 4：It utilizes intensive connection network DenseNet (Densely Connected Convolutional Networks)
Structure, which is established, generates confrontation model GAN (Generative Adversarial Nerworks), in the arbiter output end of model
The probability for fighting network output probability and being mapped as on of all categories will be originally generated using softmax functions；
Step 5：The generation confrontation model in step 4 is trained with the training set data determined in step 3, training knot
It takes arbiter weight to test test set data after beam, verifies model opener recognition effect.
3. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is that further, ultra-wide band radar used in the step 1 is 440 radar modules of PulsON, and radar operating frequency is
3.1GHz to 4.8GHz receives human body echo-signals when data acquire using two directional aerials, and data are adopted under environment indoors
Collection, acquires seven kinds of typical human actions, and selected seven kinds of human actions are respectively：Walking, boxing, ground creep, move under water,
Stand, forward direction halt jump, running, each action by every subject in triplicate, each acquisition time be 7 seconds.
4. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is that the Short Time Fourier Transform is non-stationary process to be regarded as a series of superposition of short-term stationarity signals, and short-time characteristic is
By the way that adding window is realized on time dimension, and Fourier transform is done to signal in window, obtain the time varying spectrum of signal, in short-term Fourier
The formula of transformation：
In formula, τ is time window length, and ω is angular frequency, and t is the time, and j is imaginary number, and e is natural constant, and g (t) is window function, f
(t) it is collected human body echo-signal, Gf() is the time varying spectrum after transformation.
5. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is, the method that the noise cancellation uses mean value background cancel, i.e., entire echo-signal subtract echo strength average value arrange to
Amount；The method in the determining useful signal section when being by signal m- range image determine the section for having human motion, so
Reasonable set carries out the time starting point and end point of time-frequency conversion afterwards.
6. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is that specifically, the intensity threshold that the step 2 setting is not shown will rely on the mode of artificial selection, using segmented threshold value
Method filter out noise.
7. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is specifically, to the calibration of generation radar micro-doppler image in the step 3, to mark row successively with digital " 0 " to " 6 "
It walks, box, ground is creeped, move under water, stand, forward direction is halted jump, this seven kinds actions of running, then according to 4:2:1 ratio will be through
It crosses Step 1: the image generated after two is divided into training set, verification collection and test set.
8. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is that specifically, the intensive connection network in the step 4 includes " link block " and " transition zone " two parts, specifically：Often
A connecting block structure is made of two convolutional layers and a connection operation layer, connecting block structure connect this layer before each layer spy
The input as this layer is levied, each convolutional layer is followed by one crowd of normalization operation BN (Batch Normalization) and one
It is a to correct linear unit R eLU (Rectified Linear Units) or the linear amending unit Leaky ReLU with leakage,
The expression formula of ReLU and Leaky ReLU indicates as follows respectively：
In formula, p is the input of unit；
What transition zone indicated is part between two connecting block structures, in the generator for generating confrontation model, transition zone by
One convolutional layer and a warp lamination composition；In arbiter, transition zone is by a convolutional layer and a mean value pond layer group
At.
9. special as claimed in claim 2 based on radar image and the human action opener recognition methods for generating confrontation model
Sign is that the generation confrontation model described in step 4 is made of a generator and an arbiter, and generator is from latent space
Middle stochastical sampling exports the authentic specimen that result needs imitation exercise concentration as possible as input, and the input of arbiter is then
Authentic specimen or the output for generating network, the purpose is to which the output of generator is distinguished as far as possible from authentic specimen, and
Generator will then cheat arbiter as much as possible, and two networks confront with each other, constantly adjust each layer network weight, and final purpose is
Make arbiter that can not judge whether the output result of generator is true, the object function V (D, G) that production fights network is indicated such as
Under：
Wherein, G indicates that generator, D indicate that arbiter, x indicate that input sample, z indicate the stochastic variable of input, min () table
Show that minimum operation, max () indicate that maximum operation, log () are that denary logarithm is taken to operate, E () indicates the phase
It hopes, Pdata(x) data distribution of obedience authentic specimen, P are indicatedz(z) data distribution of obedience random distribution is indicated, in addition, differentiating
The output par, c of device uses softmax functions, and essence is exactly that a K arbitrary real vectors tieed up are compressed into another K
The real vector of dimension, wherein each element value in vector is between (0,1), softmax functional forms are as follows：
In formula, zjIndicate j-th of element, zkIndicate that k-th of element, e are natural constant, σ (z)jIndicate j-th element
Softmax values；
In this way, the output of arbiter can be understood as probability of the input picture on each action classification, probability soprano is
The classification of arbiter judgement input picture.
Network training process described in step 5 uses adaptability moments estimation Adam (Adaptive Moment Estimation)
Optimize network weight, also uses gradient punishment strategy, i.e., penalty term is added in object function
Wherein λ=10,α beStochastic variable between to 1,Indicate the dummy copy that generator generates, x
Indicate true sample,Expression asks gradient, E () to indicate it is expected, then object function V (D, G) is：
Common counter " F-measure-Openness curves " assessment models effect in the opener identification of use, F-measure are fixed
Justice is as follows：
Wherein,
TP indicates that by model prediction be positive positive sample, and TN indicates that by model prediction be the negative sample born, and FP indicates pre- by model
It is positive negative sample to survey, and FN indicates that by model prediction be the positive sample born.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810177104.8A CN108520199B (en) | 2018-03-04 | 2018-03-04 | Human body action open set identification method based on radar image and generation countermeasure model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810177104.8A CN108520199B (en) | 2018-03-04 | 2018-03-04 | Human body action open set identification method based on radar image and generation countermeasure model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108520199A true CN108520199A (en) | 2018-09-11 |
| CN108520199B CN108520199B (en) | 2022-04-08 |
Family
ID=63433468
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810177104.8A Active CN108520199B (en) | 2018-03-04 | 2018-03-04 | Human body action open set identification method based on radar image and generation countermeasure model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108520199B (en) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109472757A (en) * | 2018-11-15 | 2019-03-15 | 央视国际网络无锡有限公司 | It is a kind of that logo method is gone based on the image for generating confrontation neural network |
| CN109871805A (en) * | 2019-02-20 | 2019-06-11 | 中国电子科技集团公司第三十六研究所 | A kind of electromagnetic signal opener recognition methods |
| CN109918994A (en) * | 2019-01-09 | 2019-06-21 | 天津大学 | A kind of act of violence detection method based on commercial Wi-Fi |
| CN109948532A (en) * | 2019-03-19 | 2019-06-28 | 桂林电子科技大学 | ULTRA-WIDEBAND RADAR human motion recognition method based on depth convolutional neural networks |
| CN110033043A (en) * | 2019-04-16 | 2019-07-19 | 杭州电子科技大学 | Radar range profile's based on condition production confrontation network are refused to sentence method |
| CN110052000A (en) * | 2019-04-12 | 2019-07-26 | 漳州泰里斯体育器材有限公司 | A kind of identifying processing method and system of combat sports state |
| CN110084108A (en) * | 2019-03-19 | 2019-08-02 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Pedestrian re-identification system and method based on GAN neural network |
| CN110096976A (en) * | 2019-04-18 | 2019-08-06 | 中国人民解放军国防科技大学 | Human behavior micro-Doppler classification method based on sparse migration network |
| CN110109090A (en) * | 2019-03-28 | 2019-08-09 | 北京邮电大学 | Circumstances not known multi-target detection method and device based on microwave radar |
| CN110390650A (en) * | 2019-07-23 | 2019-10-29 | 中南大学 | OCT image denoising method based on intensive connection and generation confrontation network |
| CN110532909A (en) * | 2019-08-16 | 2019-12-03 | 成都电科慧安科技有限公司 | A kind of Human bodys' response method based on three-dimensional UWB positioning |
| CN111239739A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| CN111461267A (en) * | 2019-03-29 | 2020-07-28 | 太原理工大学 | Gesture recognition method based on RFID technology |
| CN111507361A (en) * | 2019-01-30 | 2020-08-07 | 富士通株式会社 | Microwave radar-based action recognition device, method and system |
| 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 |
| CN111914919A (en) * | 2020-07-24 | 2020-11-10 | 天津大学 | Open set radiation source individual identification method based on deep learning |
| CN112200123A (en) * | 2020-10-24 | 2021-01-08 | 中国人民解放军国防科技大学 | Hyperspectral open set classification method combining dense connection network and sample distribution |
| CN112364689A (en) * | 2020-10-09 | 2021-02-12 | 天津大学 | Human body action and identity multi-task identification method based on CNN and radar image |
| CN112560778A (en) * | 2020-12-25 | 2021-03-26 | 万里云医疗信息科技（北京）有限公司 | DR image body part identification method, device, equipment and readable storage medium |
| CN112560596A (en) * | 2020-12-01 | 2021-03-26 | 中国航天科工集团第二研究院 | Radar interference category identification method and system |
| 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 |
| CN113378718A (en) * | 2021-06-10 | 2021-09-10 | 中国石油大学(华东) | Action identification method based on generation of countermeasure network in WiFi environment |
| CN113537374A (en) * | 2021-07-26 | 2021-10-22 | 百度在线网络技术（北京）有限公司 | Confrontation sample generation method |
| CN117115596A (en) * | 2023-10-25 | 2023-11-24 | 腾讯科技（深圳）有限公司 | Training method, device, equipment and medium of object action classification model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110135165A1 (en) * | 2009-06-02 | 2011-06-09 | Harry Wechsler | Robust Human Authentication Using Holistic Anthropometric and Appearance-Based Features and Boosting |
| 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 |
| 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 |
| CN107506799A (en) * | 2017-09-01 | 2017-12-22 | 北京大学 | A kind of opener classification based on deep neural network is excavated and extended method and device |
-
2018
- 2018-03-04 CN CN201810177104.8A patent/CN108520199B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110135165A1 (en) * | 2009-06-02 | 2011-06-09 | Harry Wechsler | Robust Human Authentication Using Holistic Anthropometric and Appearance-Based Features and Boosting |
| 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 |
| 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 |
| CN107506799A (en) * | 2017-09-01 | 2017-12-22 | 北京大学 | A kind of opener classification based on deep neural network is excavated and extended method and device |
Non-Patent Citations (2)
| Title |
|---|
| ÇAĞLIYAN B 等: "Micro-Doppler-based human activity classification using the mote-scale BumbleBee radar", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
| 祝依龙 等: "基于高分辨一维多普勒像的雷达目标机动检测算法", 《自动化学报》 * |
Cited By (34)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109472757A (en) * | 2018-11-15 | 2019-03-15 | 央视国际网络无锡有限公司 | It is a kind of that logo method is gone based on the image for generating confrontation neural network |
| CN109918994B (en) * | 2019-01-09 | 2023-09-15 | 天津大学 | Commercial Wi-Fi-based violent behavior detection method |
| CN109918994A (en) * | 2019-01-09 | 2019-06-21 | 天津大学 | A kind of act of violence detection method based on commercial Wi-Fi |
| CN111507361A (en) * | 2019-01-30 | 2020-08-07 | 富士通株式会社 | Microwave radar-based action recognition device, method and system |
| CN111507361B (en) * | 2019-01-30 | 2023-11-21 | 富士通株式会社 | Action recognition device, method and system based on microwave radar |
| CN109871805A (en) * | 2019-02-20 | 2019-06-11 | 中国电子科技集团公司第三十六研究所 | A kind of electromagnetic signal opener recognition methods |
| CN109871805B (en) * | 2019-02-20 | 2020-10-27 | 中国电子科技集团公司第三十六研究所 | Electromagnetic signal open set identification method |
| CN109948532A (en) * | 2019-03-19 | 2019-06-28 | 桂林电子科技大学 | ULTRA-WIDEBAND RADAR human motion recognition method based on depth convolutional neural networks |
| CN110084108A (en) * | 2019-03-19 | 2019-08-02 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Pedestrian re-identification system and method based on GAN neural network |
| CN110109090A (en) * | 2019-03-28 | 2019-08-09 | 北京邮电大学 | Circumstances not known multi-target detection method and device based on microwave radar |
| CN110109090B (en) * | 2019-03-28 | 2021-03-12 | 北京邮电大学 | Unknown environment multi-target detection method and device based on microwave radar |
| CN111461267B (en) * | 2019-03-29 | 2023-04-18 | 太原理工大学 | Gesture recognition method based on RFID technology |
| CN111461267A (en) * | 2019-03-29 | 2020-07-28 | 太原理工大学 | Gesture recognition method based on RFID technology |
| CN110052000A (en) * | 2019-04-12 | 2019-07-26 | 漳州泰里斯体育器材有限公司 | A kind of identifying processing method and system of combat sports state |
| CN110033043A (en) * | 2019-04-16 | 2019-07-19 | 杭州电子科技大学 | Radar range profile's based on condition production confrontation network are refused to sentence method |
| CN110096976A (en) * | 2019-04-18 | 2019-08-06 | 中国人民解放军国防科技大学 | Human behavior micro-Doppler classification method based on sparse migration network |
| CN110390650A (en) * | 2019-07-23 | 2019-10-29 | 中南大学 | OCT image denoising method based on intensive connection and generation confrontation network |
| CN110532909A (en) * | 2019-08-16 | 2019-12-03 | 成都电科慧安科技有限公司 | A kind of Human bodys' response method based on three-dimensional UWB positioning |
| CN111239739A (en) * | 2020-01-10 | 2020-06-05 | 上海眼控科技股份有限公司 | Weather radar echo map prediction method and device, computer equipment and storage medium |
| 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 |
| CN111914919A (en) * | 2020-07-24 | 2020-11-10 | 天津大学 | Open set radiation source individual identification method based on deep learning |
| CN112364689A (en) * | 2020-10-09 | 2021-02-12 | 天津大学 | Human body action and identity multi-task identification method based on CNN and radar image |
| CN112200123B (en) * | 2020-10-24 | 2022-04-05 | 中国人民解放军国防科技大学 | Hyperspectral open set classification method combining dense connection network and sample distribution |
| CN112200123A (en) * | 2020-10-24 | 2021-01-08 | 中国人民解放军国防科技大学 | Hyperspectral open set classification method combining dense connection network and sample distribution |
| CN112560596A (en) * | 2020-12-01 | 2021-03-26 | 中国航天科工集团第二研究院 | Radar interference category identification method and system |
| CN112560596B (en) * | 2020-12-01 | 2023-09-19 | 中国航天科工集团第二研究院 | Radar interference category identification method and system |
| CN112560778A (en) * | 2020-12-25 | 2021-03-26 | 万里云医疗信息科技（北京）有限公司 | DR image body part identification method, device, equipment and readable storage medium |
| 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 |
| 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 |
| CN113378718A (en) * | 2021-06-10 | 2021-09-10 | 中国石油大学(华东) | Action identification method based on generation of countermeasure network in WiFi environment |
| CN113537374A (en) * | 2021-07-26 | 2021-10-22 | 百度在线网络技术（北京）有限公司 | Confrontation sample generation method |
| CN113537374B (en) * | 2021-07-26 | 2023-09-08 | 百度在线网络技术（北京）有限公司 | Method for generating countermeasure sample |
| CN117115596A (en) * | 2023-10-25 | 2023-11-24 | 腾讯科技（深圳）有限公司 | Training method, device, equipment and medium of object action classification model |
| CN117115596B (en) * | 2023-10-25 | 2024-02-02 | 腾讯科技（深圳）有限公司 | Training method, device, equipment and medium of object action classification model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108520199B (en) | 2022-04-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108520199A (en) | Based on radar image and the human action opener recognition methods for generating confrontation model | |
| CN108226892B (en) | Deep learning-based radar signal recovery method in complex noise environment | |
| US20200166611A1 (en) | Detection method, detection device, terminal and detection system | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN111814875B (en) | Ship sample expansion method in infrared image based on pattern generation countermeasure network | |
| CN110045348A (en) | A kind of human motion state classification method based on improvement convolutional neural networks | |
| CN107290741A (en) | Combine the indoor human body gesture recognition method apart from time-frequency conversion based on weighting | |
| CN108492258A (en) | A kind of radar image denoising method based on generation confrontation network | |
| CN105844627B (en) | A kind of sea-surface target image background suppressing method based on convolutional neural networks | |
| CN108664894A (en) | The human action radar image sorting technique of neural network is fought based on depth convolution | |
| CN108509910A (en) | Deep learning gesture identification method based on fmcw radar signal | |
| CN110110649A (en) | Alternative method for detecting human face based on directional velocity | |
| Liu et al. | Deep learning and recognition of radar jamming based on CNN | |
| CN110007366A (en) | A kind of life searching method and system based on Multi-sensor Fusion | |
| CN110135476A (en) | A kind of detection method of personal safety equipment, device, equipment and system | |
| CN106228569A (en) | A kind of fish speed of moving body detection method being applicable to water quality monitoring | |
| Abdulatif et al. | Person identification and body mass index: A deep learning-based study on micro-Dopplers | |
| CN109444912A (en) | A kind of driving environment sensory perceptual system and method based on Collaborative Control and deep learning | |
| CN112433207A (en) | Human body identity recognition method based on two-channel convolutional neural network | |
| CN112711979A (en) | Non-contact vital sign monitoring under slow random motion based on biological radar | |
| CN107255818A (en) | A kind of submarine target quick determination method of bidimensional multiple features fusion | |
| CN108898066B (en) | Human motion detection method based on generating type countermeasure network | |
| CN107967941A (en) | A kind of unmanned plane health monitoring method and system based on intelligent vision reconstruct | |
| CN113537120B (en) | Complex convolution neural network target identification method based on complex coordinate attention | |
| Chen et al. | Human activity recognition using temporal 3DCNN based on FMCW radar |
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 |