Anti-interference millimeter-wave radar living body detection method and system based on deep learning

By using deep learning-based PMUnet and TPreUnet networks, the problems of low signal quality and slow processing speed in cold-blooded liveness detection of millimeter-wave radar are solved, achieving efficient and accurate cold-blooded liveness signal recognition, which is suitable for security and human-computer interaction.

CN118673386BActive Publication Date: 2024-11-05ZHEJIANG ANHE SCIENTIFIC INSTRUMENT CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411102841.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-05
Estimated Expiration
2044-08-13

AI Technical Summary

Technical Problem

Existing millimeter-wave radar suffers from problems such as low signal quality, insufficient anti-interference ability, slow processing speed, and lack of cold-blooded liveness datasets in cold-blooded liveness detection, resulting in low classification accuracy and difficulty in meeting the needs of real-time applications.

Method used

We employ deep learning-based PMUnet denoising network and TPreUnet classification network, combined with Physical denoising module, MUNet denoising module, MSFF multi-scale feature fusion block and PreUnet classification network, to perform signal denoising and feature extraction, achieving efficient classification of cold-blooded live animal signals.

Benefits of technology

It improves the accuracy and robustness of millimeter-wave radar liveness detection, effectively identifies cold-blooded live signals, and is suitable for fields such as security and human-computer interaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673386B_ABST
    Figure CN118673386B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for anti-interference millimeter wave radar liveness detection based on deep learning, which relates to the field of target detection technology. The method comprises the following steps: Step 1: Collect radar signal data and construct a data set; Step 2: Construct a PMUnet denoising network, use the PMUnet denoising network to denoise the data set, and obtain a denoised data set; Step 3: Construct a TPreUnet classification network, use the denoised data set to extract features using a sliding window to train the TPreUnet classification network, and obtain a liveness detection model; Step 4: Collect the radar signal to be detected, input the liveness detection model, and obtain the detection result. The present invention can increase the recognition accuracy of millimeter wave radar micro-motion signals and improve the practicability of the system by introducing deep learning for signal interference removal.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and more specifically to a deep learning-based anti-interference millimeter-wave radar living body detection method and system. Background Art

[0002] Millimeter wave radar is often used in target identification, tracking and positioning due to its high-resolution imaging capability, penetration ability and anti-interference ability. However, when identifying micro-motion signals of millimeter wave radar, the removal of interference signals is often done by empirical methods. As a non-contact biometric technology, millimeter wave radar has the advantages of strong penetration and no radiation hazards, making it gradually become one of the research hotspots in the field of biometric recognition. However, traditional millimeter-wave radar technology has technical gaps when it comes to cold-blooded biological signal recognition in complex environments, mainly including the following points: 1. When collecting weak life signals, millimeter-wave radar is often affected by vibration noise, environmental electromagnetic interference, thermal noise, and random noise, resulting in low quality of the original signal. Directly sending the original signal to the neural network for learning will cause the noise to be mistakenly learned as a feature, thereby affecting the classification accuracy; 2. The objects of recognition of traditional millimeter-wave radar life recognition systems are often warm-blooded animals such as humans and dogs. There are some technical gaps when facing cold-blooded live detection; 3. For real-time applications (such as security checkpoints or entry systems), the processing speed of existing technologies may not be sufficient to support fast and accurate live detection, which is particularly prominent in cases where large amounts of data processing and complex algorithm operations are required; 4. There is no public data set for radar data collection of cold-blooded live organisms, and traditional noise reduction models require ideal signals. Therefore, it is difficult to achieve noise reduction for radar data of cold-blooded live organisms using traditional noise reduction models, resulting in reduced classification accuracy, thereby reducing the credibility and practicality of the system.

[0003] Therefore, how to improve the efficiency and accuracy of millimeter wave radar living body detection is a problem that technical personnel in this field need to solve urgently. Summary of the invention

[0004] In view of this, the present invention provides an anti-interference millimeter-wave radar living body detection method and system based on deep learning. The introduction of deep learning for signal interference removal can increase the recognition accuracy of millimeter-wave radar micro-motion signals and improve the practicability of the system.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A method for detecting living bodies by using an anti-interference millimeter-wave radar based on deep learning, comprising the following steps:

[0007] Step 1: Collect radar signal data and build a data set; Step 2: Build a PMUnet denoising network, use the PMUnet denoising network to denoise the data set, and obtain a denoised data set; Step 3: Build a TPreUnet classification network, use the denoised data set to extract features using a sliding window to train the TPreUnet classification network, and obtain a liveness detection model; Step 4: Collect the radar signal to be detected, input it into the PMUnet denoising network for denoising, and then input it into the liveness detection model to obtain the detection result.

[0008] Preferably, the PMUnet denoising network includes a Physical denoising module, a MUNet denoising module and an MSFF multi-scale feature fusion block connected in sequence. The Physical denoising module removes the preliminary background noise of the data set through multiple filtering, and the Physical denoising module realizes phase unwrapping, thereby ensuring that the phase value is within a continuous range and avoiding the problem of phase value jump.

[0009] Preferably, the Physical noise reduction module includes an Unwrap_phase layer, a jump_threshold layer, a signal_detrend function, a Filter filter and a stack_data layer connected in sequence; the Filter filter is composed of an FIR filter and an IIR filter. The Unwrap_phase layer performs phase unwrapping; the jump_threshold detects and adjusts the phase value mutation; the signal_detrend function performs a detrending operation on the unwrapped phase data to remove the linear trend in the data to reduce the impact of slow changes; the Filter filter implements composite filtering and uses data padding to process boundary effects; stack_data stacks the processed phase data with the original amplitude data, and converts the data into npy format to facilitate subsequent deep noise reduction and data training.

[0010] Preferably, the MUNet denoising module includes a UNetConvBlock encoder, a UNetUpBlock decoder and a skip connection block connected in sequence; the UNetConvBlock encoder includes a plurality of encoding subspace modules, each encoding subspace module includes a first convolution layer, a first activation function layer, a second convolution layer, a second activation function layer and a downsampling layer connected in sequence; the UNetUpBlock decoder includes a plurality of decoding subspace modules, each decoding subspace module includes a deconvolution layer, a first convolution layer, a first activation function layer, a second convolution layer and a second activation function layer connected in sequence, and the deconvolution layer of each decoding subspace module is connected to a corresponding downsampling layer of the encoding subspace module through a skip connection module.

[0011] The convolution layer and activation function layer of the UNetConvBlock encoder are used for nonlinear feature extraction. The sampling layer can halve the spatial resolution of the feature map and increase the number of channels accordingly. The deconvolution layer of the UNetUpBlock decoder is used to double the spatial resolution of the feature map and adjust the number of channels appropriately. The output of the deconvolution layer is jump-connected with the output of the corresponding encoder stage to provide a richer information flow. The output feature map of the decoder is gradually restored to the same size and number of channels as the input image. The subspace module remaps the input feature map to a low-dimensional representation space, thereby reducing noise sensitivity and making the feature map have a higher information density and a more effective representation space. The jump connection block acts between the shallow and deep layers, retaining low-level detail information while increasing the nonlinearity of the network, allowing the network to fuse multi-scale information and reduce noise more efficiently. The encoding subspace module is a UNetConvBlock block. The UNetConvBlock encoder includes UNetConvBlock1, UNetConvBlock2, UNetConvBlock3, and UNetConvBlock4. Each UNetConvBlock includes the first convolution layer, the first activation function layer, the second convolution layer, the second activation function layer, and the downsampling layer. The MSFF multi-scale feature fusion block is inserted between different UNet ConvBlocks, which can effectively improve the network's receptive field and capture more context information, thereby improving the noise reduction performance. The decoding subspace module is a UNetUpBlock block. The UNetUpBlock decoder includes UNetUpBlock1, UNetUpBlock2, UNetUpBlock3, and UNetUpBlock4. Each UNetUpBlock includes a deconvolution layer, a first convolution layer, a first activation function layer, a second convolution layer, and a second activation function layer. Each UNetConvBlock is connected to a UNetUpBlock through a jump connection block. The jump connection block includes a UNetConvBlock block.

[0012] Preferably, the MSFF multi-scale feature fusion block includes a first convolution layer, a first activation function layer, a second convolution layer, a dilated convolution layer, an average pooling layer, a splicing layer, and a third convolution layer connected in sequence; the dilated convolution layer includes a 1×3 convolution layer with dilation rates of 6, 9, and 12. The dilated convolution layer with multiple dilation rates is used to expand the receptive field and increase the feature extraction receptive field, and then the translation invariance of the model is improved through feature splicing and average pooling operations.

[0013] Preferably, the TPreUnet classification network includes a PreUnet module, an Encoder module, a softmax layer, and two linear layers (linear1, linear2) connected in sequence. The PreUnet module is used to receive input data and pass it to the next layer; the Encoder module containing multiple blocks is used for multi-layer feature extraction and representation learning; the softmax layer and two linear layers (linear1, linear2) are used to generate the final classification output, and the output detection result is live or non-live.

[0014] Preferably, the network structure of the PreUnet module is an initial convolution layer, an intermediate convolution layer, a residual block and a classifier Tokens connected in sequence; the initial convolution layer includes two Conv1d, BatchNorm1d and Gelu activation functions connected in sequence, the data is first passed through two Conv1d and the convolution results are batch normalized using BatchNorm1D, and then the Gelu activation function is used to achieve nonlinear activation; the intermediate convolution layer includes Maxpool1d, Dropout, Conv1d and Gelu activation functions connected in sequence, Maxpool1d is used to achieve pooling operation, Dropout prevents overfitting, Conv1d and Gelu activation functions are continued to be added, and the output feature map is sent to the residual block; the classifier Tokens includes token1, token2 and token_bricks connected in sequence.

[0015] Preferably, the Encoder module includes several LayerNorm normalization layers and several Block blocks connected in sequence; each Block block includes an Attention module and a feedforward neural network MLP; the LayerNorm normalization layer includes att_norm and ffn_norm, wherein att_norm performs layer normalization on the data h in the denoised dataset to ensure that the mean and variance of the data are stable and calls the Attention module to calculate the attention weight in the multi-head attention mechanism, and then uses a residual module to connect the data h in the denoised dataset and the Attention module, and adds the data h to the output of the attention mechanism, ffn_norm performs layer normalization on the output after the attention mechanism, and passes the data through the feedforward neural network MLP, and then uses a residual module to connect the output of the attention mechanism and the output of the feedforward neural network MLP, and adds the data h to the output of the feedforward neural network MLP. The feedforward neural network MLP consists of two linear layers and an activation function. In the forward method of the Encoder class, the data will pass through all the Block blocks in sequence, and the output of each Block block will be used as the input of the next Block block. The final output data is processed by all the Block blocks, and the obtained hidden_states is normalized by the layer as the output of the Encoder module.

[0016] Preferably, in the two linear layers, linear layer linear1 constructs a fully connected layer head1, and linear layer linear2 constructs a fully connected layer head2. self.head = nn.Sequential(nn.Linear(,))Here, head1 and head2 are nn.Sequential modules composed of one linear layer respectively.

[0017] Preferably, when training the TPreUnet classification network, a cross entropy loss function and an Adam optimizer are used for optimization to obtain a liveness detection model.

[0018] The invention discloses an anti-interference millimeter-wave radar liveness detection system based on deep learning, comprising a power supply, a host computer, a data acquisition fixed base, a data acquisition angle control knob, an angle control accessory, a radar board fixed bracket, a millimeter-wave radar sensor, a signal processor, a power supply and an interactive interface; the host computer is loaded with a PMUnet noise reduction network and a TPreUnet classification network; the data acquisition angle control knob and the radar board fixed bracket are fixed on the angle control accessory, the angle control accessory is rotatably fixed on the data acquisition fixed base, the millimeter-wave radar sensor is fixed on the radar board fixed bracket, and a connected signal processor, a power supply and an interactive interface are arranged on the millimeter-wave radar sensor, and the power supply and the interactive interface are respectively connected to the power supply and the host computer; the millimeter-wave radar sensor is controlled by the data acquisition angle control knob to collect a number of radar signal data at different angles, and the data is transmitted to the host computer through the signal processor, the power supply and the interactive interface in turn, and the detection result is obtained after being processed by the PMUnet noise reduction network and the TPreUnet classification network.

[0019] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a deep learning-based anti-interference millimeter-wave radar liveness detection method and system for detecting cold-blooded liveness signals, wherein a deep learning algorithm is used to efficiently process and extract features of weak life signals collected by the millimeter-wave radar, perform anti-interference training, and realize signal binary classification (classification into living and non-living classes) to improve the accuracy and robustness of liveness detection. The present invention has broad application prospects in security, human-computer interaction and other fields. Specifically, by controlling the signal acquisition range, the millimeter-wave radar signals within the specified range are collected to construct a data set; a PMUnet denoising network with mixed anti-interference for micro-motion data is designed to denoise the data set, which is mainly reflected in the use of physical information denoising as the pre-processing of the UNetD network, realizing composite filtering and detrending operations, which will make the UNetD network more accurate in extracting features; at the same time, the UNetD network is embedded with the MSFF multi-scale feature fusion block, which is reflected in the selection of dilated convolutions of different scales to increase the receptive field, so that multi-scale features are extracted and fused, thereby capturing the characteristics of living data. On this basis, by setting appropriate subspace modules, the feature map has a higher information density and a more effective representation space, thereby effectively reducing noise; a TPreUnet classification network suitable for weak signals is designed to perform liveness detection on the collected millimeter-wave radar signals to achieve binary classification. Specifically, PreUnet is used to receive input data and pass it to the next layer; the Encoder structure containing multiple blocks is used for multi-layer feature extraction and representation learning; the softmax layer and two linear layers are used to generate the final classification output. The entire network uses a sliding window to traverse the network to extract features, which is equivalent to dividing each feature map into many small sub-maps and adding tokens. It is trained under a multi-head attention mechanism to achieve efficient classification of millimeter-wave radar cold-blooded living data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0021] Figure 1 A flow chart of a method for detecting living things using an anti-interference millimeter-wave radar based on deep learning provided by the present invention;

[0022] Figure 2 A schematic diagram of the denoising network structure of the millimeter wave radar live data set provided by the present invention;

[0023] Figure 3A schematic diagram of the classification network structure of the millimeter wave radar living body data set provided by the present invention;

[0024] Figure 4 A schematic diagram of the structure of an anti-interference millimeter-wave radar livingness detection system based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The embodiment of the present invention discloses a method for detecting living bodies using an anti-interference millimeter-wave radar based on deep learning. Figure 1 As shown, specifically including:

[0027] S1: Collect radar signal data and build a data set; S2: Build a PMUnet denoising network, use the PMUnet denoising network to denoise the data set, and obtain a denoised data set; S3: Build a TPreUnet classification network, use the denoised data set to extract features using a sliding window to train the TPreUnet classification network, and obtain a liveness detection model; S4: Collect the radar signal to be detected, input it into the PMUnet denoising network for denoising, and then input it into the liveness detection model to obtain the detection result.

[0028] On the other hand, in a specific embodiment, the steps of the anti-interference millimeter wave radar living body detection method based on deep learning are as follows:

[0029] S1: Collect radar signal data and construct a data set; collect different types of living data, including living bodies under random environmental interference, living bodies in solid containers, living bodies in liquid containers, and living bodies under solid-liquid interference; use millimeter-wave radar sensors to collect 9,000 cold-blooded living body signals and 10,000 inanimate random data, and use a horizontal 280-degree azimuth to collect data to increase data integrity; specifically, when the millimeter-wave radar sensor collects a single piece of data, it performs 280-degree surround collection on the plane to increase the information collection range; store the living body radar signal data collected under different background environmental conditions as a csv file type. The specific content of the csv file is: rows represent the time axis, columns represent position changes, and the phase and amplitude can be read from each row and column. At the same time, the corresponding data labels are assigned, with the living body data labeled 1 and the inanimate random data labeled 0. 80% of the data set is randomly taken as the training set, 10% as the validation set, and 10% as the test set to facilitate subsequent data training.

[0030] S2: Design the PMUnet denoising network: a Unet network that combines Physical denoising with the MSFF (Multi-Scale Feature Fusion) multi-scale feature fusion block; send the data set obtained in S1 into the PMUnet denoising network for denoising to obtain the denoised data.

[0031] The designed PMUnet denoising network mainly includes three architectures: Physical denoising module, MUNet denoising network, and MSFF (Multi-Scale Feature Fusion) multi-scale feature fusion block. Figure 2 shown.

[0032] In the Physical denoising module, multiple filters are used to remove the initial background noise of the data set. In this module, phase unwrapping (Unwrap_phase) needs to be implemented to ensure that the phase value is within a continuous range to avoid phase value jumps. During the unwrapping process, jump_threshold is used to detect and adjust phase value mutations. The signal_detrend function is then used to detrend the unwrapped phase data to remove linear trends in the data to reduce the impact of slow changes. In this module, Filter is used to implement composite filtering, and padding data is used to fill in the boundary effect. The Filter filter composed of FIR and IIR filter structures is used to implement low-pass filtering and reverse filtering of the data. Stack_data is used to stack the processed phase data with the original amplitude data, and the data is converted into npy format to facilitate subsequent deep denoising and data training.

[0033] In the MUNet denoising network, it includes a downsampling path (UNetConvBlock encoder), an upsampling path (UNetUpBlock decoder), and a jump connection block (JumpBlock); the upsampling path and the downsampling path are respectively composed of several subspace modules (Subspace); Unet upsampling space and downsampling space are used to realize feature extraction and noise removal.

[0034] In the downsampling path (UNetConvBlock encoder) structure, including UNetConvBlock1, UNetConvBlock2, UNetConvBlock3, and UNetConvBlock4, each UNetConvBlock contains the first layer of convolution, whose convolution kernel size is 3×3, stride is 1, padding is 1, bias is True, and the activation function uses LeakyReLU (slope is 0.2) for nonlinear feature extraction; the second layer of convolution, the convolution kernel size is 3×3, stride is 1, padding is 1, bias is True, and the activation function uses LeakyReLU (slope is 0.2); the third layer is the downsampling layer (downsampling), whose convolution kernel size is 4×4, stride is 2, padding is 1, bias is False, which can halve the spatial resolution of the feature map and increase the number of channels accordingly. Each UNetConvBlock remaps the input feature map to a low-dimensional representation space, thereby reducing noise sensitivity and making the feature map have higher information density and a more effective representation space. The size of its output feature map is halved and the number of channels increases step by step until it reaches the deepest layer.

[0035] The upsampling path (UNetUpBlock decoder) structure includes UNetUpBlock1, UNetUpBlock2, UNetUpBlock3, and UNetUpBlock4. Each UNetUpBlock contains a deconvolution layer (transposed convolution) to double the spatial resolution of the feature map and adjust the number of channels appropriately. The convolution kernel size is 2×2, the stride is 2, the padding is 0, and the bias is True. For the first layer of convolution, the convolution kernel size is 3×3, the stride is 1, the padding is 1, the bias is True, and the activation function uses LeakyReLU (with a slope of 0.2). For the second layer of convolution, the convolution kernel size is 3×3, the stride is 1, the padding is 1, the bias is True, and the activation function uses LeakyReLU. The output of the deconvolution layer is skipped with the output of the corresponding encoder stage. connection) to provide a richer information flow, the output feature map of the decoder is gradually restored to the same size and number of channels as the input image; each UNetConvBlock is connected to a UNetUpBlock through a skip connection block.

[0036] In the JumpBlock structure: it contains a UNetConvBlock block, which acts between the shallow layer and the deep layer, retains the low-level detail information, and increases the nonlinearity of the network, so that the network can fuse multi-scale information and reduce noise more efficiently.

[0037] In the MSFF (Multi-Scale Feature Fusion) multi-scale feature fusion block structure, dilated convolutions with various dilation degrees are used to expand the receptive field; first use 1×3 convolution and use Relu as the activation function, then use a 1×1 convolution layer as the first convolution for multi-scale feature extraction, and use 1×3 dilated convolutions with dilation rates of 6, 9, and 12 to increase the feature extraction receptive field; concatenate features and set an average pooling operation to improve the translation invariance of the model.

[0038] S3: Design the TPreUnet classification network, put the denoised dataset processed by S2 and the dataset constructed in S1 into the classification network for training respectively, use the sliding window to extract features to train the network, and train the best_model to achieve the classification of cold-blooded living data; the dataset in S1 participates in the network training to prove the performance of the denoising network and the necessity of its existence.

[0039] The designed TPreUnet classification network mainly includes three architectures: PreUnet accepts input data and passes it to the next layer; the Encoder structure containing multiple blocks is used for multi-layer feature extraction and representation learning; softmax and two linear layers linear1 / 2 are used to generate the final classification output; the entire network uses a sliding window to traverse the network for feature extraction, which is equivalent to dividing each feature map into many small sub-maps and adding tokens. It is trained under a multi-head attention mechanism to achieve efficient classification of millimeter-wave radar cold-blooded living data.

[0040] In PreUnet, it includes the initial convolution layer, the intermediate convolution layer, the residual block and the classifier Tokens; when initializing the convolution layer data, it first passes through two Conv1d and uses BatchNorm1D to batch normalize the convolution results, and uses the Gelu activation function to implement nonlinear activation. The first conv1d is the first layer of convolution, which is used for feature extraction. By performing convolution operations on the input data, it can capture local spatial or temporal features and maintain the dimension. Due to stride=1 and padding=1, the convolution operation will not change the width and depth of the input data, thereby maintaining the size of the data; The convolutional layer includes Maxpool1d, Dropout, Conv1d and Gelu activation functions connected in sequence. Maxpool1d is used to implement pooling operations. Dropout prevents overfitting. Conv1d and Gelu activation functions are added, and the output feature map is sent to the residual block (Resblock); Resblock contains the main path and the shortcut path. The main path (self.left) consists of two convolutional layers (nn.Conv1d) and two batch normalization layers (nn.BatchNorm1d), which are used to process input data. The shortcut path ( self.shortcut) uses 1×1 convolution to adjust the number of channels and size when the number of input and output channels does not match or the stride is not 1. It is used to add the output of the main path of the residual block and the output of the shortcut path of the residual block element by element, and applies the ReLU activation function to the addition result to increase nonlinearity; the data then enters the classifier Tokens, including token1, token2, and token_bricks connected in sequence. Different classification tokens are inserted at the beginning of the data, and token_bricks are inserted to introduce additional information and sequence features to help the network better represent and learn features; the classifier Tokens defines classification tokens, classifer_token_a, classifer_token_c, and classifier_token_bricks when the network is initialized. When the data is processed by the forward method of the model, these tokens are added to the beginning of the input data. After being processed by the Transformer encoder, the classification tokens are merged into the encoded data, so that the network can integrate the classification information into the processing flow and use these tokens as specific tags or information for processing.

[0041] In the Encoder structure, each Encoder contains multiple LayerNorm normalization layers and multiple Block blocks. Each Block block contains an Attention module and a feedforward neural network MLP. The Attention module uses a multi-head attention mechanism to extract and associate features. The feedforward neural network MLP uses the MLP class to implement nonlinear transformation and feature mapping of the fully connected layer. Multiple LayerNorm normalization layers are connected in sequence to att_norm and ffn_norm, where att_norm performs layer normalization on the input x to ensure that the mean and variance of the data are stable and calls the Attention class to calculate the attention weight in the multi-head attention mechanism, and then performs a residual connection to add the original input h to the output of the attention mechanism. ffn_norm performs layer normalization on the output after the attention mechanism and passes the data through the feedforward neural network MLP, and then performs a residual connection to add the original input h to the output of the feedforward neural network MLP. The feedforward neural network MLP consists of two linear layers and an activation function. In the forward method of the Encoder class, the data will pass through all the Block blocks in sequence, and the output of each Block block will be used as the input of the next Block block. The final output data is processed by all the Block blocks, and the obtained hidden_states is normalized by the layer as the output of the Encoder module.

[0042] In softmax and the two linear layers linear1 / 2, softmax is used to output probability distribution. Head1 constructed by linear layer linear1 and head2 constructed by linear layer linear2 are fully connected layers for classification tasks. Both head1 and head2 are sequence modules containing a single linear self.head=nn.Sequential(nn.Linear(,)); the linear transformation of head1 maps the input features to a lower dimensional space, reducing the complexity and computational complexity of the model while retaining important feature information; the linear transformation of head2 converts the reduced dimensionality features into probability distributions for category predictions, and generates the probability of each category through softmax, thereby completing the classification task.

[0043] S4: The collected millimeter-wave radar signal to be detected is input into the PMUnet denoising network for denoising, and then the trained best_model is used for detection to determine whether it is a living object or a non-living object.

[0044] Furthermore, command line parameter parsing was performed during training, using the argparse library to parse command line parameters, including batch size, training rounds, learning rate, whether to use GPU, etc.; data loading, conversion, merging and other data processing functions were set up; the data set and test set were then divided and loss functions and optimizers were added, that is, the cross entropy loss function and Adam optimizer were defined, and training was carried out after the training rounds were set.

[0045] Furthermore, during the training process, the goal is to minimize the cross entropy: that is, to minimize the cross entropy between the true distribution and the predicted distribution. When the model's prediction is close to the true probability distribution, the cross entropy approaches zero. This requirement is usually used in conjunction with optimization algorithms such as gradient descent. By gradually adjusting the model's parameters during the training process to reduce the cross entropy, the prediction is closer to the true distribution, thereby improving the classification performance.

[0046] The anti-interference millimeter-wave radar living body detection method based on deep learning of the present invention produces a data set by collecting the life signals of cold-blooded living bodies; uses the PMUnet denoising network to denoise the cold-blooded living body data set to improve the accuracy of the classification network; uses the TPreUnet classification network to effectively distinguish the negative and positive (i.e., non-living and living) of the samples, and the model has a low order of magnitude, is adaptable to a variety of scenarios, and is more intelligent and efficient in practical applications.

[0047] On the other hand, in a specific embodiment, an anti-interference millimeter wave radar living body detection system based on deep learning is as follows Figure 4 As shown, it includes: power supply 1, host computer 2, data acquisition fixed base 3, data acquisition angle control knob 4, angle control accessory 5, radar board fixed bracket 6, millimeter wave radar sensor 7, signal processor 8, power supply and interactive interface 9; data acquisition angle control knob 4 and radar board fixed bracket 6 are fixed on angle control accessory 5, angle control accessory 5 is rotatably fixed on data acquisition fixed base 3, millimeter wave radar sensor 7 is fixed on radar board fixed bracket 6, millimeter wave radar sensor 7 is provided with connected signal processor 8, power supply and interactive interface 9; power line is led from positive and negative poles of power supply 1 to power supply and interactive interface 9 respectively, and signal line is led from host computer 2 to power supply and interactive interface 9. Deploy radar startup code in host computer 2, set configuration file parameters, deploy data acquisition and decode storage code, so that the system can read data normally, and send the collected data to the host computer after proper processing.

[0048] Furthermore, when the millimeter-wave radar sensor collects a single piece of data, the data collection angle control knob 4 is used to control the millimeter-wave radar sensor 7 to perform 280-degree surround collection of the package in a plane, thereby increasing the information collection range; the live radar signal data collected under different background environmental conditions is stored as a csv file type. The specific content of the csv file is: the row represents the time axis, and the column represents the position change. The phase and amplitude can be read from each row and column. At the same time, the labels of the corresponding data are assigned to facilitate subsequent data training.

[0049] Furthermore, the millimeter-wave radar sensor 7 is used to collect life signals (tracheal contraction, micro-motion signals, etc.) of cold-blooded living bodies and make a data set; the data set is sent to the anti-interference PMUnet noise reduction network of the host computer 2 for training to achieve noise reduction processing; the noise-reduced signal is sent to the TPreUnet classification network for training, and finally it is determined whether the signal contains a living signal.

[0050] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0051] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based anti-interference millimeter wave radar liveness detection method, characterized in that: The following steps are involved: Step 1: Collect radar signal data and build a data set; Step 2: Construct a PMUnet denoising network and use it to denoise the data set to obtain a denoised data set; the PMUnet denoising network includes a Physical denoising module, a MUNet denoising module and an MSFF multi-scale feature fusion block connected in sequence; The Physical denoising module includes an Unwrap_phase layer, a jump_threshold layer, a signal_detrend function, a Filter filter and a stack_data layer connected in sequence; the Filter filter is composed of an FIR filter and an IIR filter; the MUNet denoising module includes a UNetConvBlock encoder, a UNetUpBlock decoder and a jump connection block connected in sequence; the UNetConvBlock encoder includes a number of encoding subspace modules, each of which includes a first convolution layer, a first activation function layer, a second convolution layer, a second activation function layer and a downsampling layer connected in sequence; the UNetUpBlock decoder includes a number of decoding subspace modules, each of which includes a deconvolution layer, a first convolution layer, a first activation function layer, a second convolution layer and a second activation function layer connected in sequence, and the deconvolution layer of each decoding subspace module is connected to a downsampling layer of a coding subspace module through a jump connection module; the MSFF multi-scale feature fusion block includes a first convolution layer, a first activation function layer, a second convolution layer, an expanded convolution layer, an average pooling layer, a splicing layer and a third convolution layer connected in sequence; Step 3: Construct a TPreUnet classification network, and use the denoising data set to extract features using a sliding window to train the TPreUnet classification network to obtain a liveness detection model; the TPreUnet classification network includes a PreUnet module, an Encoder module, a softmax layer, and two linear layers connected in sequence; the PreUnet module includes an initial convolutional layer, an intermediate convolutional layer, a residual block, and classifier Tokens connected in sequence; the initial convolutional layer includes two Conv1d, BatchNorm1d, and Gelu activation functions connected in sequence; the intermediate convolutional layer includes Maxpool1d, Dropout, Conv1d, and Gelu activation functions connected in sequence; the classifier Tokens includes token1, token2, and token_bricks connected in sequence; the Encoder module includes several LayerNorm normalization layers and several Block blocks connected in sequence; each Block block includes an Attention module and a feedforward neural network MLP; Step 4: Collect the radar signal to be detected, input it into the PMUnet denoising network for denoising, and then input it into the living body detection model to obtain the detection result.

2. The method for detecting living bodies by using anti-interference millimeter-wave radar based on deep learning according to claim 1, characterized in that: The dilated convolutional layers include 1×3 convolutional layers with dilation rates of 6, 9, and 12 respectively.

3. The method for detecting living bodies by using anti-interference millimeter-wave radar based on deep learning according to claim 1, characterized in that: One of the two linear layers constructs the fully connected layer head1, and the other linear layer constructs the fully connected layer head2.

4. A deep learning-based anti-interference millimeter-wave radar liveness detection system, characterized in that: The anti-interference millimeter-wave radar liveness detection method based on deep learning as described in any one of claims 1 to 3 above includes a power supply, a host computer, a data acquisition fixed base, a data acquisition angle control knob, an angle control accessory, a radar board fixed bracket, a millimeter-wave radar sensor, a signal processor, a power supply and an interactive interface; the host computer is loaded with a PMUnet noise reduction network and a TPreUnet classification network; the data acquisition angle control knob and the radar board fixed bracket are fixed on the angle control accessory, the angle control accessory is rotatably fixed on the data acquisition fixed base, the millimeter-wave radar sensor is fixed on the radar board fixed bracket, and the millimeter-wave radar sensor is provided with a connected signal processor, a power supply and an interactive interface, and the power supply and the interactive interface are respectively connected to the power supply and the host computer; the millimeter-wave radar sensor is controlled by the data acquisition angle control knob to collect a number of radar signal data at different angles, and is transmitted to the host computer through the signal processor, the power supply and the interactive interface in turn, and the detection result is obtained after being processed by the PMUnet noise reduction network and the TPreUnet classification network.

Citation Information

Patent Citations

  • In-vehicle living body detection method and system based on millimeter wave radar

    CN113093165A

  • Face living body detection model safety test method and system based on structured light

    CN116612374A

  • Radar signal processing method based on CNN-CBAM shrinkage dichotomy network

    CN117572379A

  • Millimeter wave radar fall detection method based on Transform and multi-scale convolutional neural network

    CN118050703A

  • End-to-end vehicle-mounted millimeter wave radar interference suppression method and system based on deep learning

    CN118275986A