Living body detection method and device based on CIR signal of UWB digital key

Abstract

The application discloses a kind of based on UWB digital key's CIR signal's living body detection method and device, belong to in-car living body detection technical field, this application is according to the equipment information of transmitting UWB and receiving UWB, through filtering, artificial intelligence algorithm is solved the change of environment, the respiration of human body is perceived and estimated to realize living body detection, it utilizes artificial intelligence to identify effective data in relevant signal, with higher precision, can effectively suppress false alarm, without increasing equipment cost, solve the existing in-car living body detection method mainly is based on the computer vision technology of video signal, this technique has defects in protecting user privacy, and there is the case where camera is blocked and leads to algorithm failure, or based on pressure sensor in-car living body detection technology, but this method cannot distinguish human body and goods well, there is the problem of false alarm.

Description

Technical Field

[0001] This invention belongs to the field of in-vehicle liveness detection technology, specifically relating to a liveness detection method and device based on CIR signals from a UWB digital key. Background Technology

[0002] In real-life and industrial applications, monitoring and identifying human behavior is becoming increasingly important, especially in fields such as healthcare, security surveillance, and smart homes. Passive human activity identification has gradually gained attention; it uses sensors to monitor human activity without requiring active human intervention.

[0003] UWB technology originated from pulse communication technology in the 1960s, primarily used in military radar, positioning, and communication systems with low intercept / detection rates. Key features of UWB technology include high data transmission rates (up to 1 Gbit / s), strong resistance to multipath interference, low power consumption, low cost, strong penetration, low interception rate, and spectrum sharing with other existing wireless communication systems. Therefore, UWB technology has become the preferred technology for wireless personal area network (PAN) communication. UWB digital keys are also used in automobiles. Our system uses off-the-shelf UWB devices that have been integrated into smart key solutions used by automakers.

[0004] UWB Corona Reflection (CIR) is a signal transmitted between a UWB transmitter and receiver that reflects objects in the environment. CIR reflects environmental characteristics, including any changes occurring within it, such as human movement. Attempts have been made to enhance the sensing capabilities of UWB wireless devices by constructing multi-static radar networks. Meter-level passive tracking has been demonstrated using CIR.

[0005] Existing methods for in-vehicle liveness detection primarily rely on computer vision technology based on video signals. This technology has shortcomings in protecting user privacy and is susceptible to algorithm failure due to camera obstruction. There are also in-vehicle liveness detection technologies based on pressure sensors, but these methods cannot effectively distinguish between humans and cargo, and are prone to false alarms. Summary of the Invention

[0006] To address the shortcomings of the prior art, this invention provides a method and apparatus for liveness detection based on CIR signals from a UWB digital key.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] According to a first aspect of this specification, a liveness detection method based on a UWB digital key CIR signal is provided, comprising the following steps:

[0009] Step S1: Measure the CIR signal over a time period, upsample and interpolate the complex signal using Fourier transform and its inverse transform, and align the overall CIR signal data by aligning the phase of the point with the maximum amplitude.

[0010] Step S2: The CIR signal aligned with the last overall data acquired in step S1 is filtered using a high-pass filter and a Hanning window.

[0011] Step S3: When collecting data, randomly select sampling points, and perform a fast Fourier transform on the filtered CIR signal in step S2 with the sampling points as the center to obtain the velocity-distance map;

[0012] Step S4: Based on the velocity-distance map in Step S3, a training set, a validation set, and a test set are constructed. The neural network is trained, a weight file is generated, and the weight file is deployed to an embedded Linux platform. The classification result is obtained based on the collected CIR data. The neural network, in the order of data processing, includes a data input part, a CNN network, a connection between the CNN network and the LSTM network, an LSTM network, and a result data network. The data processing result of the CNN network includes the result after the first convolution, the result after the first pooling, and the results after multiple convolutions and pooling. The connection between the CNN network and the LSTM network outputs a channel*H*W three-dimensional matrix. This channel*H*W three-dimensional matrix extracts the features in the spatial representation of a single target image. The LSTM network comprises an input gate, a forget gate, a cell state, and an output gate. The input gate determines which new information needs to be stored in the cell state. It includes a sigmoid layer and a tanh layer. The output of the sigmoid layer determines which values ​​will be updated, and the tanh layer creates a new candidate value vector that may be added to the state. The forget gate determines which information needs to be forgotten or discarded. The cell state retains information over a long period. The output gate determines the output value. The LSTM network first runs a sigmoid layer to determine which part of the cell state will be output. Then, the cell state is passed through the tanh layer and multiplied by the output of the sigmoid gate. Finally, only the part that determines the output is output.

[0013] Furthermore, step S1 specifically includes the following:

[0014] Step S11: Use two NXP NCJ29D5 devices, one as a transmitter and the other as a receiver. The frequency range of these devices is 6.0 GHz to 6.5 GHz. Use the ultra-wideband chip NXP NCJ29D5 to measure CIR information. Read the CIR information from the NXP NCJ29D5 device through the CAN bus driver. Each time, read a complex I and Q signal with a length of 100, and the sampling frequency is 20Hz. The interval between the 100 fast time signals is 1ns, corresponding to a distance of 0.3m.

[0015] Step S12: Use Fast Fourier Transform to convert the complex signal with CIR information in Step S11 from the time domain to the frequency domain, add zeros to the left and right edges of the Fast Fourier Transform output, and then perform Inverse Fast Fourier Transform to return to the time domain, completing the upsampling and interpolation of the complex signal.

[0016] Step S13: Align the sampling point with the largest amplitude to the x-axis of the complex plane to achieve overall data alignment.

[0017] Furthermore, step S2 specifically includes the following:

[0018] Step S21: The CIR signal aligned with the last acquired overall data in step S1 is subjected to high-pass filtering, and its system function is: The value of k ranges from [0,1]. When k is close to 0, the low-frequency suppression effect is strong, and only the signal corresponding to fast-moving objects will be preserved. When k is close to 1, the low-frequency suppression effect is weak, and slower-moving objects will also be preserved.

[0019] Step S22: The signal generated in step S21 after high-pass filtering is data in 128*512 format. This data is then processed using a Hanning window. The mathematical formula for the Hanning window is: The value of N is 128, consistent with the slow time scale, and the value of n is in the range of [0, N-1]. The Hanning window reduces the sidelobe level in the frequency domain representation of the signal, making it easier to distinguish close frequencies. The generated Hanning window is applied to a class of signals by column through element-wise multiplication, preparing for Fourier transform processing of CIR signals.

[0020] Furthermore, the specific content of step S3 is as follows:

[0021] During data acquisition, sampling points are randomly selected, and a Fast Fourier Transform (FFT) is performed on the filtered CIR signal from step S2, centered on these sampling points. This FFT converts the complex sequence into the magnitude of the corresponding velocity component; here, the velocity resolution... and maximum speed From the formulas respectively

[0022]

[0023] and

[0024] Confirmed, among which , The carrier frequency is 6 GHz, c is the speed of light, and fs is the sampling frequency of 20 Hz.

[0025] The formula for converting data obtained from a Fourier transform to a logarithmic coordinate system is as follows: , where x is the data obtained from the Fourier transform;

[0026] Superimposing S on the slow time dimension yields a 128*512 time-distance map as a feature. 128 represents the time dimension, corresponding to 6.4 seconds, and 512 represents the distance dimension, corresponding to 2.4 meters. After upsampling by 64 times, there are 6400 fast-time sampling points. Only the 64 points before the first peak and the subsequent 448 points are selected. The 448 points correspond to 2.1 meters after the first peak, which is an ellipsoid where the sum of the distances to the two devices is less than D+2.1 meters, where D is the distance between the two devices. Since the sampling rate is 20Hz, the 128 sampling points correspond to 6.4 seconds of time. A 30.4-second data segment is selected as the judgment criterion, and a velocity-distance map is generated every two seconds, for a total of 13 velocity-distance maps.

[0027] Furthermore, the specific content of "constructing a training set, a validation set, and a test set based on the velocity-distance map in step S3" in step S4 is as follows:

[0028] After confirming the generation method of the speed-distance map in step S3, the collected data includes data with people inside the vehicle and data with no one inside the vehicle. During data collection, training set, validation set, and test set are recorded separately, with the number of data in each set following a 10:2:2 ratio. The input data is normalized; for all data M, their mean and variance (std) are calculated. The normalization formula is as follows:

[0029] .

[0030] Furthermore, the data input part of the neural network represents a single 128*512 CNN input data, with the vertical axis representing slow time (size 128) and the horizontal axis representing fast time (size 512). Each training sample has 13 such input matrices. These 13 128*512 input matrices generate 13 inputs to the LSTM network, corresponding to a stride of 13 for the LSTM network. Each time, it receives data of one stride, and after passing through cascaded LSTM cells, it obtains the state at step i. When the model progresses to step 13, the output hidden state becomes the input of the result data network. The result data network is the final fully connected output layer, outputting a batch_size*2 matrix. The two columns represent the scores for people in the car and the scores for people not in the car, respectively. One-hot labels are used, and cross-entropy is used as the loss function.

[0031] According to a second aspect of this specification, a liveness detection device based on a CIR signal of a UWB digital key is provided, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement the liveness detection method based on a CIR signal of a UWB digital key.

[0032] The beneficial effects of this invention are as follows: Based on the information from the UWB transmitting and receiving devices, this invention uses filtering and artificial intelligence algorithms to solve for environmental changes, senses and estimates human respiration, and thus achieves liveness detection. It utilizes artificial intelligence to identify valid data in relevant signals, achieving high accuracy and effectively suppressing false alarms without increasing equipment costs. Specifically, the UWB digital key achieves decimeter-level positioning through the active positioning key or the UWB chip in the mobile phone, thereby realizing the function of sensing the approach of the car owner and automatically unlocking the vehicle. To achieve the active positioning function, UWB base stations need to be deployed on the front and rear bumpers of the car, as well as the roof of the car. This invention uses the function of sensing the channel during communication using these base stations to achieve liveness detection inside the vehicle. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart of a liveness detection method based on a UWB digital key CIR signal provided as an exemplary embodiment;

[0035] Figure 2a A state diagram of CIR before upsampling is provided for an exemplary embodiment;

[0036] Figure 2b A state diagram after CIR upsampling is provided for an exemplary embodiment;

[0037] Figure 3 A state diagram of two sets of upsampled data aligned as an exemplary embodiment;

[0038] Figure 4 A flowchart illustrating the generation of a velocity-distance map, provided as an exemplary embodiment;

[0039] Figure 5 A structural diagram of a CNN+LSTM model provided as an exemplary embodiment;

[0040] Figure 6 for Figure 3 A detailed structural diagram of the LSTM module in the middle;

[0041] Figure 7 A structural diagram of a liveness detection device based on a UWB digital key CIR signal, provided as an exemplary embodiment. Detailed Implementation

[0042] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0043] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0044] like Figure 1 As shown, a liveness detection method based on CIR signals of UWB digital keys is provided. It consists of two parts: the first part is data preprocessing, which includes reading UWB CIR signals via CAN bus, upsampling, alignment, and generating velocity-distance maps; the second part includes dataset generation, model parameter tuning, model deployment, and outputting classification results.

[0045] Specifically, the liveness detection method based on the CIR signal of the UWB digital key includes the following steps:

[0046] Step S1: Measure the CIR signal over a time period, upsample and interpolate the complex signal using Fourier transform and its inverse transform, and align the overall CIR signal data by aligning the phase of the point with the maximum amplitude.

[0047] In one embodiment, step S1 specifically includes the following:

[0048] Step S11: Two NXP NCJ29D5 devices are used, one as a transmitter and the other as a receiver. The frequency range of these devices is 6.0 GHz to 6.5 GHz. The ultra-wideband chip NXP NCJ29D5 is used to measure CIR information. The embedded Linux device reads the CIR information from the NXP NCJ29D5 device through the CAN bus driver. Each time, a complex I and Q signal with a length of 100 is read, and the sampling frequency is 20Hz. The interval between the 100 fast time signals is 1ns, corresponding to a distance of 0.3m.

[0049] In step S12, in order to increase the resolution of the signal and align the signal, further upsampling is required. That is, the complex signal with CIR information in step S11 is converted from the time domain to the frequency domain using fast Fourier transform. Zeros are added to the left and right edges of the fast Fourier transform output. Then, the inverse fast Fourier transform is performed to return to the time domain, completing the upsampling and interpolation of the complex signal.

[0050] Specifically, upsampling is achieved by inserting zeros in the frequency domain. First, the signal with a time domain length of 100 is converted to the frequency domain using an FFT (Fourier Transform), resulting in a spectrum of length 100. 2150 zeros are then inserted around the edges of this spectrum, expanding the corresponding period in the frequency domain by a factor of 64, which is equivalent to reducing the sampling interval in the time domain to 1 / 64 of its original value. Then, an inverse IFFT is used to convert the signal from the frequency domain back to the time domain, achieving a 64-fold upsampling of the signal. The signals before and after upsampling are shown below. Figure 2a and Figure 2b As shown, upsampling can improve the accuracy of information alignment in the data;

[0051] Step S13: Align the sampling point with the largest amplitude to the x-axis of the complex plane to achieve overall data alignment;

[0052] Specifically, the upsampled signal is aligned by finding the point with the largest single-sample amplitude, calculating the phase at that point, and then multiplying the entire image by... Rotate the phase clockwise to align the point of maximum amplitude with the x-axis. The states of the two sets of data before and after alignment are as follows: Figure 3 As shown, after alignment, the first 64 points and the last 448 points with the highest values ​​were selected as features. The data selection area was determined based on the signal range affected by the kicking action.

[0053] The advantage of aligning in the complex plane is that, compared to other methods of processing CIR that only process the amplitude, this method makes better use of phase information and avoids errors caused by phase jumps at specific positions when using the phase directly. In the subsequent calculation of cumulative average and cumulative variance, this invention uses complex points in the complex plane as the processing unit.

[0054] Step S2: The CIR signal, which is the last overall data collected in step S1, is filtered by a high-pass filter and a Hanning window.

[0055] In one embodiment, step S2 specifically includes the following:

[0056] Step S21: First, data 100 is obtained. Data 100 is a complex CIR data after a single upsampling of 1*512. The data is then superimposed row by row along a slow time interval for 6.4 seconds to obtain data 110. Data 110 is in the format of 128*512. The CIR signal aligned with the last acquired data in step S1 (i.e., data 110) is then subjected to a high-pass filter to remove low-frequency components. This effectively removes signals generated by stationary objects. The system function of the high-pass filter is... The zeros are located in the complex plane (1, 0), and the poles are located in the complex plane (k, 0). The value of k ranges from [0, 1]. When k is close to 0, the low-frequency suppression effect is strong, and only signals corresponding to fast-moving objects are preserved; when k is close to 1, the low-frequency suppression effect is weak, and slower-moving objects are also preserved. The chosen value for k is 0.9, because the speed corresponding to breathing is not fast, so only signals generated by completely stationary objects need to be filtered out. After filtering, the generated data is 120.

[0057] Step S22: The signal generated in step S21 after high-pass filtering is data in 128*512 format (i.e., data 120). This data is then processed using a Hanning window. The mathematical formula for the Hanning window is: The value of N is 128, consistent with the slow time scale, and the value of n ranges from [0, N-1]. The Hanning window reduces the sidelobe level in the frequency domain representation of the signal, making it easier to distinguish close frequencies. The drawback is that it also reduces frequency resolution. Then, the generated Hanning window is applied column-wise to data 120 through element-wise multiplication; that is, the 128 signals corresponding to each distance are multiplied element-wise with the Hanning window, preparing the signal for Fourier transform processing. This windowing process helps control spectral leakage, i.e., the spread of signal power from one frequency bin to adjacent frequency bins. This is a commonly used process when performing spectral analysis on a signal.

[0058] like Figure 4As shown, in step S3, when collecting data, sampling points are randomly selected, and the CIR signal filtered in step S2 is subjected to fast Fourier transform with the sampling points as the center to obtain the velocity-distance map.

[0059] In one embodiment, step S3 specifically includes the following:

[0060] During data acquisition, sampling points are randomly selected. A Fast Fourier Transform (FFT) is then performed on the filtered CIR signal (i.e., data 130) from step S2, centered on the sampling point, to obtain data 140. This FFT converts the complex sequence into the magnitude of the corresponding velocity component; here, the velocity resolution... and maximum speed From the formulas respectively

[0061]

[0062] and

[0063] Confirmed, among which , The carrier frequency is 6 GHz, c is the speed of light, and fs is the sampling frequency of 20 Hz.

[0064] This means that each cell on the vertical axis represents 128 vertical coordinate points correspond to The formula for converting data obtained from a Fourier transform to a logarithmic coordinate system is as follows: , where x is the data obtained from the Fourier transform;

[0065] Superimposing S on the slow time dimension yields a 128*512 time-distance map as a feature. 128 represents the time dimension, corresponding to 6.4 seconds, and 512 represents the distance dimension, corresponding to 2.4 meters. After upsampling by 64 times, there are 6400 fast-time sampling points. Only the 64 points before the first peak and the subsequent 448 points are selected. The 448 points correspond to 2.1 meters after the first peak, which is an ellipsoid where the sum of the distances to the two devices is less than D+2.1 meters, where D is the distance between the two devices. Since the sampling rate is 20Hz, the 128 sampling points correspond to 6.4 seconds of time. A 30.4-second data segment is selected as the judgment criterion, and a velocity-distance map is generated every two seconds, for a total of 13 velocity-distance maps.

[0066] Step S4: Based on the velocity-distance map in step S3, construct training set, validation set and test set, train neural network, generate weight file, deploy weight file to embedded Linux platform, and obtain classification result based on collected CIR data;

[0067] In one embodiment, the specific content of step S4, "based on the velocity-distance map in step S3, constructing a training set, a validation set, and a test set," is as follows:

[0068] After confirming the generation method of the speed-distance map in step S3, the collected data includes data with people in the vehicle and data with no people in the vehicle. When collecting data, the training set, validation set and test set are recorded respectively. The number of the three sets follows or approximately follows the ratio of 10:2:2. The three sets should be collected by different people. In one embodiment, after the data is collected, when generating the dataset, sampling points are randomly selected in the dataset. A total of 608 data points before and after each sampling point (30.4 seconds) are used as a single training sample. For each training sample, a set of 128*512 speed-distance maps is generated centered on each training sample every 2 seconds, generating a total of 13 speed-distance maps.

[0069] In one embodiment, to enhance the robustness of the model, the input data is normalized. For all data M, their mean and variance std are calculated. The normalization formula is as follows:

[0070] .

[0071] like Figure 5 As shown, the neural network, in the order of data processing, includes the data input part 1, the CNN network (including CNN network 2, CNN network 3 and CNN network 4), the connection part between the CNN network and the LSTM network 5, the LSTM network 6, and the result data network 7. Among them, the data processing results of the CNN network include the result after the first convolution under the action of CNN network 2, the result after the first pooling under the action of CNN network 3, and the result after multiple convolutions and pooling under the action of CNN network 4.

[0072] In one embodiment, the data input part 1 of the neural network represents a single 128*512 CNN input data, with the vertical axis representing slow time (size 128) and the horizontal axis representing fast time (size 512). Each training sample has 13 such input matrices. These 13 128*512 input matrices generate 13 inputs to the LSTM network, corresponding to a stride of 13 for the LSTM network. Each time, it receives data of one stride, and after passing through cascaded LSTM cells, it obtains the state at step i. When the model progresses to step 13, the output hidden state becomes the input to the result data network. The result data network is the final fully connected output layer, outputting a batch_size*2 matrix. The two columns represent the scores for people in the car and the scores for people not in the car, respectively. One-hot labels are used, and cross-entropy is used as the loss function.

[0073] Specifically, CNN networks 2, 3, and 4 are CNN models that receive 128*512 input data. The quotation marks between CNN networks 3 and 4 indicate that the number of hidden layers is variable. During training, this invention adjusts the number of layers, convolutional block size, stride, and number of output channels of the CNN models to achieve better results on the validation set. Finally, at the junction 5 of the LSTM network, a three-dimensional matrix of channel*H*W is output. This matrix extracts the features in the image space of a single 6.4-second image. For example, at the data input part 1, from bottom to top, the data represents no one in the car, someone in the car, and no one in the car.

[0074] The input data of the connection part 5 between the CNN network and the LSTM network is the output result of the data in the CNN network 4 after passing through the flattening layer. It is a vector of length channel*H*W. The output data of the connection part 5 between the CNN network and the LSTM network is the input of the LSTM network 6 in a single step.

[0075] The structural units of LSTM network 6 are as follows Figure 6 As shown, LSTM network 6, or Long Short-Term Memory network, is a special type of recurrent neural network (RNN) structure that performs well in processing sequential data. It is designed to address the vanishing and exploding gradient problems inherent in traditional RNNs with long-term dependencies. LSTM network 6 includes an input gate, a forget gate, a cell state, and an output gate;

[0076] The input gate determines which new information should be stored in the cell state. It consists of a sigmoid layer and a tanh layer. The output of the sigmoid layer determines which values ​​will be updated, and the tanh layer creates a new candidate value vector that may be added to the state. In an LSTM network, the formula for the input gate is:

[0077]

[0078]

[0079] in, , , , Similarly, this represents the corresponding weights and biases.

[0080] The forget gate determines which information needs to be forgotten or discarded; specifically, the forget gate learns to discard information from the cell state. In LSTM networks, the formula for the forget gate is:

[0081]

[0082] In this formula, σ represents the sigmoid activation function. and These are weights and biases. This represents the hidden state of the previous time step. This represents the input at the current time step.

[0083] Cellular state can be viewed as part of "memory," capable of retaining information over a long period and influenced by the forgetting gate and the input gate; the forgetting gate determines what information is discarded, and the input gate determines what new information is updated. The formula for the updated cell state is:

[0084]

[0085] The output gate determines the output value. The LSTM network first runs a sigmoid layer to determine which part of the cell state will be output. Then, the cell state is passed through a tanh layer (pushing the value to -1 to 1) and multiplied by the output of the sigmoid gate. Finally, only the part that determines the output is output. In LSTM network 6, the formula for the output gate is:

[0086]

[0087]

[0088] in, , Similarly, this represents the corresponding weights and biases.

[0089] In one embodiment, Figure 6 Using a single LSTM cell, this invention modifies the number of LSTM cells during model parameter tuning, cascading multiple LSTM cells to improve model performance. Simultaneously, the LSTM... as well as The length is also an adjustable parameter.

[0090] Specifically, the individual components of the resulting data network 7 are cascaded LSTM cells, and the 33 128*512 input matrices in CNN network 2 generate 33 LSTM inputs. The corresponding step size for the LSTM is 33. The LSTM receives data of one step size at a time, and the state of the i-th step is obtained after passing through cascaded LSTM cells. When the model reaches step 33, the output is... This is the input for the 8th layer.

[0091] Specifically, the final fully connected output layer, Network 7, outputs a batch_size*2 matrix. The two columns represent the scores for when someone is in the car and the scores for when no one is in the car, respectively. One-hot encoding is used, with the corresponding option in the label set to 1 and the rest to 0. Cross-entropy is chosen as the loss function, especially important for handling classification problems. For binary classification problems, the formula for cross-entropy can be defined as:

[0092]

[0093] Here, y is the true label, y' is the predicted label, and the cross-entropy loss function measures the difference between the predicted probability distribution and the true probability distribution. The cross-entropy loss function has a value of 0 when the prediction is completely correct and an value of infinity when the prediction is completely wrong. Therefore, the goal when training a model is usually to minimize the cross-entropy loss function. Because cross-entropy is used, a sigmoid activation function needs to be added after the fully connected layer; however, this function is already integrated into the cross-entropy loss function in Python.

[0094] Specifically, during training, the parameter update optimizer used in this invention is ADAM. It combines two extensions of gradient descent: RMSProp (Root Mean Square Propagation) and Momentum. The ADAM algorithm achieves adaptive adjustment of the learning rate by calculating the first-order and second-order moment estimates of the gradient. Specifically: First-order moment estimate: also known as the exponential moving average of the gradient, corresponding to the Momentum method, which can increase the momentum of the optimization process. Second-order moment estimate: also known as the exponential moving average of the squared gradient, corresponding to the RMSProp method, which can adjust the learning rate.

[0095] Preferably, in terms of model deployment, after the model is trained, a `best_checkpoint.pth` file is generated. The RK3399 embedded Linux system running Ubuntu 18.04 is used as the model deployment platform. A PyTorch runtime environment is configured in Ubuntu, and the model parameters are loaded from the generated weight file `best_checkpoint.pth`. The model runs on the local port 20019, using Python's Flask library to read, parse, and return HTTP information and judgment results. The embedded Linux platform reads CIR data via a CAN driver, preprocesses the data, sends requests to the local port 20019 via Python's `requests` library, and receives the request results.

[0096] This invention utilizes information from UWB transmitting and receiving devices to filter and employ artificial intelligence algorithms to determine environmental changes and estimate human respiration, thereby achieving liveness detection. It leverages artificial intelligence to identify valid data within relevant signals, achieving high accuracy and effectively suppressing false alarms without increasing equipment costs. Specifically, the UWB digital key achieves decimeter-level positioning via an active positioning key or UWB chip in a mobile phone, enabling automatic unlocking upon sensing the approach of the vehicle owner. To achieve active positioning, UWB base stations need to be deployed on the front and rear bumpers and the roof of the vehicle. This invention uses the channel sensing function of these base stations during communication to achieve liveness detection inside the vehicle.

[0097] Corresponding to the aforementioned embodiments of the liveness detection method based on CIR signals of UWB digital keys, the present invention also provides embodiments of a liveness detection device based on CIR signals of UWB digital keys.

[0098] See Figure 7 The present invention provides a liveness detection device based on CIR signal of UWB digital key, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement the liveness detection method based on CIR signal of UWB digital key in the above embodiment.

[0099] The liveness detection device based on the CIR signal of a UWB digital key, as described in this invention, can be applied to any device with data processing capabilities, such as a computer. The device can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 7 The diagram shown is a hardware structure diagram of any device with data processing capabilities, including the CIR signal liveness detection device based on a UWB digital key according to the present invention. (Except for...) Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0100] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0101] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0102] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the liveness detection method based on a CIR signal of a UWB digital key as described in the above embodiments.

[0103] The computer-readable storage medium can be an internal storage unit of any data-processing device in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data-processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., mounted on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store computer programs and other programs and data required by any data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0104] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0105] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0106] It should be understood that although the terms first, second, third, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of one or more embodiments of this specification, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "in response to a determination," or "when," or "in the event of a determination."

[0107] The above are merely preferred embodiments of one or more embodiments of this specification and are not intended to limit the scope of one or more embodiments of this specification. Any modifications made within the spirit and principles of one or more embodiments of this specification are permitted.

Claims

1. A liveness detection method based on CIR signals from a UWB digital key, characterized in that, Includes the following steps: Step S1: Measure the CIR signal over a time period, upsample and interpolate the complex signal using Fourier transform and its inverse transform, and align the overall CIR signal data by aligning the phase of the point with the maximum amplitude. Step S2: The CIR signal aligned with the last overall data acquired in step S1 is filtered using a high-pass filter and a Hanning window. Step S3: When collecting data, randomly select sampling points, and perform a fast Fourier transform on the filtered CIR signal in step S2 with the sampling points as the center to obtain the velocity-distance map; Step S4: Based on the velocity-distance map in Step S3, a training set, a validation set, and a test set are constructed. The neural network is trained, a weight file is generated, and the weight file is deployed to an embedded Linux platform. The classification result is obtained based on the collected CIR data. The neural network, in the order of data processing, includes a data input part, a CNN network, a connection between the CNN network and the LSTM network, an LSTM network, and a result data network. The data processing result of the CNN network includes the result after the first convolution, the result after the first pooling, and the results after multiple convolutions and pooling. The connection between the CNN network and the LSTM network outputs a channel*H*W three-dimensional matrix. This channel*H*W three-dimensional matrix extracts the features in the spatial representation of a single target image. The LSTM network comprises an input gate, a forget gate, a cell state, and an output gate. The input gate determines which new information needs to be stored in the cell state. It includes a sigmoid layer and a tanh layer. The output of the sigmoid layer determines which values ​​will be updated, and the tanh layer creates a new candidate value vector that may be added to the state. The forget gate determines which information needs to be forgotten or discarded. The cell state retains information over a long period. The output gate determines the output value. The LSTM network first runs a sigmoid layer to determine which part of the cell state will be output. Then, the cell state is passed through the tanh layer and multiplied by the output of the sigmoid gate. Finally, only the part that determines the output is output.

2. The liveness detection method based on CIR signal of UWB digital key according to claim 1, characterized in that, Step S1 specifically includes the following: Step S11: Two NXP NCJ29D5 devices are used, one as a transmitter and the other as a receiver. The frequency range of these devices is 6.0 GHz to 6.5 GHz. The CIR information is measured using the ultra-wideband chip NXP NCJ29D5. The CIR information in the NCJ29D5 device is read through the CAN bus. Each reading is a complex I and Q signal with a length of 100, and the sampling frequency is 20Hz. The interval between the 100 fast time signals is 1ns, corresponding to a distance of 0.3m. Step S12: Use Fast Fourier Transform to convert the complex signal with CIR information in Step S11 from the time domain to the frequency domain, add zeros to the left and right edges of the Fast Fourier Transform output, and then perform Inverse Fast Fourier Transform to return to the time domain, completing the upsampling and interpolation of the complex signal. Step S13: Align the sampling point with the largest amplitude to the x-axis of the complex plane to achieve overall data alignment.

3. The liveness detection method based on CIR signals using a UWB digital key according to claim 2, characterized in that, Step S2 specifically includes the following: Step S21: The CIR signal aligned with the last acquired overall data in step S1 is subjected to high-pass filtering, and its system function is: The value of k ranges from [0,1]. When k is close to 0, the low-frequency suppression effect is strong, and only the signal corresponding to fast-moving objects will be preserved. When k is close to 1, the low-frequency suppression effect is weak, and slower-moving objects will also be preserved. Step S22: The signal generated in step S21 after high-pass filtering is data in 128*512 format. This data is then processed using a Hanning window. The mathematical formula for the Hanning window is: The value of N is 128, consistent with the slow time scale, and the value of n is in the range of [0, N-1]. The Hanning window reduces the sidelobe level in the frequency domain representation of the signal, making it easier to distinguish close frequencies. The generated Hanning window is applied to a class of signals by column through element-wise multiplication, preparing for Fourier transform processing of CIR signals.

4. The liveness detection method based on CIR signal of UWB digital key according to claim 1, characterized in that, The specific details of step S3 are as follows: During data acquisition, sampling points are randomly selected, and a Fast Fourier Transform (FFT) is performed on the filtered CIR signal from step S2, centered on these sampling points. This FFT converts the complex sequence into the magnitude of the corresponding velocity component; here, the velocity resolution... and maximum speed From the formulas respectively and Confirmed, among which , The carrier frequency is 6 GHz, c is the speed of light, and fs is the sampling frequency of 20 Hz. The formula for converting data obtained from a Fourier transform to a logarithmic coordinate system is as follows: , where x is the data obtained from the Fourier transform; Superimposing S on the slow time dimension yields a 128*512 time-distance map as a feature. 128 represents the time dimension, corresponding to 6.4 seconds, and 512 represents the distance dimension, corresponding to 2.4 meters. After upsampling by 64 times, there are 6400 fast-time sampling points. Only the 64 points before the first peak and the subsequent 448 points are selected. The 448 points correspond to 2.1 meters after the first peak, which is an ellipsoid where the sum of the distances to the two devices is less than D+2.1 meters, where D is the distance between the two devices. Since the sampling rate is 20Hz, the 128 sampling points correspond to 6.4 seconds of time. A 30.4-second data segment is selected as the judgment criterion, and a velocity-distance map is generated every two seconds, for a total of 13 velocity-distance maps.

5. The liveness detection method based on CIR signal of UWB digital key according to claim 1, characterized in that, The specific content of "constructing training, validation, and test sets based on the velocity-distance map in step S3" in step S4 is as follows: After confirming the generation method of the speed-distance map in step S3, the collected data includes data with people inside the vehicle and data with no one inside the vehicle. During data collection, training set, validation set, and test set are recorded separately, with the number of data in each set following a 10:2:2 ratio. The input data is normalized; for all data M, their mean and variance (std) are calculated. The normalization formula is as follows: 。 6. The liveness detection method based on CIR signal of UWB digital key according to claim 1, characterized in that, The data input part of the neural network represents a single 128*512 CNN input data. The vertical axis represents slow time with a size of 128, and the horizontal axis represents fast time with a size of 512. Each training sample has 13 such input matrices. The 13 128*512 input matrices generate 13 inputs to the LSTM network, corresponding to a stride of 13 for the LSTM network. Each time, it receives data of one stride, and after passing through cascaded LSTM cells, it obtains the state of the i-th step. When the model progresses to the 13th step, the output hidden state is the input of the result data network. The result data network is the final fully connected output layer, which outputs a batch_size*2 matrix. The two columns are the scores for people in the car and the scores for people not in the car, respectively. One-hot labels are used, and cross-entropy is used as the loss function.

7. A liveness detection device based on a CIR signal from a UWB digital key, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it is used to implement the liveness detection method based on the CIR signal of the UWB digital key as described in any one of claims 1-6.

Images