A heart rate determination method and system based on millimeter wave radar and neural network

Abstract

The present invention discloses a heart rate determination method and system based on millimeter-wave radar and neural networks, which relates to the field of heart rate detection technology. The method comprises: synchronously collecting raw radar echo data and raw electrocardiogram (ECG) signals; using an initial sliding window to determine a target ECG signal based on the raw ECG signals to generate multiple initial instantaneous heart rates; performing data removal according to high-frequency body movement time to output the target radar echo data and the target instantaneous heart rate; constructing an initial radar complex data matrix corresponding to each target instantaneous heart rate based on the target radar echo data using the initial sliding window; performing data equalization based on the initial radar complex data matrices belonging to multiple preset heart rate intervals; determining a training radar complex data matrix and a true instantaneous heart rate; determining a trained heart rate prediction network based on the training radar complex data matrix and the true instantaneous heart rate; and predicting the measured instantaneous heart rate of the measured radar echo data. The above scheme is beneficial for improving the accuracy of heart rate determination.

Description

Technical Field

[0001] The present invention relates to the field of heart rate detection technology, and in particular to a heart rate determination method and system based on millimeter wave radar and neural network. Background Art

[0002] Heart rate is an important component of human vital signal detection. Traditional detection methods mainly rely on wearable sensors or adhesive electrodes that directly contact the human body to achieve detection. In recent years, with the development of millimeter-wave radar, non-contact radar has unique advantages over other detection methods and has broad application prospects. Therefore, non-contact human vital signal detection has gradually become a research hotspot in some emerging application fields.

[0003] In the existing heart rate determination process, the human body's IQ data is first collected through millimeter-wave radar. The chest displacement signal is extracted from the IQ data and the heartbeat signal is separated using methods such as filtering, VMD, and DWT. The heart rate is then calculated based on the heartbeat signal. However, since the human heartbeat movement is very weak and there is interference from human movements such as breathing, there is a problem of insufficient feature extraction from the radar data, resulting in low heart rate determination accuracy. Summary of the Invention

[0004] The present invention provides a heart rate determination method and system based on millimeter wave radar and neural network, which solves the technical problem that existing heart rate determination schemes are prone to insufficient radar data feature extraction, resulting in low heart rate determination accuracy.

[0005] A first aspect of the present invention provides a heart rate determination method based on millimeter wave radar and a neural network, comprising:

[0006] Acquire synchronously collected original radar echo data and original ECG signals;

[0007] Preprocessing the original ECG signal to generate a target ECG signal, and performing R-peak positioning on the target ECG signal using an initial sliding window to determine multiple initial instantaneous heart rates;

[0008] According to the high-frequency body movement time determined by the amplitude of the original radar echo data, the original radar echo data and each of the initial instantaneous heart rates are subjected to data elimination, and target radar echo data and target instantaneous heart rate are output;

[0009] intercepting the target radar echo data through the initial sliding window to construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target;

[0010] If data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, data equalization is performed based on the target radar echo data and the initial sliding window to determine a training radar complex data matrix and a corresponding true instantaneous heart rate;

[0011] Determining a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and determining a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training;

[0012] The measured radar echo data is obtained, and the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data is input into the trained heart rate prediction network to output the measured instantaneous heart rate.

[0013] Optionally, if data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, performing data equalization based on the target radar echo data and the initial sliding window to determine the training radar complex data matrix and the corresponding true instantaneous heart rate includes:

[0014] If data imbalance is determined based on the initial radar complex data matrix to which the plurality of preset heart rate intervals belong, a marginal missing heart rate interval is determined;

[0015] When the edge-missing heart rate interval is a low heart rate interval, data windowing is performed on the target radar echo data through a low heart rate sliding window and linear interpolation is performed to determine a plurality of low heart rate radar complex data matrices;

[0016] Determining multiple initial low instantaneous heart rates of the target electrocardiogram signal based on the low heart rate sliding window, and performing multiplication operations on each of the initial low instantaneous heart rates using a window length ratio of the low heart rate sliding window to the initial sliding window to generate target low instantaneous heart rates corresponding to each of the low heart rate radar complex data matrices;

[0017] When the edge-missing heart rate interval is a high heart rate interval, slicing and down-sampling the target radar echo data using a high heart rate sliding window to determine a plurality of high heart rate radar complex data matrices;

[0018] constructing a plurality of initial high instantaneous heart rates of the target ECG signal according to the high heart rate sliding window, and multiplying each of the initial high instantaneous heart rates by a window length ratio of the high heart rate sliding window to the initial sliding window, respectively, to obtain a target high instantaneous heart rate associated with each of the high heart rate radar complex data matrices;

[0019] If data balance is determined according to each of the preset heart rate intervals based on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix, a training radar complex data matrix is ​​formed through sparse stratified sampling, and the corresponding true instantaneous heart rate is determined.

[0020] Optionally, determining a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and determining a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training includes:

[0021] After performing first-order filtering on the training radar complex data matrix, static clutter is filtered out using a moving target indication algorithm to determine a filtered radar complex data matrix;

[0022] calculating phase changes in the filtered radar complex data matrix and converting each phase change into a displacement to construct a chest displacement matrix;

[0023] Using an IIR bandpass filter to filter the chest cavity displacement matrix to generate a heartbeat feature matrix, and performing matrix normalization on the heartbeat feature matrix to determine a spatiotemporal map of the heartbeat feature matrix;

[0024] Performing image enhancement on the heartbeat feature matrix spatiotemporal map to generate a training enhanced heartbeat feature matrix spatiotemporal map;

[0025] Inputting the training enhanced heartbeat feature matrix spatiotemporal map into the heart rate prediction network to be trained, and outputting the predicted instantaneous heart rate;

[0026] The heart rate prediction network to be trained is iteratively optimized based on the loss function value calculated based on the predicted instantaneous heart rate and the true instantaneous heart rate to determine a trained heart rate prediction network.

[0027] Optionally, obtaining measured radar echo data, inputting a measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into a trained heart rate prediction network, and outputting the measured instantaneous heart rate includes:

[0028] When measured radar echo data is collected, a measured sliding window associated with the measured radar echo data is used to perform data segmentation to construct a measured radar complex data matrix;

[0029] Determine a spatiotemporal map of a measured enhanced heartbeat feature matrix based on the measured radar complex data matrix;

[0030] The heart rate is predicted by the trained heart rate prediction network on the measured enhanced heartbeat feature matrix spatiotemporal map, and the target instantaneous heart rate is output.

[0031] Optionally, the processing of the heart rate prediction network includes:

[0032] Perform channel cutting on the input enhanced heartbeat feature matrix spatiotemporal map to determine multi-channel input features;

[0033] Perform preliminary channel interaction according to the multi-channel input features, and output channel interaction features;

[0034] Performing convolution extraction on the channel interaction features to generate low-level features;

[0035] Performing continuous residual learning and residual attention processing based on the low-level features to construct residual enhancement features;

[0036] Adopting the residual enhancement feature to perform adaptive feature pooling to determine the global feature;

[0037] The global features are sequentially subjected to linear transformation, ReLU nonlinear mapping, and Dropout regularization, and then linearly transformed to output the instantaneous heart rate.

[0038] Optionally, the adopting the residual enhancement feature to perform adaptive feature pooling to determine the global feature includes:

[0039] Performing global average pooling using the residual enhancement feature to determine an average pooling feature;

[0040] Performing global maximum pooling according to the residual enhancement feature and outputting the maximum pooling feature;

[0041] Concatenating the average pooling feature and the maximum pooling feature to construct a pooling concatenation feature;

[0042] Performing a softmax operation after linear transformation based on the pooled splicing features to generate an average pooling weight and a maximum pooling weight;

[0043] The average pooling feature and the average pooling weight, the maximum pooling feature and the maximum pooling weight are weighted and then concatenated to output a global feature.

[0044] A second aspect of the present invention provides a heart rate determination system based on millimeter wave radar and a neural network, comprising:

[0045] A data acquisition module is used to obtain synchronously collected original radar echo data and original electrocardiogram signals;

[0046] an ECG signal processing module, configured to pre-process the original ECG signal to generate a target ECG signal, and perform R-peak positioning on the target ECG signal using an initial sliding window to determine a plurality of initial instantaneous heart rates;

[0047] a data removal module, configured to remove the raw radar echo data and each of the initial instantaneous heart rates according to the high-frequency body movement time determined by the amplitude of the raw radar echo data, and output target radar echo data and target instantaneous heart rate;

[0048] a data matrix construction module, configured to intercept the target radar echo data through the initial sliding window and construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target;

[0049] a training data determination module configured to, if data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, perform data equalization based on the target radar echo data and the initial sliding window to determine a training radar complex data matrix and a corresponding true instantaneous heart rate;

[0050] A model training module is used to determine a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and to determine a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training;

[0051] The heart rate measurement module is used to obtain the measured radar echo data, input the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into the trained heart rate prediction network, and output the measured instantaneous heart rate.

[0052] A third aspect of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the heart rate determination method based on millimeter wave radar and neural network as described in any one of the above items.

[0053] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the heart rate determination method based on millimeter-wave radar and neural network as described in any one of the above items.

[0054] A fifth aspect of the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the heart rate determination method based on millimeter-wave radar and neural network as described in any one of the above items.

[0055] It can be seen from the above technical solutions that the present invention has the following advantages:

[0056] The above solution of the present invention provides a heart rate determination method based on millimeter wave radar and neural network, comprising: obtaining synchronously collected original radar echo data and original electrocardiogram signals; preprocessing the original electrocardiogram signals to generate target electrocardiogram signals, using an initial sliding window to perform R peak positioning on the target electrocardiogram signals to determine multiple initial instantaneous heart rates; according to the high-frequency body movement time determined by the amplitude of the original radar echo data, data is removed from the original radar echo data and each initial instantaneous heart rate, and the target radar echo data and the target instantaneous heart rate are output; data is intercepted on the target radar echo data through the initial sliding window to construct the initial radar echo corresponding to each target instantaneous heart rate. A complex data matrix is ​​obtained; if data imbalance is determined based on the initial radar complex data matrix belonging to multiple preset heart rate intervals, data equalization is performed based on the target radar echo data and the initial sliding window to determine the training radar complex data matrix and the corresponding true instantaneous heart rate; a training enhanced heartbeat feature matrix spatiotemporal map is determined based on the training radar complex data matrix, and a trained heart rate prediction network is determined using the training enhanced heartbeat feature matrix spatiotemporal map and the true instantaneous heart rate training; measured radar echo data is obtained, and the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data is input into the trained heart rate prediction network to output the measured instantaneous heart rate. In the above scheme, data preprocessing and data equalization are performed based on radar echo data and electrocardiogram signals, making full use of data from various spatial distances of the radar to construct training data, and the heart rate prediction network is trained using the training data. The heart rate prediction network then fully extracts the heartbeat features contained in the radar echo data collected by the millimeter wave radar, which is beneficial to improving the accuracy of heart rate determination. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 A flowchart of a method for determining heart rate based on millimeter-wave radar and neural network provided by an embodiment of the present invention;

[0059] Figure 2 A schematic diagram of a data collection process according to an embodiment of the present invention;

[0060] Figure 3 A comparison chart of various signals including body motion data provided by an embodiment of the present invention;

[0061] Figure 4 A schematic diagram of a spatiotemporal map of a chest cavity displacement matrix provided by an embodiment of the present invention;

[0062] Figure 5 A schematic diagram of a spatiotemporal map of a heartbeat feature matrix provided by an embodiment of the present invention;

[0063] Figure 6 An image enhancement effect diagram of the spatiotemporal atlas of the heartbeat feature matrix provided by an embodiment of the present invention;

[0064] Figure 7 A graph showing the change in loss and accuracy during the training process of a heart rate prediction network provided by an embodiment of the present invention;

[0065] Figure 8 The network framework of the heart rate prediction network provided by the embodiment of the present invention Figure 1 ;

[0066] Figure 9 The network framework of the heart rate prediction network provided by the embodiment of the present invention Figure 2 ;

[0067] Figure 10 A schematic diagram of the structure of the adaptive pooling layer provided in an embodiment of the present invention;

[0068] Figure 11 This is a structural block diagram of a heart rate determination system based on millimeter wave radar and neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] Embodiments of the present invention provide a heart rate determination method and system based on millimeter-wave radar and neural network, which are used to solve the technical problem that existing heart rate determination solutions are prone to insufficient radar data feature extraction, resulting in low heart rate determination accuracy.

[0070] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below 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 making creative work are within the scope of protection of the present invention.

[0071] See also Figure 1 , Figure 1 A flowchart of the steps of a heart rate determination method based on millimeter wave radar and neural network provided in an embodiment of the present invention.

[0072] The present invention provides a heart rate determination method based on millimeter wave radar and neural network, comprising:

[0073] Step 101: Acquire synchronously collected original radar echo data and original electrocardiogram signals.

[0074] It should be noted that if Figure 2 As shown, this embodiment uses a millimeter-wave radar device to collect data from a resting subject lying quietly in bed to obtain raw radar echo data (radar IQ data). At the same time, an ECG acquisition device with reliable accuracy is used to collect the resting subject's raw electrocardiogram signal (ECG signal). The main consideration is that the raw radar echo data is mixed with signals such as the resting human heartbeat and breathing, and the synchronously collected raw ECG signal can serve as a benchmark to provide an accurate heartbeat signal reference.

[0075] Step 102: Preprocess the original ECG signal to generate a target ECG signal, and use an initial sliding window to perform R-peak positioning on the target ECG signal to determine multiple initial instantaneous heart rates.

[0076] It should be noted that the PT method is used to calculate the corresponding initial instantaneous heart rate for the original ECG signal. In the specific implementation, the original ECG signal is first preprocessed by low-pass filtering, band-pass filtering, differential operation and square operation in sequence to determine the target ECG signal. Then, a preset initial sliding window is used to perform sliding window detection on the target ECG signal and perform R peak positioning. The instantaneous heart rate is calculated according to the R peak position in each initial sliding window and recorded as the initial instantaneous heart rate. The initial sliding window can be set to a window length of 12s and a step length of 1s.

[0077] Step 103 : According to the high-frequency body movement time determined by the amplitude of the original radar echo data, the original radar echo data and each initial instantaneous heart rate are removed, and the target radar echo data and the target instantaneous heart rate are output.

[0078] It should be noted that, since the original radar echo data may contain poorly collected data, it is necessary to perform pre-screening. In this embodiment, high-frequency body motion factors caused by violent body motion and environmental influences are used as the screening basis, such as Figure 3 As shown, the degree of subject's body movement is judged according to the amplitude of the original radar echo data, and an amplitude threshold is set to determine the high-frequency body movement time that causes data abnormality. The original radar echo data is then removed according to the high-frequency body movement time. Simultaneously, the corresponding initial instantaneous heart rate is also removed, thereby determining the target radar echo data and the target instantaneous heart rate.

[0079] It is understood that radar echo data (IQ data) includes in-phase component (I) and quadrature component (Q). In IQ data, amplitude is usually used to represent signal strength, which can be measured by To calculate the amplitude.

[0080] Step 104: intercept the target radar echo data through the initial sliding window to construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target.

[0081] It should be noted that in order to determine the instantaneous heart rate of each target as the target radar echo data corresponding to the heart rate label, the data is intercepted by sliding the initial sliding window on the target radar echo data, and the target radar echo data of each initial sliding window is organized into a matrix form and recorded as the initial radar complex data matrix.

[0082] Step 105: If data imbalance is determined based on the initial radar complex data matrices belonging to the multiple preset heart rate intervals, data equalization is performed based on the target radar echo data and the initial sliding window to determine the training radar complex data matrix and the corresponding true instantaneous heart rate.

[0083] Step 105 includes the following sub-steps:

[0084] S11, if data imbalance is determined based on the initial radar complex data matrix to which the plurality of preset heart rate intervals belong, determining an edge missing heart rate interval;

[0085] S12. When the edge-missing heart rate interval is a low heart rate interval, windowing the target radar echo data using a low heart rate sliding window and performing linear interpolation to determine multiple low heart rate radar complex data matrices;

[0086] S13, determining multiple initial low instantaneous heart rates of the target ECG signal based on the low heart rate sliding window, multiplying each of the initial low instantaneous heart rates by a window length ratio of the low heart rate sliding window to the initial sliding window, to generate target low instantaneous heart rates corresponding to each low heart rate radar complex data matrix;

[0087] S14. When the edge-missing heart rate interval is a high heart rate interval, slicing and down-sampling the target radar echo data using a high heart rate sliding window to determine multiple high heart rate radar complex data matrices;

[0088] S15, constructing multiple initial high instantaneous heart rates of the target ECG signal based on the high heart rate sliding window, multiplying each initial high instantaneous heart rate by the window length ratio of the high heart rate sliding window to the initial sliding window, and obtaining the target high instantaneous heart rate associated with each high heart rate radar complex data matrix;

[0089] S16. If data balance is determined according to each preset heart rate interval based on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix, a training radar complex data matrix is ​​formed through sparse stratified sampling, and the corresponding true instantaneous heart rate is determined.

[0090] It should be noted that each initial radar complex data matrix corresponds to a target instantaneous heart rate as a heart rate label. According to the heart rate label, the initial radar complex data matrix belonging to each interval is counted according to multiple preset heart rate (HR) intervals to determine whether there is data imbalance;

[0091] If the data is balanced, sparse stratified sampling is performed based on each initial radar complex data matrix to form a training radar complex data matrix, and the target instantaneous heart rate corresponding to the training radar complex data matrix is ​​determined as the true instantaneous heart rate to make the sample distribution uniform;

[0092] If the data is unbalanced, observing the number of radar complex data matrices in each preset heart rate interval can reveal intervals with missing data. Since the radar complex data matrices in the medium heart rate interval must be relatively large in the preset heart rate interval, the main task is to determine the edge missing heart rate intervals to expand the edge data and thus complete data balancing. The edge missing heart rate intervals include the low heart rate interval and the high heart rate interval relative to the medium heart rate interval;

[0093] When the edge missing heart rate interval is a low heart rate interval, the target radar echo data is windowed using a preset low heart rate sliding window, wherein the window length of the low heart rate sliding window is smaller than the window length of the initial sliding window, and the target radar echo data of each low heart rate sliding window is stretched and enlarged to the window length of the initial sliding window by linear interpolation to generate multiple low heart rate radar complex data matrices, and the target electrocardiogram signal is R-peak located using the low heart rate sliding window to calculate multiple initial low instantaneous heart rates, and the window length ratio determined by performing a ratio operation between the low heart rate sliding window and the initial sliding window is used as the stretching ratio of the target radar echo data in the low heart rate sliding window, and each initial low instantaneous heart rate is scaled according to the window length ratio, thereby obtaining multiple target low instantaneous heart rates as heart rate labels corresponding to each low heart rate radar complex data matrix;

[0094] When the edge missing heart rate interval is a high heart rate interval, the target radar echo data is sliced ​​through a preset high heart rate sliding window, wherein the window length of the high heart rate sliding window is greater than the window length of the initial sliding window, and the target radar echo data of each high heart rate sliding window is compressed and narrowed to the window length of the initial sliding window by downsampling to generate multiple high heart rate radar complex data matrices, and the target electrocardiogram signal is R-peak located through the high heart rate sliding window to calculate multiple initial high instantaneous heart rates, and the window length ratio determined by the ratio operation between the high heart rate sliding window and the initial sliding window is used as the compression ratio of the target radar echo data in the high heart rate sliding window, and each initial high instantaneous heart rate is scaled according to the window length ratio to determine multiple target high instantaneous heart rates as heart rate labels corresponding to each high heart rate radar complex data matrix;

[0095] After completing the above-mentioned data balancing processing, classification and statistics are performed again based on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix according to each preset heart rate interval until the statistical results show that the radar complex data matrices of each preset heart rate interval are basically balanced. In order to make the sample distribution more uniform, sparse stratified sampling is performed on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix to form a training radar complex data matrix, and the corresponding true instantaneous heart rate is determined.

[0096] Step 106: Determine a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and determine a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training.

[0097] Step 106 includes the following sub-steps:

[0098] After first-order filtering of the training radar complex data matrix, static clutter is filtered out using a moving target indication algorithm to determine the filtered radar complex data matrix;

[0099] Calculate the phase changes in the filtered radar complex data matrix and convert each phase change into displacement to construct the chest displacement matrix;

[0100] An IIR bandpass filter is used to filter the chest displacement matrix to generate a heartbeat feature matrix, and the heartbeat feature matrix is ​​normalized to determine the spatiotemporal map of the heartbeat feature matrix.

[0101] Perform image enhancement on the heartbeat feature matrix spatiotemporal map to generate a training enhanced heartbeat feature matrix spatiotemporal map;

[0102] Input the training enhanced heartbeat feature matrix spatiotemporal map into the heart rate prediction network to be trained, and output the predicted instantaneous heart rate;

[0103] The heart rate prediction network to be trained is iteratively optimized based on the loss function value calculated based on the predicted instantaneous heart rate and the true instantaneous heart rate to determine the trained heart rate prediction network.

[0104] It should be noted that after completing the data selection to determine the training radar complex data matrix and the true instantaneous heart rate, further processing is performed based on the training radar complex data matrix to obtain the heartbeat data. First, each training radar complex data matrix is ​​subjected to first-order filtering to obtain a smoothed radar complex data matrix. The smoothed radar complex data matrix is ​​filtered out of static clutter according to the moving target indication (MTI) algorithm to determine the filtered radar complex data matrix. Then, the phase change between each matrix element in the filtered radar complex data matrix is ​​calculated, and the phase change is converted into displacement, thereby constructing the chest displacement matrix. The spatiotemporal spectrum of the chest displacement matrix is ​​shown as follows: Figure 4 As shown, the chest displacement matrix is ​​then processed through a 0.8-2Hz IIR bandpass filter to generate a heartbeat feature matrix, and then the matrix is ​​normalized to obtain the following Figure 5 Finally, the obtained heartbeat feature matrix spatiotemporal map is subjected to image enhancement to obtain the trained enhanced heartbeat feature matrix spatiotemporal map. Image enhancement includes histogram equalization and dark enhancement. The enhancement effect is as follows: Figure 6 As shown;

[0105] After the heart rate prediction network to be trained is built, the training enhanced heartbeat feature matrix spatiotemporal map is input into the heart rate prediction network to be trained, and the corresponding predicted instantaneous heart rate is output. The preset loss function is used to calculate the loss function value based on the predicted instantaneous heart rate and the actual instantaneous heart rate. The change curves of the training loss (train_loss) and training accuracy (train_acc) during the training process are shown as follows: Figure 7 As shown, if the loss function value has not converged, a suitable optimizer is selected to iteratively optimize the heart rate prediction network to be trained with the goal of minimizing the loss function value until the loss function value converges, and the trained heart rate prediction network is determined. During the training process, the validation set can also be used to determine the validation loss (val_loss) and validation accuracy (val_acc) to evaluate the model performance and prevent overfitting. In one implementation, the loss function can use the mean squared error loss function (MSE Loss): , where is the sample size, is the real instantaneous heart rate, To predict instantaneous heart rate.

[0106] In a specific embodiment, this embodiment provides a heart rate prediction network that introduces an attention mechanism and is designed with multi-channel residuals. The network architecture is shown in FIG. Figure 8 and Figure 9 As shown in the figure, it includes the channel interaction layer (ChannelInteraction), feature initialization layer (Initial processe), residual attention layer (ResLayer), adaptive pooling layer (AdaptivePooling), and prediction output layer (Classifier); the processing process of the heart rate prediction network includes:

[0107] S21. Perform channel cutting on the input enhanced heartbeat feature matrix spatiotemporal map to determine multi-channel input features.

[0108] It should be noted that the enhanced heartbeat feature matrix spatiotemporal map x processed by the heart rate prediction network is cut into 31 channels to determine the multi-channel input features. It can be regarded as a grayscale image with a dimension of (1, 31, 300, 20), where "1" represents that it has only grayscale values, "31" corresponds to cutting into 31 channel inputs, the width "300" corresponds to the time dimension, and the height "20" corresponds to the space dimension. It can be understood that this embodiment has been verified through a large number of experiments that segmentation into 31 channels can achieve the best prediction effect while occupying relatively few computing resources.

[0109] S22. Perform preliminary channel interaction based on the multi-channel input features and output channel interaction features.

[0110] It should be noted that the channel interaction layer is used to perform preliminary channel interaction on the multi-channel input features, where the channel interaction layer includes a 3×3 convolution unit, a BatchNorm2d unit and a ReLU activation function. After the multi-channel input features are convolved by the convolution unit, the BatchNorm2d unit is used for two-dimensional batch normalization, and ReLU nonlinear mapping is performed based on the ReLU activation function, thereby completing preliminary channel interaction to enrich the channel features and outputting the channel interaction features.

[0111] S23. Perform convolution extraction on channel interaction features to generate low-level features.

[0112] It should be noted that the channel interaction features are convolutionally extracted through the feature initialization layer, which includes a convolution unit (the convolution kernel size is 7×7, the stride is 2 and the padding is 3), a BatchNorm2d unit, a ReLU activation function and a 3×3 global maximum pooling MaxPool unit (the stride is 2). After the channel interaction features are extracted using the convolution unit, the feature map size is reduced by half ( / 2), and two-dimensional batch normalization is performed through the BatchNorm2d unit. After ReLU nonlinear processing is performed through the ReLU activation function, maximum pooling is performed based on the global maximum pooling unit to obtain low-level features. At this time, the feature map size is further reduced by half, reducing the amount of computation for subsequent processing.

[0113] S24. Perform continuous residual learning and residual attention processing based on low-level features to construct residual enhancement features.

[0114] It should be noted that the low-level features are input into the cascaded multiple residual attention layers for continuous residual learning and residual attention processing. Each residual attention layer includes a residual block (ResNet Block) and a channel attention unit (Channel Attention). In the processing of each residual attention layer, after residual learning is performed on the input residual input features through each residual attention layer, the channel attention unit is used to input the channel weight, and the channel weight is multiplied by the residual input feature to obtain the corresponding residual output feature, thereby enhancing the important features and outputting the residual enhanced features.

[0115] S25. Adaptive feature pooling is performed using residual enhancement features to determine global features.

[0116] Optionally, sub-step S25 includes:

[0117] Use the residual enhancement feature to perform global average pooling and determine the average pooling feature;

[0118] Perform global maximum pooling based on the residual enhancement features and output the maximum pooling features;

[0119] Concatenate the average pooling feature and the maximum pooling feature to construct the pooling concatenation feature;

[0120] After performing linear transformation based on the pooled splicing features, a softmax operation is performed to generate average pooling weights and maximum pooling weights;

[0121] The average pooling feature and the average pooling weight, the maximum pooling feature and the maximum pooling weight are weighted respectively and then concatenated to output the global feature.

[0122] It should be noted that adaptive feature pooling is performed on the residual enhancement features based on the adaptive pooling layer. In one implementation, the structure of the adaptive pooling layer is as follows: Figure 10 As shown in the figure, it includes a global maximum pooling unit (Maxpool), a global average pooling unit (Avgpool) and a weight learning unit (weight). The weight learning unit includes a fully connected subunit and a softmax activation function. In the specific implementation, the global maximum pooling unit is used to perform global maximum pooling on the residual enhancement feature to generate a maximum pooling feature. The global average pooling unit is used to perform global average pooling on the residual enhancement feature to output the average pooling feature. The average pooling feature and the maximum pooling feature are concatenated (Concatenate) to construct a pooled splicing feature. After the pooled splicing feature is linearly transformed by the fully connected subunit, the softmax activation function is used to perform the softmax operation, and the average pooling weight and the maximum pooling weight are output. They are weighted with the average pooling feature and the maximum pooling feature respectively and combined together to generate a compact feature containing global information, namely a global feature.

[0123] S26. After linear transformation, ReLU nonlinear mapping and Dropout regularization are performed on the global features in sequence, linear transformation is performed to output the instantaneous heart rate.

[0124] It should be noted that heart rate prediction is performed based on global features through the prediction output layer, where the prediction output layer includes a fully connected unit (FC) and a ReLU activation function. After the global features are linearly transformed using the fully connected unit, ReLU nonlinear mapping is performed based on the ReLU activation function, and after Dropout regularization, the fully connected unit is linearly changed again to complete the prediction task and output the estimated instantaneous heart rate.

[0125] Step 107: Obtain measured radar echo data, input the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into the trained heart rate prediction network, and output the measured instantaneous heart rate.

[0126] Step 107 includes the following sub-steps:

[0127] When the measured radar echo data is collected, the measured sliding window associated with the measured radar echo data is used to segment the data and construct the measured radar complex data matrix;

[0128] Determine the spatiotemporal map of the measured enhanced heartbeat feature matrix based on the measured radar complex data matrix;

[0129] The trained heart rate prediction network is used to predict the heart rate of the measured enhanced heartbeat feature matrix spatiotemporal map and output the target instantaneous heart rate.

[0130] It should be noted that after determining the trained heart rate prediction network, when the measured radar echo data of the heart rate to be determined is collected, the measured radar echo data is slidingly segmented according to the preset measured sliding window to construct a measured radar complex data matrix, and according to the step of determining the training enhanced heartbeat feature matrix spatiotemporal spectrum in step 106, the measured enhanced heartbeat feature matrix spatiotemporal spectrum corresponding to the measured radar complex data matrix is ​​determined, and input into the trained heart rate prediction network for heart rate evaluation, thereby outputting the target instantaneous heart rate.

[0131] In an embodiment of the present invention, data preprocessing and data equalization are performed based on radar echo data and electrocardiogram signals, and the data of each spatial distance of the radar are fully utilized to construct training data. The heart rate prediction network is trained with the training data, and the heart rate prediction network is used to fully extract the heartbeat features contained in the radar echo data collected by the millimeter wave radar, which is beneficial to improving the accuracy of heart rate determination.

[0132] See also Figure 11 , Figure 11This is a structural block diagram of a heart rate determination system based on millimeter wave radar and neural network provided by an embodiment of the present invention.

[0133] The present invention provides a heart rate determination system based on millimeter wave radar and neural network, comprising:

[0134] The data acquisition module 1101 is used to obtain the synchronously collected original radar echo data and original ECG signal;

[0135] The ECG signal processing module 1102 is configured to pre-process the original ECG signal to generate a target ECG signal, and use an initial sliding window to perform R-peak positioning on the target ECG signal to determine multiple initial instantaneous heart rates;

[0136] A data removal module 1103 is configured to remove the original radar echo data and each initial instantaneous heart rate according to the high-frequency body motion time determined by the amplitude of the original radar echo data, and output the target radar echo data and the target instantaneous heart rate;

[0137] The data matrix construction module 1104 is used to intercept the target radar echo data through the initial sliding window and construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target;

[0138] a training data determination module 1105 for performing data equalization based on the target radar echo data and the initial sliding window to determine the training radar complex data matrix and the corresponding true instantaneous heart rate if data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals;

[0139] A model training module 1106 is configured to determine a training enhanced heartbeat feature matrix spatiotemporal map based on a training radar complex data matrix, and determine a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training;

[0140] The heart rate measurement module 1107 is used to obtain the measured radar echo data, input the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into the trained heart rate prediction network, and output the measured instantaneous heart rate.

[0141] Optionally, the training data determination module 1105 is specifically configured to:

[0142] If data imbalance is determined based on the initial radar complex data matrix to which the plurality of preset heart rate intervals belong, a marginal missing heart rate interval is determined;

[0143] When the edge missing heart rate interval is a low heart rate interval, the target radar echo data is windowed and linearly interpolated using a low heart rate sliding window to determine multiple low heart rate radar complex data matrices;

[0144] Determine multiple initial low instantaneous heart rates of the target ECG signal based on the low heart rate sliding window, use the window length ratio of the low heart rate sliding window to the initial sliding window, and perform multiplication operation on each initial low instantaneous heart rate to generate the target low instantaneous heart rate corresponding to each low heart rate radar complex data matrix;

[0145] When the edge missing heart rate interval is a high heart rate interval, a high heart rate sliding window is used to slice and downsample the target radar echo data to determine multiple high heart rate radar complex data matrices;

[0146] Constructing multiple initial high instantaneous heart rates of the target ECG signal based on the high heart rate sliding window, multiplying the window length ratio of the high heart rate sliding window to the initial sliding window by each initial high instantaneous heart rate, corresponding to the target high instantaneous heart rate associated with each high heart rate radar complex data matrix;

[0147] If data balance is determined according to each preset heart rate interval based on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix, a training radar complex data matrix is ​​formed through sparse stratified sampling, and the corresponding true instantaneous heart rate is determined.

[0148] Optionally, the model training module 1106 is specifically configured to:

[0149] After first-order filtering of the training radar complex data matrix, static clutter is filtered out using a moving target indication algorithm to determine the filtered radar complex data matrix;

[0150] Calculate the phase changes in the filtered radar complex data matrix and convert each phase change into displacement to construct the chest displacement matrix;

[0151] An IIR bandpass filter is used to filter the chest displacement matrix to generate a heartbeat feature matrix, and the heartbeat feature matrix is ​​normalized to determine the spatiotemporal map of the heartbeat feature matrix.

[0152] Perform image enhancement on the heartbeat feature matrix spatiotemporal map to generate a training enhanced heartbeat feature matrix spatiotemporal map;

[0153] Input the training enhanced heartbeat feature matrix spatiotemporal map into the heart rate prediction network to be trained, and output the predicted instantaneous heart rate;

[0154] The heart rate prediction network to be trained is iteratively optimized based on the loss function value calculated based on the predicted instantaneous heart rate and the true instantaneous heart rate to determine the trained heart rate prediction network.

[0155] Optionally, the heart rate measurement module 1107 is specifically configured to:

[0156] When the measured radar echo data is collected, the measured sliding window associated with the measured radar echo data is used to segment the data and construct the measured radar complex data matrix;

[0157] Determine the spatiotemporal map of the measured enhanced heartbeat feature matrix based on the measured radar complex data matrix;

[0158] The trained heart rate prediction network is used to predict the heart rate of the measured enhanced heartbeat feature matrix spatiotemporal map and output the target instantaneous heart rate.

[0159] Optionally, the processing of the heart rate prediction network includes:

[0160] Perform channel cutting on the input enhanced heartbeat feature matrix spatiotemporal map to determine multi-channel input features;

[0161] Perform preliminary channel interaction based on multi-channel input features and output channel interaction features;

[0162] Perform convolution extraction on channel interaction features to generate low-level features;

[0163] Continuous residual learning and residual attention processing are performed based on low-level features to construct residual enhancement features;

[0164] Adaptive feature pooling is performed using residual enhancement features to determine global features;

[0165] The global features are sequentially subjected to linear transformation, ReLU nonlinear mapping, and Dropout regularization, and then linearly transformed to output the instantaneous heart rate.

[0166] Optionally, adaptive feature pooling is performed using residual enhancement features to determine global features, including:

[0167] Use the residual enhancement feature to perform global average pooling and determine the average pooling feature;

[0168] Perform global maximum pooling based on the residual enhancement features and output the maximum pooling features;

[0169] Concatenate the average pooling feature and the maximum pooling feature to construct the pooling concatenation feature;

[0170] After performing linear transformation based on the pooled splicing features, a softmax operation is performed to generate average pooling weights and maximum pooling weights;

[0171] The average pooling feature and the average pooling weight, the maximum pooling feature and the maximum pooling weight are weighted respectively and then concatenated to output the global feature.

[0172] An embodiment of the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the heart rate determination method based on millimeter wave radar and neural network as in any of the above embodiments.

[0173] An embodiment of the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the heart rate determination method based on millimeter wave radar and neural network as in any of the above embodiments are implemented.

[0174] An embodiment of the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the heart rate determination method based on millimeter wave radar and neural network as in any of the above embodiments.

[0175] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0176] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0177] 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 these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0178] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0179] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0180] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A heart rate determination method based on millimeter wave radar and neural network, characterized in that: include: Acquire synchronously collected original radar echo data and original ECG signals; Preprocessing the original ECG signal to generate a target ECG signal, and performing R-peak positioning on the target ECG signal using an initial sliding window to determine multiple initial instantaneous heart rates; According to the high-frequency body movement time determined by the amplitude of the original radar echo data, the original radar echo data and each of the initial instantaneous heart rates are subjected to data elimination, and target radar echo data and target instantaneous heart rate are output; intercepting the target radar echo data through the initial sliding window to construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target; If data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, data equalization is performed based on the target radar echo data and the initial sliding window to determine a training radar complex data matrix and a corresponding true instantaneous heart rate; Determining a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and determining a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training; The measured radar echo data is obtained, and the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data is input into the trained heart rate prediction network to output the measured instantaneous heart rate.

2. The heart rate determination method based on millimeter wave radar and neural network according to claim 1, characterized in that: If data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, data equalization is performed based on the target radar echo data and the initial sliding window to determine a training radar complex data matrix and a corresponding true instantaneous heart rate, including: If data imbalance is determined based on the initial radar complex data matrix to which the plurality of preset heart rate intervals belong, a marginal missing heart rate interval is determined; When the edge-missing heart rate interval is a low heart rate interval, data windowing is performed on the target radar echo data through a low heart rate sliding window and linear interpolation is performed to determine a plurality of low heart rate radar complex data matrices; Determining multiple initial low instantaneous heart rates of the target electrocardiogram signal based on the low heart rate sliding window, and performing multiplication operations on each of the initial low instantaneous heart rates using a window length ratio of the low heart rate sliding window to the initial sliding window to generate target low instantaneous heart rates corresponding to each of the low heart rate radar complex data matrices; When the edge-missing heart rate interval is a high heart rate interval, slicing and down-sampling the target radar echo data using a high heart rate sliding window to determine a plurality of high heart rate radar complex data matrices; constructing a plurality of initial high instantaneous heart rates of the target ECG signal according to the high heart rate sliding window, and multiplying each of the initial high instantaneous heart rates by a window length ratio of the high heart rate sliding window to the initial sliding window, respectively, to obtain a target high instantaneous heart rate associated with each of the high heart rate radar complex data matrices; If data balance is determined according to each of the preset heart rate intervals based on the initial radar complex data matrix, the low heart rate radar complex data matrix, and the high heart rate radar complex data matrix, a training radar complex data matrix is ​​formed through sparse stratified sampling, and the corresponding true instantaneous heart rate is determined.

3. The heart rate determination method based on millimeter wave radar and neural network according to claim 1, characterized in that: The method of determining a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and determining a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training, includes: After performing first-order filtering on the training radar complex data matrix, static clutter is filtered out using a moving target indication algorithm to determine a filtered radar complex data matrix; calculating phase changes in the filtered radar complex data matrix and converting each phase change into a displacement to construct a chest displacement matrix; Using an IIR bandpass filter to filter the chest cavity displacement matrix to generate a heartbeat feature matrix, and performing matrix normalization on the heartbeat feature matrix to determine a spatiotemporal map of the heartbeat feature matrix; Performing image enhancement on the heartbeat feature matrix spatiotemporal map to generate a training enhanced heartbeat feature matrix spatiotemporal map; Inputting the training enhanced heartbeat feature matrix spatiotemporal map into the heart rate prediction network to be trained, and outputting the predicted instantaneous heart rate; The heart rate prediction network to be trained is iteratively optimized based on the loss function value calculated based on the predicted instantaneous heart rate and the true instantaneous heart rate to determine a trained heart rate prediction network.

4. The heart rate determination method based on millimeter wave radar and neural network according to claim 1, characterized in that: The step of obtaining measured radar echo data, inputting a measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into a trained heart rate prediction network, and outputting the measured instantaneous heart rate includes: When measured radar echo data is collected, a measured sliding window associated with the measured radar echo data is used to perform data segmentation to construct a measured radar complex data matrix; Determine a spatiotemporal map of a measured enhanced heartbeat feature matrix based on the measured radar complex data matrix; The heart rate is predicted by the trained heart rate prediction network on the measured enhanced heartbeat feature matrix spatiotemporal map, and the target instantaneous heart rate is output.

5. The heart rate determination method based on millimeter wave radar and neural network according to claim 1, characterized in that: The processing of the heart rate prediction network includes: Perform channel cutting on the input enhanced heartbeat feature matrix spatiotemporal map to determine multi-channel input features; Perform preliminary channel interaction according to the multi-channel input features, and output channel interaction features; Performing convolution extraction on the channel interaction features to generate low-level features; Performing continuous residual learning and residual attention processing based on the low-level features to construct residual enhancement features; Adopting the residual enhancement feature to perform adaptive feature pooling to determine the global feature; The global features are sequentially subjected to linear transformation, ReLU nonlinear mapping, and Dropout regularization, and then linearly transformed to output the instantaneous heart rate.

6. The heart rate determination method based on millimeter wave radar and neural network according to claim 5, characterized in that: The adaptive feature pooling using the residual enhancement feature to determine the global feature includes: Performing global average pooling using the residual enhancement feature to determine an average pooling feature; Performing global maximum pooling according to the residual enhancement feature and outputting the maximum pooling feature; Concatenating the average pooling feature and the maximum pooling feature to construct a pooling concatenation feature; Performing a softmax operation after linear transformation based on the pooled splicing features to generate an average pooling weight and a maximum pooling weight; The average pooling feature and the average pooling weight, the maximum pooling feature and the maximum pooling weight are weighted and then concatenated to output a global feature.

7. A heart rate determination system based on millimeter wave radar and neural network, characterized in that: include: A data acquisition module is used to obtain synchronously collected original radar echo data and original electrocardiogram signals; an ECG signal processing module, configured to pre-process the original ECG signal to generate a target ECG signal, and perform R-peak positioning on the target ECG signal using an initial sliding window to determine a plurality of initial instantaneous heart rates; a data removal module, configured to remove the raw radar echo data and each of the initial instantaneous heart rates according to the high-frequency body movement time determined by the amplitude of the raw radar echo data, and output target radar echo data and target instantaneous heart rate; a data matrix construction module, configured to intercept the target radar echo data through the initial sliding window and construct an initial radar complex data matrix corresponding to the instantaneous heart rate of each target; a training data determination module configured to, if data imbalance is determined based on the initial radar complex data matrices belonging to the plurality of preset heart rate intervals, perform data equalization based on the target radar echo data and the initial sliding window to determine a training radar complex data matrix and a corresponding true instantaneous heart rate; A model training module is used to determine a training enhanced heartbeat feature matrix spatiotemporal map based on the training radar complex data matrix, and to determine a trained heart rate prediction network using the training enhanced heartbeat feature matrix spatiotemporal map and the real instantaneous heart rate training; The heart rate measurement module is used to obtain the measured radar echo data, input the measured enhanced heartbeat feature matrix spatiotemporal map constructed based on the measured radar echo data into the trained heart rate prediction network, and output the measured instantaneous heart rate.

8. A computer device, characterized in that: It includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor performs the steps of the heart rate determination method based on millimeter wave radar and neural network as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the heart rate determination method based on millimeter wave radar and neural network are implemented as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the heart rate determination method based on millimeter wave radar and neural network are implemented as described in any one of claims 1 to 6.

Images