CN110096976A - Human behavior micro-Doppler classification method based on sparse migration network - Google Patents
Human behavior micro-Doppler classification method based on sparse migration network Download PDFInfo
- Publication number
- CN110096976A CN110096976A CN201910311070.1A CN201910311070A CN110096976A CN 110096976 A CN110096976 A CN 110096976A CN 201910311070 A CN201910311070 A CN 201910311070A CN 110096976 A CN110096976 A CN 110096976A
- Authority
- CN
- China
- Prior art keywords
- network
- micro
- doppler
- sparse
- human body
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Radar Systems Or Details Thereof (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a human behavior micro-Doppler classification method based on a sparse migration network, which comprises the steps of firstly, determining an initial structure of a neural network, and pre-training in an ImageNet natural image database; then, modifying the full connection layer of the network to make the network output layer unit consistent with the classification number of the micro Doppler to be classified; then, migration training is carried out on a small number of micro Doppler image sets based on a sparse constraint criterion, and redundant areas of the network are marked in the training process; and finally, pruning the redundant area of the network, namely greatly reducing the complexity of the network on the premise of not influencing the identification performance, thereby obtaining a light-weight deep network.
Description
Technical field
The present invention relates to radar target micro-Doppler effect analysis field, be it is a kind of based on sparse migration network to human body not
The method for carrying out feature extraction and classification with micro-doppler spectrogram caused by behavior.
Background technique
Human body target detection radar is applied to battle reconnaissance earliest, the demands such as anti-terrorism in recent years, post-disaster search and rescue, medical monitoring
Appearance expanded the application of human body target detection radar.Whether there is or not hairs by easy detection human body target for radar human detection function
It opens up and determines target position, developing deeply to human body attitude identifies.Since human motion belongs to non-rigid motion, different action processes
In each limbs there are respective regular motion features, the movement velocity of each limbs can be captured by radar, be embodied in transmitting and
The difference on the frequency for receiving echo is anisotropic, i.e. Doppler frequency shift.The fine motion of each limbs, that is, micro-doppler frequency displacement, also referred to as micro-doppler effect
It answers.Micro-Doppler effect adjusts the distance, light condition and background complexity and insensitive, can effectively make up the hard of optical sensor
Part limitation provides strong foundation for analysis estimation human motion characteristic under complex scene.
Micro-doppler frequency displacement is a kind of time-varying frequency displacement, can be by carrying out time-frequency conversion to radar return, in T/F group
At two-dimentional spectrogram on characterized.By carrying out feature extraction and analysis to time-frequency spectrum, human body behavior can be identified
And classification.In radar return, the time-frequency spectral intensity of trunk is much larger than four limbs, and the temporal characteristics of limbs fine motion capture and divide
It is larger to analyse difficulty.It is exercised supervision study by using the convolutional neural networks of deep layer to human body micro-doppler time-frequency spectrum, it can be right
Each class behavior is effectively distinguished.But deep neural network usually require a large amount of training samples (it is generally necessary to it is thousands of so on
Ten thousand), the parameter of neural network is equally numerous and jumbled, and training complexity is high.Micro-doppler based on deep neural network is analyzed practical
In application process, need to reduce emphatically computational complexity and amount of training data constraint.
Summary of the invention
The technical problem to be solved by the present invention is in view of the shortcomings of the prior art, provide a kind of based on sparse migration network
Human body behavior micro-doppler classification method, reduce computational complexity and amount of training data constraint.
In order to solve the above technical problems, the technical scheme adopted by the invention is that: it is a kind of based on the light of sparse migration network
Quantify human body behavior micro-doppler recognition methods.First with the large-scale open source natural image database such as ImageNet, to depth
Neural network carries out pre-training, optimizes the weighting parameter of network；Then in the micro-doppler time-frequency modal data of a small amount of (50/class)
Library carries out network migration training, can the cyberspace that can cut of determination simultaneously during parameter optimization；Finally by network
In the region that can cut cut, finely tuned, can be obtained required lightweight network, which can be to the micro- of different behaviors
Doppler's time-frequency spectrum is effectively distinguished, to realize the Classification and Identification of human body behavior.
Basic ideas of the invention are: firstly, the initial configuration of neural network is determined, in ImageNet natural image data
Pre-training is carried out in library；Then, the full articulamentum for modifying network makes network output layer unit and quasi- classification micro-doppler classification number
Unanimously；Then, transfer training is carried out on a small amount of micro-doppler image set based on sparse constraint criterion, in training process, to net
The redundant area of network is labeled；Finally, the redundant area of network is subjected to beta pruning, it can be in the premise for not influencing recognition performance
Under, network complexity is greatly reduced, to obtain light-weighted depth network.
Technical solution of the present invention includes following processing step:
The first step determines the initial network structure of deep neural network
Initial configuration of the depth convolutional neural networks Vgg-19 as network is chosen, the network is by 16 layers of convolutional layer and 3 layers
Full articulamentum composition, the parameter of each layer of network carry out random initializtion.19 layer network stacks gradually, before be convolutional layer, most
Three layers are full articulamentum afterwards.
Second step, natural image database carry out network parameter pre-training
Determining deep neural network is known in the classification that ImageNet natural image database carries out all kinds of images first
Not, by ImageNet training sample abundant deep layer network is trained up, the parameters of network are optimized.
Third step, the full articulamentum adjustment of network
Since the image of ImageNet natural image database shares 1000 types, with human body behavior type to be identified
It is usually inconsistent, need to the output unit of the full articulamentum (refering in particular to the last layer of Vgg-19 neural network herein) of network into
Row modification.
4th step, the transfer training based on human body micro-doppler time-frequency spectrum under sparse constraint
The modified network of full articulamentum is subjected to further supervised learning in human body micro-Doppler feature database.It learns
During habit, the image difference of one side Learning from Nature image and radar return time-frequency spectrum improves the transfer ability between data, separately
On the one hand the significance level for determining each section in network, is marked the lower network channel of importance.
5th step, depth network pruning and fine tuning
According to the importance of quasi- cutting ratio and each component part of network, network is accordingly cut.Net after cutting
Road can be obtained light-weighted network by 10-20 repetition training in micro-doppler time-frequency spectrum data set for actually making
With.
Compared with prior art, the advantageous effect of present invention is that: the present invention can be significantly reduced neural network instruction
Practice required data set scale, the memory space of neural network, and computational complexity in actual use.Data set rule
Mould is reduced to 50/class by the training sample of 300/class of original usual need, and the memory space and computational complexity of network can
To be cut according to demand, (memory space can compress corresponding to computational complexity when VGG16 network pruning scale reaches 80%
Amplitude), the classification accuracy of network only declines within 1%, reduces the constraint of computational complexity and amount of training data.
Detailed description of the invention
Fig. 1 is processing flow schematic diagram of the present invention.
Fig. 2 is the network structure of deep neural network.
Fig. 3 be different constraint factors under, numeric distribution.
Fig. 4 is to cut network area under different cutting ratios.
Fig. 5 is performance comparison figure before and after network pruning.
Specific embodiment
The present invention proposes the process flow of the lightweight human body behavior micro-doppler recognition methods based on sparse migration network
As shown in Figure 1.The present invention is further explained below.
The first step determines the initial network structure of deep neural network
Initial configuration of the depth convolutional neural networks Vgg-19 as network is chosen, the network is by 16 convolutional layers and 3
Full articulamentum composition, specific network structure are as shown in Figure 2.Wherein 3x3Conv indicates convolution of the convolution kernel having a size of 3x3 size
Layer, 64,128,256,512 be respectively the output channel number of network；MaxPool is maximum pond layer, i.e., image is in the layer network
It can be maximized to the region (2x2) is closed on, to play the effect of picture size compression；FC is full articulamentum, and 4096,1000 refer to
The output unit number of network, Softmax are classification function.The parameter of each layer of network carries out random initializtion.
Second step, natural image database carry out network parameter pre-training
Determining deep neural network is known in the classification that ImageNet natural image database carries out all kinds of images first
Not, learning process is supervised training method, by each ginseng for making deep layer network by ImageNet training sample abundant
Number can train up, to improve the feature learning and Classification and Identification performance of network.
Third step, the full articulamentum adjustment of network
Since the image of ImageNet natural image database shares 1000 types, with human body behavior type to be identified
It is usually inconsistent, need to the output unit of the full articulamentum (refering in particular to the last layer of Vgg-19 neural network herein) of network into
Row modification.The full connection output layer unit number of Vgg-19 is 1000, need to be revised as the human body behavior type number phase with quasi- classification
Unanimously.
4th step, the transfer training based on human body micro-doppler time-frequency spectrum under sparse constraint
The modified network of full articulamentum is subjected to further supervised learning in human body micro-Doppler feature database.It learns
During habit, the image difference of one side Learning from Nature image and radar return time-frequency spectrum improves the transfer ability between data, separately
On the one hand the significance level for determining each section in network, is marked the lower network channel of importance.
Batch data normalization operation (Batch Normalization) is carried out first between each layer network, batch data
Normalized operation process are as follows:
1, batch data mean value computation
It wherein, is the set of a lot data, input data is that quantity is
2, batch data variance calculates
3, batch data normalizes
4, the flexible translation of batch data
After operation, the parameter learnt corresponding with each network layer can be obtained, the number for controlling and measuring is passed through
It is worth size, that is, can determine the significance level of each network layer.The process of sparse optimization be by add about L1 norm
The process of the transfer learning in time-frequency spectrum is controlled as bound term.
Specifically, the loss function of transfer learning are as follows:
Wherein, it is the data and label of training dataset respectively, is neural network, be the weight of neural network, is to intersect
Entropy function, to constrain term coefficient, the numeric distribution controlled (under different constraint factors, numeric distribution as shown in Figure 3) passes through
The method declined using gradient, can be obtained the value of the parameter sum at loss function minimum.
5th step, depth network pruning and fine tuning
For the neural network after transfer training can according to the importance of quasi- cutting ratio and each component part of network,
Network is accordingly cut.Since each network output channel is corresponding with a particular parameter value, can incite somebody to action
Numerical ordering is carried out, numerical value is smaller, and the network structure at this is more inessential, can be cut.It as a result, can be to neural network
Accurately cut.When the cutting ratio of network is excessive (usually more than the 30% of original structure), the network after reduction need to be existed
The repetition training for carrying out 10-20 times in micro-doppler time-frequency spectrum data set again, can be obtained light-weighted network thus for reality
Border uses.Fig. 4 gives under different cutting ratios, the network area of cutting.Fig. 5 is the performance comparison before and after network pruning.
The above is only a kind of application example of the invention, and the behavior classification being directed in the present invention includes but is not limited to go
It walks, run, shaking one's fists, standing, rebounding, creeping, kicking, eight kinds of skip-forwards movements.The primitive network that can be cut include but
It is not limited to Vgg-19 neural network, convolutional neural networks can be used the present invention and cut.
Claims (6)
1. a kind of human body behavior micro-doppler classification method based on sparse migration network, which comprises the following steps:
1) initial configuration for determining neural network carries out pre-training in ImageNet natural image database；
2) the full articulamentum for modifying neural network keeps neural network output layer unit consistent with quasi- classification micro-doppler classification number；
3) transfer training is carried out on a certain number of micro-doppler image sets based on sparse constraint criterion, it is right in training process
The redundant area of neural network is labeled；
4) redundant area of neural network is subjected to beta pruning, to obtain light-weighted depth network.Light-weighted depth network
Output layer number of network node it is consistent with the human body behavior number of quasi- classification, pass through the light-weighted depth network of detection and export weight
Maximum nodal scheme determines corresponding human body behavior classification.
2. the human body behavior micro-doppler classification method according to claim 1 based on sparse migration network, feature exist
In in step 1), choosing initial configuration of the depth convolutional neural networks Vgg-19 as neural network, the initial configuration includes
16 layers of convolutional layer and 3 layers of full articulamentum, 19 layer network stack gradually, before be convolutional layer, last three layers be full articulamentum；Just
Each layer parameter of beginning structure carries out random initializtion.
3. the human body behavior micro-doppler classification method according to claim 2 based on sparse migration network, feature exist
In the specific implementation process of step 2) includes: to carry out to the output unit of the last layer of depth convolutional neural networks Vgg-19
The full connection output layer unit number of Vgg-19 is revised as consistent with the human body behavior type number of quasi- classification by modification.
4. the human body behavior micro-doppler classification method according to claim 1 based on sparse migration network, feature exist
In in step 4), the specific implementation process by the redundant area progress beta pruning of neural network includes: each net of neural network
Network output channel is corresponding with a particular parameter value γ, γ is carried out numerical ordering, the network at the small preceding N% of logarithm
It is cut.
5. the human body behavior micro-doppler classification method according to claim 4 based on sparse migration network, feature exist
In by the ascending arrangement of parameter value γ, network corresponding to the numerical value to preceding 30% is cut.
6. the human body behavior micro-doppler classification method according to claim 4 based on sparse migration network, feature exist
In when the cutting ratio of network is more than the 30% of original structure, by the network after reduction in micro-doppler time-frequency spectrum data set
10~20 repetition trainings are carried out, again thus to obtain light-weighted depth network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910311070.1A CN110096976A (en) | 2019-04-18 | 2019-04-18 | Human behavior micro-Doppler classification method based on sparse migration network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910311070.1A CN110096976A (en) | 2019-04-18 | 2019-04-18 | Human behavior micro-Doppler classification method based on sparse migration network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110096976A true CN110096976A (en) | 2019-08-06 |
Family
ID=67445142
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910311070.1A Pending CN110096976A (en) | 2019-04-18 | 2019-04-18 | Human behavior micro-Doppler classification method based on sparse migration network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110096976A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111931914A (en) * | 2020-08-10 | 2020-11-13 | 北京计算机技术及应用研究所 | Convolutional neural network channel pruning method based on model fine tuning |
| CN113158880A (en) * | 2021-04-19 | 2021-07-23 | 中国海洋大学 | Deep learning-based student classroom behavior identification method |
| CN113269134A (en) * | 2021-06-17 | 2021-08-17 | 中国空间技术研究院 | Abnormal broadcast identification model and construction method and use method thereof |
| WO2022127907A1 (en) * | 2020-12-17 | 2022-06-23 | Moffett Technologies Co., Limited | System and method for domain specific neural network pruning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
| 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 |
| CN107292246A (en) * | 2017-06-05 | 2017-10-24 | 河海大学 | Infrared human body target identification method based on HOG PCA and transfer learning |
| CN108388850A (en) * | 2018-02-08 | 2018-08-10 | 天津大学 | A kind of human motion recognition method based on k arest neighbors and micro-Doppler feature |
| CN108520199A (en) * | 2018-03-04 | 2018-09-11 | 天津大学 | Based on radar image and the human action opener recognition methods for generating confrontation model |
-
2019
- 2019-04-18 CN CN201910311070.1A patent/CN110096976A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
| 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 |
| CN107292246A (en) * | 2017-06-05 | 2017-10-24 | 河海大学 | Infrared human body target identification method based on HOG PCA and transfer learning |
| CN108388850A (en) * | 2018-02-08 | 2018-08-10 | 天津大学 | A kind of human motion recognition method based on k arest neighbors and micro-Doppler feature |
| CN108520199A (en) * | 2018-03-04 | 2018-09-11 | 天津大学 | Based on radar image and the human action opener recognition methods for generating confrontation model |
Non-Patent Citations (2)
| Title |
|---|
| ZHUANG LIU ET AL.: "Learning Efficient Convolutional Networks through Network Slimming", 《ARXIV:1708.06519V1》 * |
| 袁延鑫 等: "基于卷积神经网络和微动特征的人体步态识别技术", 《信号处理》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111931914A (en) * | 2020-08-10 | 2020-11-13 | 北京计算机技术及应用研究所 | Convolutional neural network channel pruning method based on model fine tuning |
| WO2022127907A1 (en) * | 2020-12-17 | 2022-06-23 | Moffett Technologies Co., Limited | System and method for domain specific neural network pruning |
| CN116438544A (en) * | 2020-12-17 | 2023-07-14 | 墨芯国际有限公司 | System and method for domain-specific neural network pruning |
| CN116438544B (en) * | 2020-12-17 | 2024-04-30 | 墨芯国际有限公司 | System and method for domain-specific neural network pruning |
| CN113158880A (en) * | 2021-04-19 | 2021-07-23 | 中国海洋大学 | Deep learning-based student classroom behavior identification method |
| CN113269134A (en) * | 2021-06-17 | 2021-08-17 | 中国空间技术研究院 | Abnormal broadcast identification model and construction method and use method thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110096976A (en) | Human behavior micro-Doppler classification method based on sparse migration network | |
| CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
| CN107145830B (en) | Hyperspectral image classification method based on spatial information enhancing and deepness belief network | |
| CN109492556A (en) | Synthetic aperture radar target identification method towards the study of small sample residual error | |
| CN109829541A (en) | Deep neural network incremental training method and system based on learning automaton | |
| CN109635790A (en) | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution | |
| CN106651915B (en) | The method for tracking target of multi-scale expression based on convolutional neural networks | |
| CN108734208A (en) | Multi-source heterogeneous data fusion system based on multi-modal depth migration study mechanism | |
| CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
| CN106909938B (en) | Visual angle independence behavior identification method based on deep learning network | |
| CN110222634A (en) | A kind of human posture recognition method based on convolutional neural networks | |
| CN109272107A (en) | A method of improving the number of parameters of deep layer convolutional neural networks | |
| CN107909008A (en) | Video target tracking method based on multichannel convolutive neutral net and particle filter | |
| CN113012811B (en) | Traditional Chinese medicine syndrome diagnosis and health evaluation method combining deep convolutional network and graph neural network | |
| CN106980830A (en) | One kind is based on depth convolutional network from affiliation recognition methods and device | |
| CN111582397A (en) | CNN-RNN image emotion analysis method based on attention mechanism | |
| CN111401261B (en) | Robot gesture recognition method based on GAN-CNN framework | |
| CN110852369B (en) | Hyperspectral image classification method combining 3D/2D convolutional network and adaptive spectrum unmixing | |
| CN113569881A (en) | Self-adaptive semantic segmentation method based on chain residual error and attention mechanism | |
| CN107945210A (en) | Target tracking algorism based on deep learning and environment self-adaption | |
| CN112633257A (en) | Potato disease identification method based on improved convolutional neural network | |
| CN106980831A (en) | Based on self-encoding encoder from affiliation recognition methods | |
| CN108596044B (en) | Pedestrian detection method based on deep convolutional neural network | |
| CN110275147A (en) | Human behavior micro-Doppler classification and identification method based on migration depth neural network | |
| CN112668486A (en) | Method, device and carrier for identifying facial expressions of pre-activated residual depth separable convolutional network |
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: 20190806 |