Non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar

Abstract

The invention discloses a non-contact health and safety dynamic monitoring method based on a pulse ultra-wideband radar. The method comprises the steps of pulse ultra-wideband signal preprocessing, tumble behavior detection based on improved Yolo v11, human body heartbeat signal sensing based on Transform and the like. According to the invention, a Yolo v11 network is improved from a backbone network and a neck structure, a fusion channel space attention mechanism is added, and a voting mechanism is adopted to detect a falling behavior. A Transformer encoder uses a one-dimensional convolutional layer to extract time features related to heartbeat, uses a batch normalization layer to make a training process more effective and stable, uses a leakage ReLU layer as an activation function, maps encoding features to potential distribution through completely connected layers, each transposed convolutional layer of a decoder upsamples input by 4 times, and the decoder upsamples the input by 4 times. And transforming the decoding features through a one-dimensional convolutional layer, and executing nonlinear decomposition and heartbeat waveform recovery.

Description

Technical Field

[0001] The present invention relates to the technical field of personal health and safety detection, and specifically to a full-time non-contact personal health and safety dynamic monitoring method based on pulse ultra-wideband (IR-UWB) radar. Background Art

[0002] With the continuous development of society, the demand for monitoring people's own health status is increasing, and medical and health monitoring has become a huge market. Especially in the current increasingly aging society, non-contact human vital signs monitoring can not only conveniently monitor the health status of elderly users, but also screen for early-stage diseases and continuously monitor chronic diseases. This is a huge market worth hundreds of millions of yuan, with enormous economic benefits.

[0003] Traditional personal health and safety detection technologies are mainly divided into two types: contact sensing and non-contact sensing.

[0004] Contact sensing involves continuous signal monitoring using different sensor devices. This requires users to carry cumbersome wearable devices, such as using a medical breathing belt to monitor the breathing status of the person being tested, or using professional medical instruments to obtain an electrocardiogram of the human body to determine whether there is a risk of heart disease.

[0005] Non-contact sensing, based on computer vision, typically uses a camera to capture video of the person being sensed. The camera then analyzes subtle color changes in the facial area of ​​the person in the video to determine the subject's heart rate. This method is susceptible to obstruction and infringes on the subject's privacy. Acoustic sensing solutions, often integrated into smartphones, detect the acoustic signals emitted by the user's breathing to achieve acoustic respiration sensing.

[0006] Compared to the aforementioned sensing technologies, pulsed ultra-wideband radar technology can achieve high-precision personal health and safety monitoring, is relatively low-cost, and is more suitable for use in people's daily lives. However, current methods for dynamic monitoring of personal health and safety based on pulsed ultra-wideband (IR-UWB) radar are still immature. Therefore, a non-contact dynamic monitoring method for personal health and safety based on pulsed ultra-wideband radar is urgently needed to improve this technology. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention proposes a non-contact dynamic health and safety monitoring method based on pulse ultra-wideband radar, which has the advantages of non-contact vital sign signal extraction, generation of human respiratory signal waveform, separation of respiratory and heartbeat signals based on deep learning algorithm, and perception of fall behavior recognition based on YOLO.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar, comprising the following steps:

[0009] Step 1: Pulse UWB signal preprocessing

[0010] S1. Collect pulse ultra-wideband signals and set them to zero to eliminate the content in the blind area of ​​pulse ultra-wideband signal perception;

[0011] S2, use Hampel filtering to eliminate the outliers in the pulse ultra-wideband signal sequence, and then use SG filtering to make the entire complete time series smoother;

[0012] S3. Feature extraction and compression of pulse ultra-wideband signals are performed using an I / Q complex plane projection method, specifically:

[0013] S3-1: Construct a projection surface Y on the I\Q complex plane. The projection surface is a straight line on the I\Q plane.

[0014] S3-2: When calculating the projection of the human body reflection signal on the projection surface, assuming the projection coordinate is [cosθ sinθ], the projection surface corresponding to the reflection signal is Y = [Rea(H)Image(H)] * [cosθ sinθ] T , where Real(H) and Image(H) distributions are the real and imaginary parts of the pulse UWB signal H;

[0015] S3-3: Change the step size of θ according to the step size Δθ = π / 50, and obtain 100 corresponding projection component features from the step size range of [0, 2π];

[0016] S3-4: Finally, the projection component feature with the largest variance is selected as the reflection feature of the human body sign signal.

[0017] Step 2: Fall behavior detection based on improved Yolo v11

[0018] S1. Added the fused spatial channel attention mechanism, added the CBAM attention mechanism to the backbone network, used ResNext to replace the residual network in Yolo v11, and added the fused channel spatial attention mechanism in ResNext;

[0019] S2. Improved the backbone network and neck structure of the Yolo v11 network, replacing the LeakyReLu activation function in the backbone network with the hard-swish activation function;

[0020] S3, the original SPP structure was replaced by the ASPP structure, and the original neck structure in Yolo v11 was replaced by BiFPN and the improved pre-BiFPN;

[0021] S4, based on the improved Yolo v11 network structure, comprehensively considers the aspect ratio, the center position of the body, and the falling speed of the center point to determine whether a fall has occurred. IR-UWB radar is used for multi-angle fall detection, and a voting mechanism is used to detect radar data associated with fall behavior.

[0022] Step 3: Transformer-based human heartbeat signal perception

[0023] S1. The encoder applies a one-dimensional convolutional layer to extract temporal features related to heartbeats. Secondly, a batch normalization layer is used to make the training process more efficient and stable. Finally, a leaky ReLU layer is used as the activation function to add nonlinearity to the network. The encoded features are mapped to the latent distribution through a fully connected layer.

[0024] S2. The decoder uses a stack of one-dimensional transposed convolutional layers, batch normalization layers, and leaky ReLU layers. Each transposed convolutional layer upsamples the input by a factor of 4, increases the feature map length, and expands the potential representation to a longer sequence. The decoded features are transformed by a one-dimensional convolutional layer to match the output dimension, and the sigmoid function is used as the activation function.

[0025] Furthermore, the pulse ultra-wideband signal transceiver has feedthrough at close range, resulting in high-intensity noise interference in the pulse ultra-wideband signal within a certain distance, completely covering the target information within the target segment. This distance is considered to be a blind spot for pulse ultra-wideband signal perception.

[0026] Furthermore, the Transformer-based deep learning model uses signal representation in the complex in-phase / quadrature (I / Q) domain, takes the two components of the original complex RF signal as input, and leverages the universal approximation capability of self-attention and the generative potential of variational inference in the latent space to perform nonlinear decomposition and waveform recovery.

[0027] Furthermore, the Transformer encoder receives the I / Q signal r and encodes it into a potential representation h. Since r consists of complex numbers, two parallel encoders with the same structure are designed to process the in-phase component I and the orthogonal component Q of r, respectively. The decoder is used as an extended model to reconstruct the heartbeat waveform.

[0028] Compared with the existing technology, the technical solution of this application has the following beneficial effects:

[0029] 1. The present invention adopts non-contact vital sign signal extraction. On the basis of using radio frequency equipment to analyze the surrounding environment, it uses intelligent signal processing methods to separate the body reflection signal of the sensing target.

[0030] 2. The present invention realizes non-contact human breathing detection by generating a human breathing signal waveform and using a pulse ultra-wideband radio frequency device to detect the pulse interval of the human body reflection signal.

[0031] 3. The present invention separates the respiratory and heartbeat signals based on a deep learning algorithm. Compared with the respiratory signal, the heartbeat signal has a smaller amplitude and a faster interval, and is often drowned out by the human breathing process. An intelligent signal separation method is used to separate the heartbeat signal waveform from the original signal.

[0032] 4. The present invention is based on YOLO's fall behavior recognition. In order to better realize the real-time detection of human fall behavior, the YOLO model is used to analyze human behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall technical flow chart of the present invention;

[0034] Figure 2 Schematic diagram of projection of the I\Q complex plane of the present invention;

[0035] Figure 3 This is the network structure diagram of Yolo v11 of the present invention;

[0036] Figure 4 This is a schematic diagram of the voting algorithm for decision-making in the present invention;

[0037] Figure 5 This is the block structure diagram of the present invention based on Transformer;

[0038] Figure 6 Comparison results of fall detection based on improved Yolo v11;

[0039] Figure 7 Comparison results of human heartbeat signal perception based on Transformer. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] The non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar in this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0042] Step 1: Pulse UWB signal preprocessing

[0043] S1. Collect pulse ultra-wideband signals and set them to zero to eliminate the content in the blind area of ​​pulse ultra-wideband signal perception;

[0044] S2, use Hampel filtering to eliminate the outliers in the pulse ultra-wideband signal sequence, and then use SG filtering to make the entire complete time series smoother;

[0045] S3, using the I\Q complex plane projection method to extract and compress the pulse ultra-wideband signal;

[0046] Step 2: Fall behavior detection based on improved Yolo v11

[0047] S1. Added the fused spatial channel attention mechanism, added the CBAM attention mechanism to the backbone network, used ResNext to replace the residual network in Yolo v11, and added the fused channel spatial attention mechanism in ResNext;

[0048] S2. Improved the backbone network and neck structure of the Yolo v11 network, replacing the LeakyReLu activation function in the backbone network with the hard-swish activation function;

[0049] S3, the original SPP structure was replaced by the ASPP structure, and the original neck structure in Yolo v11 was replaced by BiFPN and the improved pre-BiFPN;

[0050] S4, based on the improved Yolo v11 network structure, comprehensively considers the aspect ratio, the center position of the body, and the falling speed of the center point to determine whether a fall has occurred. IR-UWB radar is used for multi-angle fall detection, and a voting mechanism is used to detect falls associated with IR-UWB radar data.

[0051] Step 3: Transformer-based human heartbeat signal perception

[0052] S1. The encoder applies a one-dimensional convolutional layer to extract temporal features related to heartbeats. Secondly, a batch normalization layer is used to make the training process more efficient and stable. Finally, a leaky ReLU layer is used as the activation function to add nonlinearity to the network. The encoded features are mapped to the latent distribution through a fully connected layer.

[0053] S2. The decoder uses a stack of one-dimensional transposed convolutional layers, batch normalization layers, and leaky ReLU layers. Each transposed convolutional layer receives m input channels and adopts a kernel size of n. Each transposed convolutional layer upsamples the input by a factor of 4, increases the feature map length, and expands the potential representation to a longer sequence. The decoded features are transformed by a one-dimensional convolutional layer to match the output dimension, and the sigmoid function is used as the activation function.

[0054] like Figure 2 As shown in the figure, as the human body continues to breathe, the signal shows a periodic "arc" state change. The human heartbeat signal is often obscured by the human breathing signal. To better extract the human life signal, the I / Q complex plane projection method is used to achieve data feature extraction and compression. The algorithm steps of the I / Q complex plane projection method are as follows:

[0055] 1) Construct a projection surface Y on the I\Q complex plane, which is a straight line on the I\Q plane;

[0056] 2) When calculating the projection of the human body reflection signal on the projection surface, assuming the projection coordinates are [cosθ sinθ], then the projection surface corresponding to the reflection signal is Y = [Rea(H)Image(H)] * [cosθ sinθ] T , where Real(H) and Image(H) distributions are the real and imaginary parts of the pulse UWB signal H;

[0057] 3) Change the step size of θ by Δθ = π / 50, and obtain 100 corresponding projection component features from the step size range of [0, 2π];

[0058] 4) Finally, the projection component feature with the largest variance is selected as the reflection feature of the human body sign signal.

[0059] like Figure 3 As shown in Figure 1, Yolo v11 mainly consists of two parts: the backbone network and the neck structure. The backbone network consists of the Focus structure, the Cross-Stage Partial Connection (CSP) structure, and the SPP structure. The Focus structure fuses image information onto channels through a slicing operation and then performs feature extraction on it, making feature extraction more comprehensive. The CSP structure divides the input feature map into two sub-maps, one of which is sent to the convolution layer for feature extraction, while the other sub-map is directly concat-operated with the convolution result. This can effectively reduce the computational complexity of classification problems and improve the learning ability of convolutional neural networks (CNNs). The SPP structure uses window pooling of different sizes to improve feature extraction. The neck structure guides the shallow network through the deep network to obtain more advanced semantic information.

[0060] In this example, the CBAM attention mechanism was added to the backbone network, and the residual network in Yolov11 was replaced with ResNext. A fused channel spatial attention mechanism was also added to ResNext. Comparative experiments with gradual additions verified that adding the attention mechanism to the backbone network effectively improved human recognition and reduced the interference of background images on human positions. Adding the attention mechanism to the ResNext residual structure improved the network's ability to recognize small human targets and enhanced fall detection at a distance from the radar.

[0061] Specifically, the Leaky ReLu activation function in the backbone network was replaced with the hard-swish activation function, which has good nonlinear expression capabilities and low computational complexity, allowing the network to better fit the data model. The ASPP structure can increase the network's receptive field and improve the network's ability to extract features. The ASPP structure replaced the original SPP structure. BiFPN and the improved pre-BiFPN were used to replace the original neck structure in Yolo v11, improving the network's ability to extract initial features and enabling the network to easily and quickly perform multi-scale feature fusion. Even with the increase in network structure, real-time detection can still be achieved. Improvements to the Yolo v11 network structure effectively improved the accuracy of network fall detection.

[0062] like Figure 4 As shown in the figure, when the accuracy of the neural network algorithm is high for single-frame human fall recognition, the fall situation in a single frame is no longer used as the alarm condition. Instead, a more complex calculation method is used to determine the fall behavior. The aspect ratio, the center position of the human body and the falling speed of the center point are comprehensively considered to comprehensively judge whether the fall behavior has occurred. In addition, IR-UWB radar is used to perform fall detection from multiple angles, and a voting algorithm is used to determine the occurrence of fall behavior.

[0063] like Figure 5 As shown in Figure 2, the Transformer is essentially a more complex encoder-decoder model. It's based on the Attention mechanism. Compared to models based on CNNs and recurrent neural networks (RNNs), the Attention mechanism can address the long-distance dependency issues inherent in RNNs and their variants, while also supporting parallel computation.

[0064] In this embodiment, the Transformer consists of an encoder and a decoder. The encoder receives an I / Q signal r and encodes it into a latent representation h. Since r consists of complex numbers, two parallel encoders with identical structures are designed to process r's in-phase component I and quadrature component Q, respectively. The encoder first applies a one-dimensional convolutional layer to extract temporal features related to heartbeats. A batch normalization layer is then used to make the training process more efficient and stable. A leaky ReLU layer is then used as the activation function to add nonlinearity to the network. Finally, a fully connected layer maps the encoded features to the latent distribution.

[0065] Specifically, the decoder is used as an extended model to reconstruct the heartbeat waveform, using a stack of one-dimensional transposed convolutional layers, batch normalization layers, and leaky ReLU layers. Each transposed convolutional layer receives m input channels and adopts a kernel size of n. As each transposed convolutional layer upsamples the input by a factor of 4 (i.e., stride 4), the feature map length increases and the potential representation is expanded to a longer sequence. Finally, the decoded features are transformed by a one-dimensional convolutional layer to match the output dimension, and the sigmoid function is used as the activation function.

[0066] In summary, the present invention adopts non-contact vital sign signal extraction, and uses an intelligent signal processing method to separate the body reflection signal of the sensing target on the basis of using radio frequency equipment to detect the surrounding environment. By generating the human breathing signal waveform and using pulse ultra-wideband radio frequency equipment, the pulse interval of the human reflection signal is detected to achieve non-contact human breathing detection. The breathing and heartbeat signals are separated based on the deep learning algorithm. Compared with the breathing signal, the heartbeat signal has a smaller amplitude and a faster interval, and is often submerged in the human breathing process. The intelligent signal separation method is used to separate the heartbeat signal waveform from the original signal. The fall behavior recognition based on YOLO realizes the real-time detection of human fall behavior. Figure 6 and Figure 7 This shows the improvement effect of the present invention.

[0067] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar, characterized in that: The steps include: Step 1: Pulse UWB signal preprocessing S1. Collect the pulse ultra-wideband (IR-UWB) signal and set it to zero to eliminate the content in the blind area of ​​the pulse ultra-wideband signal perception; S2, use Hampel filtering to eliminate the outliers in the pulse ultra-wideband signal sequence, and then use SG filtering to make the entire complete time series smoother; S3, using the I\Q complex plane projection method to extract and compress the pulse ultra-wideband signal; Step 2: Fall behavior detection based on improved Yolo v11 S1. Added a fused spatial channel attention mechanism, added a convolutional block attention module (CBAM) attention mechanism to the backbone network, replaced the residual network in Yolo v11 with the aggregated residual transform of deep neural networks (ResNext), and added a fused channel spatial attention mechanism in ResNext; S2. Improved the backbone network and neck structure of the Yolo v11 network, replacing the Leaky ReLu activation function in the backbone network with the hard-swish activation function; S3. The original spatial pyramid pooling (SPP) structure was replaced with the atrous spatial pyramid pooling (ASPP) structure, and the original neck structure in Yolo v11 was replaced with the bidirectional feature pyramid network (BiFPN) and the improved pre-BiFPN. S4, based on the improved Yolo v11 network structure, comprehensively considers the aspect ratio, the center position of the body, and the falling speed of the center point to determine whether a fall has occurred. IR-UWB radar is used for multi-angle fall detection, and a voting mechanism is used to detect radar data associated with fall behavior. Step 3: Transformer-based human heartbeat signal perception S1. The encoder applies a one-dimensional convolutional layer to extract temporal features related to heartbeats. Secondly, a batch normalization layer is used to make the training process more efficient and stable. Finally, a leaky ReLU layer is used as the activation function to add nonlinearity to the network. The encoded features are mapped to the latent distribution through a fully connected layer. S2. The decoder uses a stack of one-dimensional transposed convolutional layers, batch normalization layers, and leaky ReLU layers. Each transposed convolutional layer upsamples the input by a factor of 4, increases the feature map length, and expands the potential representation to a longer sequence. The decoded features are transformed by a one-dimensional convolutional layer to match the output dimension, and the sigmoid function is used as the activation function.

2. The non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar according to claim 1 is characterized in that: The algorithm steps based on the I\Q complex plane projection method are as follows: 1) Construct a projection surface Y on the I\Q complex plane, which is a straight line on the I\Q plane; 2) Calculate the projection of the human body reflection signal on the projection surface. Assuming the projection coordinate is [cosθ sinθ], the projection surface corresponding to the reflection signal is Y = [Rea(H) Image(H)] * [cosθ sinθ] T , where Real(H) and Image(H) are the real and imaginary parts of the pulse UWB signal H, respectively; 3) Change the step size of θ by Δθ = π / 50, and obtain 100 corresponding projection component features from the step size range of [0, 2π]; 4) Finally, the projection component feature with the largest variance is selected as the reflection feature of the human body sign signal.

3. The non-contact health and safety dynamic monitoring method based on pulse ultra-wideband radar according to claim 1 is characterized in that: The Transformer-based deep learning model uses signal representation in the complex in-phase / quadrature (I / Q) domain, takes the two components of the original complex RF signal as input, and utilizes the universal approximation capability of self-attention and the generative potential of variational inference in the latent space to perform nonlinear decomposition and waveform recovery. The Transformer encoder receives the I\Q signal r and encodes it into a latent representation h. Since r consists of complex numbers, two parallel encoders with the same structure are designed to process the in-phase component I and the quadrature component Q of r, respectively. The decoder is used as an extended model for reconstructing the heartbeat waveform.

Images