Human posture estimation and non-contact vital sign monitoring method and system based on CSI

Abstract

The application relates to the technical field of wireless sensing, and discloses a human posture estimation and non-contact vital sign monitoring method and system based on CSI. WiFi channel state information (CSI) is collected, and dynamic CSI feature vectors are obtained through preprocessing. Features are mapped to human key point feature space through Transformer cross attention, spatial reasoning is performed through graph convolutional neural network fusion of human skeleton topology constraints, and human posture estimation results are output. Respiratory rate and heart rate are extracted based on multi-band frequency domain analysis, and vital sign monitoring results are output after being optimized in combination with posture information. Without cameras and wearable devices, non-contact human perception that is privacy-safe and wall-penetrating can be realized, and the method has the advantages of high precision, strong generalization and low-cost deployment, and is suitable for smart home, medical monitoring, emergency rescue and other scenes.

Description

Technical Field

[0001] This application relates to the field of wireless sensing technology, and in particular to a non-contact method and system for human posture estimation and vital signs monitoring based on CSI. Background Technology

[0002] With the rapid development of smart homes, healthcare, public safety and other fields, the demand for human posture estimation and vital sign monitoring is increasing. Existing technologies are mainly divided into the following four categories, all of which have unavoidable technical defects.

[0003] The first category is posture recognition technology based on computer vision. The mainstream solution uses RGB cameras combined with deep convolutional neural networks to detect key points on the human body. Although this solution has high detection accuracy, it has serious drawbacks: First, it directly collects human image information, which poses a serious risk of privacy leakage in privacy-sensitive scenarios such as homes and medical settings, making it unacceptable to users. Second, visible light cannot penetrate walls, furniture, ruins, and other obstructions, making it completely ineffective in scenarios such as monitoring through walls and disaster relief. Third, it is highly dependent on lighting conditions, and the recognition accuracy drops sharply in low-light environments such as nighttime and smoke, resulting in insufficient robustness. Fourth, large-scale deployment is costly, requiring multiple devices to work together and complex wiring, making maintenance difficult.

[0004] The second category is vital sign monitoring technology based on wearable devices. Existing solutions mostly use devices such as smart bracelets and chest strap sensors to collect vital signs through PPG and accelerometer sensors. However, these solutions have the following drawbacks: First, they require the monitored person to wear them close to their body, resulting in poor comfort during long-term use and low user compliance in scenarios such as sleep monitoring. Second, the devices have limited battery life, requiring frequent charging and easily leading to data interruption due to forgetting to wear them. Third, they cannot adapt to emergency scenarios. In disaster relief, emergency triage, and other scenarios, it is impossible to require the monitored person to wear the device immediately, resulting in a delayed response. Fourth, the cost of monitoring multiple people increases linearly with the number of people, making it difficult to scale up applications.

[0005] The third category is human body sensing technology based on millimeter-wave radar. Some solutions use 60GHz FMCW millimeter-wave radar to achieve human body detection and attitude estimation. Its drawbacks are: first, the hardware cost is high, and the cost of a single device is high, making it difficult to deploy on a large scale in civilian applications; second, the signal is highly directional, and the coverage angle of a single device is narrow, requiring multiple devices to work together to achieve whole-house coverage; third, the high-frequency signal is greatly attenuated by walls, and the ability to detect through walls is weak, making it unsuitable for obstructed scenarios; fourth, it requires dedicated hardware equipment and cannot reuse the existing widely deployed WiFi infrastructure, resulting in poor deployment flexibility.

[0006] The fourth category is human body sensing technology based on WiFi signals. There are currently a few sensing schemes based on WiFi signals, most of which use Received Signal Strength Indicator (RSSI) to detect the presence of human bodies. They can only determine whether there are people present and cannot achieve precise posture estimation.

[0007] In summary, existing technologies cannot simultaneously achieve privacy protection, wear-free operation, through-wall monitoring capabilities, and low-cost scalable deployment. Furthermore, they struggle to simultaneously achieve high-precision human pose estimation and vital sign monitoring. Additionally, they suffer from core technical challenges such as weak cross-environment generalization ability, insufficient robustness in occluded scenarios, and high computational overhead, making them difficult to adapt to edge device deployments. Summary of the Invention

[0008] To address the aforementioned technical issues, this application provides a method and system for non-contact monitoring of human posture estimation and vital signs based on CSI.

[0009] Firstly, this application provides a method and system for non-contact monitoring of human posture and vital signs based on CSI, employing the following technical solution: A non-contact method for human posture estimation and vital sign monitoring based on CSI includes the following steps: S1, Collect CSI data of WiFi channel, the CSI data includes amplitude and phase information of OFDM subcarriers corresponding to multiple sets of transmit-receive antenna pairs; S2, preprocess the CSI data to obtain a dynamic CSI feature vector; S3, through the Transformer cross attention mechanism, the dynamic CSI feature vector is mapped to the human body key point feature space to obtain the initial features of each key point corresponding to the human skeleton topology map; S4. Based on the human skeleton topology map, spatial reasoning is performed on the initial features of each key point through a graph convolutional neural network to obtain the refined features of key points that integrate spatial constraint relationships. S5. Based on the refined features of the key points, the 3D coordinates and confidence of each key point are obtained by regression, and the human pose estimation result is output after time-series smoothing. S6, based on the preprocessed CSI data, extracts the raw data of respiratory rate and heart rate in parallel with steps S3 to S5; after step S5 is completed, optimizes the respiratory rate and heart rate data by combining the exercise intensity corresponding to the human posture estimation results, and outputs the vital signs monitoring results.

[0010] Optionally, the preprocessing in step S2 specifically includes: performing linear calibration on the phase information of the original CSI data to eliminate hardware phase offset and carrier frequency offset, and completing phase unwrapping; based on the Fresnel zone electromagnetic wave propagation model, separating the static multipath component and dynamic multipath component in the CSI data, and extracting the dynamic component corresponding to human activity; fusing amplitude, phase, phase difference, Doppler frequency shift and power spectral density information to construct a dynamic CSI feature vector.

[0011] Optionally, in step S4, the human skeleton topology is a COCO standard 17 keypoint skeleton topology, which includes 17 nodes and 16 bone connection edges. The graph convolutional neural network uses two stacked graph convolutional layers, performs message passing through a symmetric normalized adjacency matrix, aggregates the feature information of adjacent nodes, and retains the initial feature information through residual connections.

[0012] Optionally, step S6 specifically includes: performing a 0.1-0.5Hz bandpass filter on the preprocessed CSI amplitude sequence, and obtaining the respiratory rate by Fast Fourier Transform (FFT) peak detection; performing phase unwrapping on the preprocessed CSI phase sequence, and then performing a 0.8-2.0Hz bandpass filter, selecting the subcarrier with the highest signal-to-noise ratio, and obtaining the heart rate by FFT peak detection; calculating the motion intensity based on the human posture estimation results, classifying the motion state based on the motion intensity, and performing filtering optimization or effectiveness determination on the respiratory rate and heart rate based on the motion state, while simultaneously completing the detection of abnormal vital signs.

[0013] Optionally, it also includes an environment adaptation step: Real-time monitoring of the difference between the current CSI feature distribution and the baseline environment feature distribution; triggering environment adaptation when the difference exceeds a preset threshold. By using a low-rank adaptive mechanism, the parameters of the inserted low-rank matrix are adjusted only while the weights of the backbone model are frozen. Combined with elastic weight reinforcement regularization to prevent catastrophic forgetting, the model parameters are updated through online gradient accumulation, enabling rapid adaptation to new environments.

[0014] Secondly, this application provides a CSI-based non-contact human posture estimation and vital sign monitoring system, which adopts the following technical solution: The CSI-based human posture estimation and vital sign non-contact monitoring system includes a CSI data acquisition module, a signal preprocessing module, a cross-modal feature mapping module, a graph neural network inference module, a posture estimation output module, and a vital sign extraction module that are connected in sequence via communication. The CSI data acquisition module is used to acquire CSI data of the WiFi channel. The CSI data includes amplitude and phase information of the corresponding OFDM subcarriers for multiple sets of transmit-receive antenna pairs. The signal preprocessing module is used to preprocess the CSI data and output a dynamic CSI feature vector; The cross-modal feature mapping module is used to map the dynamic CSI feature vector to the human keypoint feature space through the Transformer cross attention mechanism, and output the initial features of each keypoint corresponding to the human skeleton topology map. The graph neural network inference module is used to perform spatial inference on the initial features of each key point based on the human skeleton topology map, and output the refined features of key points that integrate spatial constraint relationships. The pose estimation output module is used to refine the features of the key points, regress to obtain the 3D coordinates and confidence of each key point, and output the human pose estimation result after time-series smoothing. The vital signs extraction module is used to extract respiratory rate and heart rate data based on preprocessed CSI data through multi-band frequency domain analysis, optimize the respiratory rate and heart rate data by combining the exercise intensity corresponding to the human posture estimation results, and output vital signs monitoring results.

[0015] Optionally, it also includes an adaptive learning module, which is communicatively connected to the signal preprocessing module, the cross-modal feature mapping module, and the graph neural network inference module, respectively. The adaptive learning module is used to detect the drift of environmental feature distribution, complete the online environmental adaptation of the model through a low-rank adaptive mechanism, and prevent catastrophic forgetting of the model through elastic weight consolidation regularization.

[0016] Optionally, it also includes a multi-static fusion module, which is communicatively connected to the CSI data acquisition module and is used to fuse CSI data collected from multiple WiFi access points.

[0017] Optionally, the CSI data acquisition module uses a WiFi device that supports CSI data export, including at least one of an ESP32-S3 chip and an Intel 5300 network card; all data processing and inference processes of the system are completed on the local edge device.

[0018] In summary, this application includes at least the following beneficial technical effects: This application achieves human body perception through WiFi channel state information, eliminating the need for image data collection and fundamentally avoiding privacy risks. Furthermore, it eliminates the need for the monitored person to wear any devices, significantly improving user compliance and broadening its applicability. Utilizing the wall-penetrating properties of WiFi signals, stable monitoring can be achieved even in obstructed environments such as walls, furniture, and ruins, perfectly adapting to special applications such as disaster relief and monitoring through walls. Through phase-amplitude joint analysis and Fresnel zone electromagnetic wave propagation modeling, effective information related to human activity within WiFi signals is fully extracted. Combined with cross-modal feature mapping and graph neural network-based human skeleton topological constraints, this effectively improves posture detection. The system boasts high accuracy in state estimation and robustness in occluded scenarios; it can simultaneously achieve human pose estimation and vital sign monitoring without requiring additional hardware or computational overhead, significantly reducing system deployment and usage costs; through a low-rank adaptive online learning mechanism, it can quickly adapt to different deployment environments without the need for extensive labeled data and model retraining, significantly improving the system's cross-environment generalization ability and engineering feasibility; the overall solution can directly reuse existing widely deployed WiFi infrastructure, has low computational overhead, and is adaptable to various edge device deployments, combining practicality and economy, and can be widely applied in multiple fields such as smart homes, medical health monitoring, emergency rescue, and personnel safety monitoring. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the non-contact monitoring method for human posture estimation and vital signs in this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0021] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] This application discloses a non-contact method for human posture estimation and vital sign monitoring based on CSI, referring to... Figure 1 This includes the following steps: S1, Collect CSI data of WiFi channel, the CSI data includes amplitude and phase information of OFDM subcarriers corresponding to multiple sets of transmit-receive antenna pairs; S2, preprocess the CSI data to obtain a dynamic CSI feature vector; S3, through the Transformer cross attention mechanism, the dynamic CSI feature vector is mapped to the human body key point feature space to obtain the initial features of each key point corresponding to the human skeleton topology map; S4. Based on the human skeleton topology map, spatial reasoning is performed on the initial features of each key point through a graph convolutional neural network to obtain the refined features of key points that integrate spatial constraint relationships. S5. Based on the refined features of the key points, the 3D coordinates and confidence of each key point are obtained by regression, and the human pose estimation result is output after time-series smoothing. S6, based on the preprocessed CSI data, extracts the raw data of respiratory rate and heart rate in parallel with steps S3 to S5; after step S5 is completed, optimizes the respiratory rate and heart rate data by combining the exercise intensity corresponding to the human posture estimation results, and outputs the vital signs monitoring results.

[0023] In this embodiment, the following specific solution is adopted: The first step is CSI data acquisition.

[0024] WiFi devices supporting CSI data export, including the ESP32-S3 chip and Intel 5300 network card, are used to collect channel state information of WiFi signals. The WiFi devices operate in the 2.4GHz band (for wall-penetrating scenarios) or the 5GHz band (for high-precision scenarios), following the 802.11n protocol. Each transmit-receive (TX-RX) antenna pair collects CSI data from 56 effective OFDM subcarriers. The data format is a complex matrix, containing amplitude and phase information for each subcarrier. The sampling rate is set according to the scenario: a typical sampling rate of 100Hz for attitude recognition scenarios, 50Hz for vital sign monitoring scenarios, and can be reduced to 20Hz for emergency rescue scenarios to reduce power consumption.

[0025] The CSI data expression is: ; in, The serial number of the transmit-receive antenna pair. , This represents the total number of antenna pairs; The sequence number of the OFDM subcarrier. , This represents the total number of subcarriers, with a typical value of 56. For the first The antenna pair The signal amplitude of each subcarrier; For the first The antenna pair The signal phase of each subcarrier; The unit is imaginary. The data sampling rate can be adjusted within the range of 20Hz-200Hz depending on the application scenario.

[0026] The second step is CSI data preprocessing.

[0027] This step is divided into three sub-steps: phase calibration, dynamic component extraction, and feature construction. Phase calibration: Using the 28th subcarrier in the middle position as the reference subcarrier, calculate the phase slope corresponding to the carrier frequency offset, and perform linear compensation on the original phase to eliminate hardware phase offset and carrier frequency offset; unwrap the calibrated phase of adjacent frames, subtract 2π when the adjacent phase difference is greater than π, and add 2π when it is less than -π to eliminate 2π phase jump and ensure the continuity of phase information.

[0028] Dynamic component extraction: Based on the first Fresnel zone model, the static baseline of CSI data is calculated through a sliding window. The window size is set according to the environmental stability, with 50-100 frames in a stable environment and 20-30 frames in a dynamic environment. The static baseline is subtracted from the original CSI data to obtain the dynamic CSI components caused by human activity. Then, the frequency range corresponding to human activity is retained by bandpass filtering. The passband for posture recognition is 0.1-20Hz, the passband for fall detection is extended to 0.1-30Hz, and the passbands for vital sign monitoring are 0.1-0.5Hz and 0.8-2.0Hz, respectively.

[0029] Feature construction: The amplitude information is Z-score normalized, the phase difference between adjacent frames is calculated to represent the motion speed, the Doppler frequency shift is calculated based on the phase difference, the power spectral density of the amplitude sequence is calculated by FFT, and finally the amplitude, phase, phase difference, Doppler frequency shift and power spectral density information are concatenated to obtain a 56×5-dimensional dynamic CSI feature vector, corresponding to a total dimension of 280.

[0030] Phase calibration is used to eliminate hardware phase offset and carrier frequency offset, complete phase unwrapping, and eliminate 2π phase jumps; dynamic component extraction is based on the first Fresnel zone electromagnetic wave propagation model, separating static multipath and dynamic multipath components. The formula for calculating the radius of the first Fresnel zone is: ; in, The radius of the first Fresnel zone; The wavelength of the WiFi signal; This refers to the distance from the transmitting antenna to the human body. This refers to the distance from the human body to the receiving antenna.

[0031] The static baseline is calculated by sliding window, and the dynamic components corresponding to human activities are extracted. Noise is filtered out by bandpass filtering. The feature construction integrates amplitude, phase, phase difference, Doppler frequency shift and power spectral density information to generate a multi-dimensional dynamic CSI feature vector.

[0032] The third step is cross-modal feature mapping.

[0033] The Transformer cross-attention mechanism maps dynamic CSI feature vectors to the human keypoint feature space. First, a linear projection layer maps the 280-dimensional CSI feature vectors to a 64-dimensional feature embedding space. Then, a 4-head cross-attention mechanism is used, with 17 learnable human keypoint query vectors as query Q (17×64 dimension) and CSI feature embeddings as keys K and values ​​V (N×64 dimension), where N is the number of antenna pairs. Attention weights are calculated and features are aggregated, ultimately outputting 17×64-dimensional initial keypoint features, which correspond one-to-one with the 17 human keypoints in the COCO standard.

[0034] The fourth step is spatial reasoning using graph neural networks.

[0035] This step performs spatial reasoning based on the COCO standard's 17-keypoint human skeleton topology graph. The human skeleton topology graph contains 17 nodes and 16 bone connection edges. The 17 nodes of the topology graph are nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle. The 16 edges correspond to bone connection relationships, including nose-left eye, nose-right eye, left shoulder-left elbow, left elbow-left wrist, etc.

[0036] Spatial reasoning is performed using a two-layer stacked graph convolutional neural network.

[0037] First, construct the corresponding adjacency matrix. Then, symmetric normalization is performed to obtain the normalized adjacency matrix. Adjacency matrix In the code, if there is a skeletal connection between nodes, the corresponding element is 1; otherwise, it is 0.

[0038] First, the adjacency matrix is ​​symmetrically normalized to obtain the normalized adjacency matrix. ; ; in, Adjacency matrix The corresponding degree matrix.

[0039] The first layer of graph convolution aggregates the first-order neighbor features of each node. The activation function outputs the first layer of features; The second-layer graph convolution aggregates second-order neighbor features and propagates global skeleton constraints. The activation function outputs the second layer of features; Finally, through residual connections, the second-layer features are added to the initial keypoint features to obtain a 17×64-dimensional refined keypoint feature that incorporates spatial constraints. The specific calculation formula is as follows: ; ; Finally, the keypoint refined features are obtained through residual connections: ; in, Initial feature matrix for key points; , is a trainable weight matrix; It is a linear rectification activation function; , These are the output features of the first and second layer graph convolutions, respectively.

[0040] Step 5: Pose estimation output.

[0041] Regression is achieved through two independent fully connected layers. The first fully connected layer is a coordinate regression head, taking 64-dimensional features as input and outputting the 3D coordinates (x, y, z) of each keypoint, resulting in a 17×3 coordinate matrix. The second fully connected layer is a confidence regression head, taking 64-dimensional features as input, and outputting the detection confidence of each keypoint after passing through a sigmoid activation function, ranging from 0 to 1, resulting in a 17×1 confidence matrix. Keypoint coordinates with confidence scores below 0.3 are set to zero to filter out noise. Then, Kalman filtering is used to temporally smooth the keypoint coordinates across consecutive frames. The Kalman gain is set according to the motion speed: 0.5-0.7 for fast motion scenes and 0.2-0.4 for slow motion scenes, ultimately outputting a stable human pose estimation result.

[0042] Step 6: Monitoring vital signs.

[0043] Executed in parallel with attitude estimation, based on preprocessed CSI data, respiratory rate and heart rate are extracted through multi-band frequency domain analysis, and the corresponding exercise intensity is optimized by combining the attitude estimation results. Specifically, this includes: Respiratory rate extraction: The preprocessed CSI amplitude sequence is bandpass filtered at 0.1-0.5Hz to retain the frequency band corresponding to human respiration. The filtered signals of 56 subcarriers are summed to enhance the signal-to-noise ratio. Then, FFT is used for spectrum analysis to find the strongest peak of the power spectrum in the 0.1-0.5Hz frequency band. The frequency corresponding to the peak is multiplied by 60 and converted into a respiratory rate in breaths per minute. Heart rate extraction: After unwrapping the preprocessed CSI phase sequence, a bandpass filter of 0.8-2.0Hz is applied to retain the frequency band corresponding to the human heartbeat. The signal-to-noise ratio of each subcarrier is calculated, and the subcarrier signal with the highest signal-to-noise ratio is selected. Spectral analysis is performed using FFT to find the strongest peak of the power spectrum in the 0.8-2.0Hz frequency band. The frequency corresponding to the peak is multiplied by 60 and converted into heart rate in beats per minute. Data optimization and anomaly detection: Human motion intensity is calculated based on changes in keypoint coordinates across consecutive frames. The motion intensity range is 0-1. When the motion intensity is ≤0.2, it is considered a stationary state, and the respiratory rate and heart rate are directly output. When the motion intensity is 0.2 < ≤0.6, it is considered a moderate activity state, and the data is smoothed and optimized using Kalman filtering. When the motion intensity is >0.6, it is considered a strenuous activity state, and the data is marked as invalid. Anomaly thresholds are also set: when the respiratory rate is <5 breaths / min or >35 breaths / min, or the heart rate is <40 beats / min or >140 beats / min, an anomaly alarm is triggered.

[0044] Furthermore, this method also includes an environment adaptation step, which achieves rapid adaptation to the new environment through a low-rank adaptation mechanism, specifically including: Environment drift detection: The KL divergence between the current environment CSI feature distribution and the training environment baseline feature distribution is calculated in real time. At the same time, the HNSW nearest neighbor distance between the current CSI feature and the baseline fingerprint database is calculated. When the KL divergence is >0.5 or the nearest neighbor distance is >0.8, it is determined to be an environment drift and environment adaptation is triggered. Low-rank adaptive update: Freeze all weights of the backbone model and insert trainable low-rank matrices into the convolutional and fully connected layers of the model. With B, the encoder layer rank r=4, the decoder layer rank r=8, and the output layer rank r=16, only the parameters of the low-rank matrix are updated; Forgetting-prevention regularization: Elastic weights are used to reinforce regularization. The importance of parameters is weighted based on the Fisher information matrix. Changes in parameters that are of high importance in the old environment are penalized more strongly to prevent the model from forgetting knowledge of the old environment when adapting to the new environment. Online gradient accumulation: Gradients are calculated using small batches of real-time data of 1-10 frames. Parameters are updated every 10 steps after accumulating gradients to avoid instability in parameter updates caused by noise in a single frame. Adaptation to a new environment can be completed in 10-30 seconds.

[0045] In this embodiment, the core of the above content is the calculation of LoRA effective weights: ; in, The effective weights of the layer; Pre-train weights for the frozen backbone model; This is the LoRA scaling factor, used to control the strength of the adaptation. For dimension A trainable low-rank matrix, where B is a matrix of dimension 1. Trainable low-rank matrices Let be the rank of the low-rank matrix, and , The output dimension of the weight matrix. is the input dimension of the weight matrix. Elastic weights are used to reinforce regularization and prevent catastrophic forgetting. Parameter updates are completed through online gradient accumulation, enabling adaptation to new environments without retraining the backbone model.

[0046] On the other hand, the present invention provides a CSI-based human posture estimation and non-contact monitoring system for vital signs, which is used to implement the above method. It includes a CSI data acquisition module, a signal preprocessing module, a cross-modal feature mapping module, a graph neural network inference module, and a posture estimation output module that are connected in sequence, as well as a vital signs extraction module that is connected in communication with the signal preprocessing module and the posture estimation output module. An adaptive learning module and a multi-static fusion module can also be added according to application requirements.

[0047] The specific implementation methods of the above modules are as follows: CSI data acquisition module: The core adopts the ESP32-S3 chip, which has low single-chip cost and supports WiFi CSI data acquisition. It can be powered by battery or wired power and evenly deployed at multiple points in the monitoring area. It communicates with the edge server via WiFi to synchronously acquire CSI data with a clock synchronization error of <5ms. The multi-static fusion module can fuse CSI data from 3-6 WiFi access points and dynamically allocate weights based on the signal-to-noise ratio and viewing angle geometry to achieve 360-degree full-view coverage and eliminate single-view occlusion.

[0048] Signal preprocessing module: Deployed in the edge server, it has built-in processing units for phase calibration, sliding window filtering, normalization, and multipath component extraction. After receiving the raw CSI data, it completes the preprocessing and outputs dynamic CSI feature vectors with a processing delay of <1ms.

[0049] The cross-modal feature mapping module, graph neural network inference module, and pose estimation output module together constitute the model inference engine, which can be deployed on edge servers or cloud servers. It has built-in pre-trained Transformer and graph convolutional neural network models, takes dynamic CSI feature vectors as input, and outputs human pose estimation results. It supports INT8 quantization compression, reducing the model size by 75%, and can be adapted to edge devices such as Raspberry Pi and ESP32-S3.

[0050] Vital signs extraction module: Deployed in parallel with the model inference engine, it has built-in bandpass filtering unit, FFT spectrum analysis unit, and motion state judgment unit. It takes preprocessed CSI data and attitude estimation results as input and outputs vital signs monitoring results and abnormal alarms.

[0051] Adaptive learning module: It has built-in environment drift detection unit, LoRA parameter update unit, and EWC regularization unit, which can automatically trigger online model adaptation when environment drift is detected, without the need for manual intervention and labeled data.

[0052] All data acquisition, processing, and inference processes are completed locally without uploading to the cloud. Communication uses TLS 1.3 encryption to ensure data security and privacy protection. At the same time, it provides RESTful API and WebSocket real-time push interface to support integration with third-party systems such as smart home, medical monitoring system, and emergency rescue platform.

[0053] The following are examples of implementation scenarios for this application in specific situations: Example 1 of implementation scenario: Home-based elderly care scenario: The hardware configuration of this embodiment is as follows: 1 WiFi router that supports CSI export, 3 ESP32-S3 nodes deployed in the four corners of the room ceiling, and 1 Raspberry Pi4 as an edge inference server.

[0054] The implementation steps are as follows: First, complete the hardware deployment and network connection. Deploy the Docker container of this system on the edge server. Complete the parameter configuration through the web interface. Set the sampling rate to 50Hz, the working frequency band to 2.4GHz, the bandpass limit to 30Hz to adapt to fall detection, and the confidence threshold to 0.25 to improve the recall rate. The system automatically collects a static baseline of an empty room for 30 minutes to complete the construction of the environmental fingerprint database and the adaptation of the SONA module (adaptive learning module). The system runs continuously, outputting the posture estimation result once per second and the vital signs data once every 15 seconds. When a fall or abnormal breathing / heart rate is detected, an alarm is immediately pushed to the family members through the preset interface.

[0055] This embodiment enables 24-hour non-contact monitoring of the elderly in a home setting, without the need for wearing any devices, eliminating the risk of privacy leaks, and achieving stable monitoring through furniture and walls.

[0056] Example 2 of implementation scenario: Disaster relief and search and rescue scenario: The hardware configuration of this embodiment is as follows: 6 battery-powered portable ESP32-S3 nodes, 1 laptop computer as a field inference server, and 1 portable WiFi hotspot.

[0057] The implementation steps are as follows: The rescue team completes the circular deployment of 6 nodes around the ruins within 5 minutes, starts the system and configures it to wall-penetrating mode, sets the sampling rate to 20Hz, the operating frequency band to 2.4GHz, turns off heart rate detection, focuses on respiratory signal extraction, and sets the FFT window to 20 seconds to extract weak respiratory signals from strong noise; the system collects CSI data, detects the respiratory signal source and completes the location through a multi-static fusion algorithm, automatically draws a heat map to mark the location of the trapped person, marks it immediately when a respiratory signal is detected, and triggers a red alarm when the respiratory rate is lower than 5 breaths / minute to guide the rescue team to prioritize rescue.

[0058] This embodiment can penetrate concrete walls and rubble layers to achieve non-contact life detection and positioning of trapped personnel. Compared with traditional life detectors, it has stronger applicability in obstructed scenarios.

[0059] Implementation Scenario Example 3: Multi-person monitoring scenario in a hospital ward: The hardware configuration of this embodiment is as follows: two WiFi routers are deployed at both ends of the ward, 12 ESP32-S3 nodes are used for cross-coverage of two per bed, and one server is deployed at the nurse station.

[0060] The implementation steps are as follows: The system collects baseline data of empty beds, marks the bed number after the patient is admitted, separates the signal of each bed through multi-static geometric positioning, and the SONA module creates an independent adaptive for each bed; the system simultaneously monitors the vital signs and posture of 6 patients, outputs monitoring data every 10 seconds, and displays the respiratory rate and heart rate curves of each patient in real time on the large screen at the nurse station. When respiratory arrest or abnormal heart rate is detected, an alarm is triggered immediately; the system automatically calibrates with the bedside monitor data every hour to ensure monitoring accuracy.

[0061] This embodiment enables non-contact vital sign monitoring for multiple patients, eliminating the need for electrodes and chest straps, thus improving patient comfort and reducing the workload of medical staff.

[0062] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for human pose estimation and non-contact monitoring of vital signs based on CSI, characterized in that, Includes the following steps: S1, Collect CSI data of WiFi channel, the CSI data includes amplitude and phase information of OFDM subcarriers corresponding to multiple sets of transmit-receive antenna pairs; S2, preprocess the CSI data to obtain a dynamic CSI feature vector; S3, through the Transformer cross attention mechanism, the dynamic CSI feature vector is mapped to the human body key point feature space to obtain the initial features of each key point corresponding to the human skeleton topology map; S4. Based on the human skeleton topology map, spatial reasoning is performed on the initial features of each key point through a graph convolutional neural network to obtain the refined features of key points that integrate spatial constraint relationships. S5. Based on the refined features of the key points, the 3D coordinates and confidence of each key point are obtained by regression, and the human pose estimation result is output after time-series smoothing. S6, based on the preprocessed CSI data, extract the raw respiratory rate and heart rate data in parallel with steps S3 to S5; After step S5 is completed, the respiratory rate and heart rate data are optimized based on the exercise intensity corresponding to the human posture estimation results, and the vital signs monitoring results are output.

2. The method for non-contact monitoring of human posture and vital signs based on CSI according to claim 1, characterized in that, The preprocessing in step S2 specifically includes: performing linear calibration on the phase information of the original CSI data to eliminate hardware phase offset and carrier frequency offset, and completing phase unwrapping; based on the Fresnel zone electromagnetic wave propagation model, separating the static multipath component and dynamic multipath component in the CSI data, and extracting the dynamic component corresponding to human activity; and fusing amplitude, phase, phase difference, Doppler frequency shift and power spectral density information to construct a dynamic CSI feature vector.

3. The method for non-contact monitoring of human posture and vital signs based on CSI according to claim 1, characterized in that, In step S4, the human skeleton topology is a COCO standard 17 keypoint skeleton topology, which includes 17 nodes and 16 bone connection edges. The graph convolutional neural network uses two stacked graph convolutional layers, performs message passing through a symmetric normalized adjacency matrix, aggregates the feature information of adjacent nodes, and retains the initial feature information through residual connections.

4. The method for non-contact monitoring of human posture and vital signs based on CSI according to claim 1, characterized in that, Step S6 specifically includes: performing a 0.1-0.5Hz bandpass filter on the preprocessed CSI amplitude sequence, and obtaining the respiratory rate through Fast Fourier Transform (FFT) peak detection; performing phase unwrapping on the preprocessed CSI phase sequence, and then performing a 0.8-2.0Hz bandpass filter, selecting the subcarrier with the highest signal-to-noise ratio, and obtaining the heart rate through FFT peak detection; calculating the motion intensity based on the human posture estimation results, classifying the motion state based on the motion intensity, and performing filtering optimization or effectiveness determination on the respiratory rate and heart rate based on the motion state, while simultaneously completing the detection of abnormal vital signs.

5. The method for non-contact monitoring of human posture and vital signs based on CSI according to claim 1, characterized in that, It also includes environmental adaptation steps: Real-time monitoring of the difference between the current CSI feature distribution and the baseline environment feature distribution; triggering environment adaptation when the difference exceeds a preset threshold. By using a low-rank adaptive mechanism, the parameters of the inserted low-rank matrix are adjusted only while the weights of the backbone model are frozen. Combined with elastic weight reinforcement regularization to prevent catastrophic forgetting, the model parameters are updated through online gradient accumulation, enabling rapid adaptation to new environments.

6. A CSI-based non-contact human posture estimation and vital sign monitoring system, characterized in that, It includes a CSI data acquisition module, a signal preprocessing module, a cross-modal feature mapping module, a graph neural network inference module, and an attitude estimation output module that are connected in sequence, as well as a vital signs extraction module that is connected in communication with the signal preprocessing module and the attitude estimation output module. The CSI data acquisition module is used to acquire CSI data of the WiFi channel. The CSI data includes amplitude and phase information of the corresponding OFDM subcarriers for multiple sets of transmit-receive antenna pairs. The signal preprocessing module is used to preprocess the CSI data and output a dynamic CSI feature vector; The cross-modal feature mapping module is used to map the dynamic CSI feature vector to the human keypoint feature space through the Transformer cross attention mechanism, and output the initial features of each keypoint corresponding to the human skeleton topology map. The graph neural network inference module is used to perform spatial inference on the initial features of each key point based on the human skeleton topology map, and output the refined features of key points that integrate spatial constraint relationships. The pose estimation output module is used to refine the features of the key points, regress to obtain the 3D coordinates and confidence of each key point, and output the human pose estimation result after time-series smoothing. The vital signs extraction module is used to extract respiratory rate and heart rate data based on preprocessed CSI data through multi-band frequency domain analysis, optimize the respiratory rate and heart rate data by combining the exercise intensity corresponding to the human posture estimation results, and output vital signs monitoring results.

7. The CSI-based non-contact human posture estimation and vital sign monitoring system according to claim 6, characterized in that, It also includes an adaptive learning module, which is communicatively connected to the signal preprocessing module, the cross-modal feature mapping module, and the graph neural network inference module. The adaptive learning module is used to detect the drift of environmental feature distribution, complete the online environmental adaptation of the model through a low-rank adaptive mechanism, and prevent catastrophic forgetting of the model through elastic weight consolidation regularization.

8. The CSI-based non-contact human posture estimation and vital sign monitoring system according to claim 6, characterized in that, It also includes a multi-static fusion module, which is communicatively connected to the CSI data acquisition module and is used to fuse CSI data collected from multiple WiFi access points.

9. The CSI-based non-contact human posture estimation and vital sign monitoring system according to claim 6, characterized in that, The CSI data acquisition module uses a WiFi device that supports CSI data export, including at least one of the ESP32-S3 chip and Intel 5300 network card; all data processing and inference processes of the system are completed on the local edge device.

Images