Micro-motion three-dimensional perception method and system based on WiFi channel state information

Abstract

This invention discloses a micro-motion 3D perception method and system based on WiFi channel state information, relating to the field of wireless signal sensing technology. The steps are as follows: First, CSI data is collected and preprocessed to output the CSI amplitude matrix and phase matrix. Second, common noise is canceled using the phase difference of adjacent antenna pairs to construct a linear mapping model between displacement increment and phase change, achieving displacement calculation. Third, a spatiotemporal TransFormer model is designed, and through a joint temporal and spatial attention mechanism, CSI amplitude, phase difference, displacement increment, and Doppler frequency shift features are fused to output an initial 3D trajectory. Fourth, a self-supervised learning mechanism based on physical consistency constraints is used to optimize the initial 3D trajectory. Finally, a dynamic path filter is constructed, and extended Kalman filtering and Bayesian multipath fusion are combined to perform real-time filtering and correction on the optimized trajectory, outputting the final 3D trajectory. This achieves high-precision, low-latency trajectory reconstruction without annotation in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of wireless signal sensing technology, and in particular to a micro-motion three-dimensional sensing method and system based on WiFi channel state information. Background Technology

[0002] Wireless sensing technology, as a non-contact sensing paradigm, has developed in tandem with the development of wireless communication itself. Its core idea is to extract and analyze valuable physical world information from environmental channel disturbances, traditionally considered noise. The development of wireless sensing technology is currently at a critical crossroads. Existing technologies have failed to fundamentally solve the inherent defects of Channel State Information (CSI) signals, and are caught in a "triple dilemma" between accuracy, cost, and generalization ability. Current WiFi sensing technologies mainly fall into two categories: classification tasks and regression tasks. Classification tasks: These are relatively mature technologies, with implemented applications including fall detection and gesture recognition. The core is to extract statistical features of CSI amplitude / phase to train models such as CNNs and LSTMs to output discrete labels. However, CNN methods, due to the limitations of their local receptive field, struggle to capture long-distance temporal dependencies in signals and perform poorly when processing modulated signals with complex temporal characteristics. The task at hand is to focus on continuous trajectory reconstruction, but it is still in the laboratory research stage. Existing solutions mostly rely on customized hardware, such as multi-antenna arrays and high-sampling-rate oscilloscopes, or can only achieve centimeter-level accuracy, which is difficult to meet the needs of micro-motion monitoring.

[0003] Existing WiFi-CSI trajectory reconstruction solutions suffer from the following four types of significant technical flaws:

[0004] 1. Phase ambiguity leads to unreliable displacement calculations: CSI phase is affected by carrier frequency offset, sampling time offset, and hardware phase offset, and directly using it for displacement calculation will cause a "phase entanglement" problem. This makes the original phase information unusable for displacement calculation. Existing technologies use a "phase unrolling" algorithm to deal with this problem, but in complex multipath environments, due to noise interference, the error rate of phase unrolling is as high as 30%, resulting in displacement calculation deviations exceeding 1 cm.

[0005] 2. Multipath interference masks micro-motion characteristics: In indoor environments, WiFi signals are reflected by objects such as walls and furniture, forming multipath components. When these components are superimposed on the direct signal, the CSI amplitude / phase fluctuates drastically, causing micro-motion signals to be submerged by noise. Most current research uses "static multipath elimination" methods (such as background subtraction), which essentially assume the environment is static. They cannot handle dynamic multipath changes caused by people moving around or doors and windows opening and closing, leading to complete failure of trajectory reconstruction in real dynamic environments and drastic trajectory jumps.

[0006] 3. Deep models lack spatiotemporal fusion capabilities: Traditional deep learning models cannot simultaneously capture the long-range temporal dependence of CSI signals and spatial antenna interactions. CSI data is inherently spatiotemporally coupled, carrying information in both time and space dimensions. However, traditional model architectures have inherent flaws, making effective joint modeling difficult.

[0007] 4. High deployment costs due to reliance on labeled data: Existing trajectory reconstruction solutions mostly use supervised learning, which requires the collection of paired data through optical equipment, resulting in extremely high labeling costs.

[0008] Therefore, for those skilled in the art, how to overcome the above-mentioned defects and achieve sub-centimeter-level trajectory reconstruction under commercial WiFi hardware is an urgent problem to be solved. Summary of the Invention

[0009] The purpose of this invention is to provide a micro-motion 3D perception method and system based on WiFi channel state information. Through the deep integration of phase difference domain analysis, spatiotemporal Transformer modeling, self-supervised constraint learning and dynamic path filtering, it achieves high-precision and low-latency trajectory reconstruction without annotation in complex scenarios, providing reliable technical support for fields such as medical rehabilitation, smart homes, and metaverse interaction.

[0010] To achieve the above objectives, the present invention provides the following solution: On one hand, it provides a micro-motion 3D perception method based on WiFi channel state information, the specific steps of which include the following:

[0011] Collect CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix;

[0012] The phase difference between adjacent antenna pairs is obtained from the phase matrix. After denoising the phase difference, a linear mapping model between the displacement increment and the phase difference change is constructed, and the global displacement increment and Doppler frequency shift characteristics are output.

[0013] Design a spatiotemporal Transformer model, standardize the CSI amplitude matrix, phase difference, displacement increment and Doppler frequency shift features into input tensors, and output the initial three-dimensional trajectory through a temporal-spatial joint attention mechanism;

[0014] The initial 3D trajectory is optimized by using a self-supervised learning mechanism based on physical consistency constraints for unlabeled training.

[0015] A dynamic path filter is constructed, and the optimized trajectory is filtered and corrected in real time by combining extended Kalman filtering and Bayesian multipath fusion technology to output the final three-dimensional trajectory.

[0016] Preferably, the CSI data is collected at a sampling rate of not less than 1000Hz through a hardware platform consisting of a commercial WiFi network card that supports CSI extraction and an edge computing device.

[0017] Preferably, the specific steps for preprocessing the CSI data are as follows:

[0018] The acquired CSI data undergoes carrier frequency offset compensation and sampling time offset compensation, and the Dynamic Time Warping (DTW) algorithm is used to achieve time synchronization of multi-antenna data streams; the compensation formula is:

[0019] ;

[0020] in, The compensated phase, For the original phase, f_c = 5180MHz, This is the sampling time offset. For carrier frequency offset, This refers to hardware phase offset.

[0021] By adopting the above technical solution, the following beneficial technical effects are achieved: millimeter-level phase calculation under commercial WiFi hardware eliminates the need for customized hardware; through innovative signal processing algorithms, systematic errors such as CFO / STO are fundamentally offset, establishing a stable and accurate "phase-displacement" mapping model. This reduces hardware costs and makes high-precision sensing capabilities accessible to every ordinary WiFi device.

[0022] Preferably, the instantaneous phase difference between adjacent antenna pairs under the same subcarrier is obtained from the phase matrix, and the formula for calculating the phase difference is:

[0023] ;

[0024] in, Let be the phase difference between the i-th and j-th antennas at time t under the k-th subcarrier. , These represent the receiving phases of the two antennas, respectively.

[0025] The phase difference is smoothed using the subcarrier sliding window averaging method.

[0026] Preferably, the global displacement increment is calculated as follows:

[0027] Calculate the signal-to-noise ratio of each subcarrier-antenna pair element;

[0028] A signal-to-noise ratio (SNR) threshold is set, and units with SNR values ​​higher than the threshold are weighted and normalized. Finally, the global displacement increment is obtained by weighted summation.

[0029] Preferably, the spatiotemporal Transformer model adopts an encoder-decoder structure. The encoder sequentially includes a temporal window attention layer, a spatial antenna attention layer, and a multi-scale feature fusion layer. Residual connections and layer normalization ensure training stability. The decoder is a multilayer perceptron (MLP) that maps the high-dimensional features output by the encoder into a continuous sequence of three-dimensional spatial coordinates, i.e., the initial three-dimensional trajectory.

[0030] Preferably, the physical consistency constraints include phase conservation, trajectory smoothing, and Doppler matching; wherein, the phase conservation loss function is:

[0031] ;

[0032] in, Let i be the predicted displacement increment for the i-th frame. Let λ be the cumulative sum of predicted displacement increments from frame 1 to frame t, where λ is the wavelength of the WiFi signal. Let be the CSI phase measurement value of the t-th frame. The CSI phase measurement value for the first frame;

[0033] The trajectory smoothing loss function is:

[0034] ;

[0035] Where d_t is the predicted displacement increment of frame t, d{t-1} is the predicted displacement increment of frame t-1, and T is the total number of frames in a training batch.

[0036] The Doppler consistency loss function is:

[0037] ;

[0038] Where d_t is the displacement increment sequence, Dopper is the Doppler frequency shift sequence, and FFT() is the Fast Fourier Transform, which converts the time-domain signal into a frequency-domain signal.

[0039] The self-supervised learning mechanism is a weighted sum of the phase conservation loss function, the trajectory smoothing loss function, and the Doppler consistency loss function:

[0040] ;

[0041] By adopting the above technical solution, the following beneficial technical effects are achieved: a model training paradigm without labeled data; a self-supervised or weakly supervised learning framework that does not rely on optically labeled data; utilizing the physical laws of the signal itself (such as phase conservation and Doppler effect) as supervisory signals to achieve high-precision model training; significantly reducing deployment costs and barriers to entry, and eliminating dependence on expensive labeled data and cumbersome processes.

[0042] Preferably, a high-confidence observation path is selected from the direct path and multiple reflection paths through Bayesian inference, and the process noise and observation noise covariance of the extended Kalman filter are adaptively adjusted based on the CSI signal quality and path confidence to output the final smooth three-dimensional trajectory.

[0043] On the other hand, a micro-motion 3D perception system based on WiFi channel state information is provided, including a CSI acquisition and synchronization module, a phase difference domain displacement calculation module, an initial trajectory calculation module, a trajectory optimization module, and a correction module; wherein...

[0044] The CSI acquisition and synchronization module is used to acquire CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix.

[0045] The phase difference domain displacement calculation module is used to obtain the phase difference between adjacent antenna pairs from the phase matrix, denoise the phase difference, construct a linear mapping model between the displacement increment and the phase difference change, and output the global displacement increment and Doppler frequency shift characteristics.

[0046] The initial trajectory calculation module is used to design a spatiotemporal Transformer model, which standardizes the CSI amplitude, phase difference, displacement increment and Doppler frequency shift features into input tensors, and outputs the initial three-dimensional trajectory through a temporal-spatial joint attention mechanism.

[0047] The trajectory optimization module is used to optimize the initial three-dimensional trajectory by performing label-free training using a self-supervised learning mechanism based on physical consistency constraints.

[0048] The correction module is used to construct a dynamic path filter, which combines extended Kalman filtering and Bayesian multipath fusion technology to perform real-time filtering and correction on the optimized trajectory and output the final three-dimensional trajectory.

[0049] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] By leveraging Channel State Information (CSI) from commercial WiFi devices, sub-centimeter-level high-precision reconstruction of human 3D trajectories can be achieved. This technology can be applied to complex scenarios such as medical and health monitoring and smart home interaction. It overcomes the limitations of existing technologies, such as phase ambiguity, multipath interference, insufficient spatiotemporal fusion of models, and high annotation costs. Through technologies such as phase difference domain analysis, spatiotemporal Transformer modeling, self-supervised trajectory constraints, and dynamic path filtering, continuous trajectory reconstruction without additional sensors, with low power consumption and high robustness can be achieved. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2 This is a schematic diagram of the spatiotemporal Transformer model framework of the present invention;

[0054] Figure 3 This is a schematic diagram of the self-supervised trajectory constraint learning mechanism of the present invention;

[0055] Figure 4 This is a schematic diagram of the dynamic path filter (DPF) of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The purpose of this invention is to provide a micro-motion 3D perception method based on WiFi channel state information, such as... Figure 1 As shown, the specific steps include the following:

[0058] S1. Acquire CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix;

[0059] S2. Obtain the phase difference between adjacent antenna pairs from the phase matrix, denoise the phase difference, construct a linear mapping model between the displacement increment and the phase difference change, and output the global displacement increment and Doppler frequency shift characteristics.

[0060] S3. Design a spatiotemporal Transformer model, standardize the CSI amplitude matrix, phase difference, displacement increment and Doppler frequency shift features into input tensors, and output the initial three-dimensional trajectory through a temporal-spatial joint attention mechanism;

[0061] S4. Utilize a self-supervised learning mechanism based on physical consistency constraints to perform label-free training and optimize the initial 3D trajectory;

[0062] S5. Construct a dynamic path filter, and combine extended Kalman filtering and Bayesian multipath fusion technology to perform real-time filtering and correction on the optimized trajectory, and output the final three-dimensional trajectory.

[0063] Furthermore, in S1, a hardware platform consisting of a commercial WiFi network card supporting CSI extraction and an edge computing device is used to collect raw CSI data from multiple antennas at a sampling rate of no less than 1000Hz. Carrier frequency offset compensation and sampling time offset compensation are performed on the collected complex CSI data, and the Dynamic Time Warping (DTW) algorithm is used to achieve high-precision time synchronization of the multi-antenna data streams, outputting the synchronized CSI amplitude matrix and phase matrix. Specifically, S1 includes the following:

[0064] S11. Hardware and data acquisition specifications are as follows:

[0065] Transmitter (Tx): Commercial WiFi router, supporting 802.11ac / ax protocol, prioritizing 5GHz band (wavelength ≈ 6cm), subcarrier number K = 30-114.

[0066] Receiver (Rx): A network card that supports CSI extraction, with a sampling rate of 800-1200Hz and the number of antennas N=2-4. If AoA measurement is required, a multi-antenna array is used.

[0067] Sampling configuration: Each frame of CSI data is a complex matrix (subcarrier × antenna), and the timestamp is synchronized via NTP / PTP, with hardware triggering to achieve alignment of <1ms.

[0068] Data format: CSI data is saved in .mat or .npy format, and metadata (timestamps, frame identifiers, etc.) is recorded in JSON.

[0069] S12, CFO / STO Compensation and Noise Filtering: Carrier Frequency Offset (CFO) and Sampling Time Offset (STO) are estimated using pilot subcarriers. The compensation formula is as follows:

[0070] ;

[0071] in, For the original phase, f_c = 5180MHz, This is the sampling time offset. For carrier frequency offset, This addresses hardware phase shift. After compensation, a 5th-order Butterworth low-pass filter (cutoff frequency 20Hz) is used to perform zero-phase filtering on the CSI amplitude and phase to suppress high-frequency noise. Finally, the DTW algorithm is used to align the timing of multiple antennas to ensure spatiotemporal consistency.

[0072] In this embodiment of the invention, the hardware and data acquisition specifications are as follows: The transmitting end uses a commercial WiFi router (such as TP-Link AX3000) that supports the 802.11ax protocol, preferentially using the 5GHz band (center frequency 5180MHz), with a bandwidth of 40 / 80MHz and 114 subcarriers. The receiving end uses an Intel AX210 network card, supporting 3×3 MIMO antenna configuration, and extracts raw CSI data through a Linux kernel patch (linux-80211n-csitool).

[0073] The sampling configuration is as follows: the sampling rate is set to 1000Hz, and each frame of CSI data is a complex matrix (114 subcarriers × 3 antennas). The timestamp is synchronized via NTP / PTP protocol, and external GPIO hard triggering achieves alignment accuracy of <1ms.

[0074] The original data format is as follows: Single-frame CSI: complex matrix (114,3), representing the number of subcarriers × the number of antennas; Window data: real tensor (T,114,3,2), where T is the frame number and the last dimension represents the amplitude and phase.

[0075] Metadata (JSON): {timestamp, frame_id, location_fix}, where timestamp (microsecond-level timestamp), frame_id (frame identifier), and location_fix (location assistance information).

[0076] Furthermore, a phase difference domain displacement calculation module is constructed in S2. From the synchronized phase matrix, the instantaneous phase difference between adjacent antenna pairs under the same subcarrier is obtained. A sliding window averaging method is used to smooth and denoise the phase difference along the subcarrier dimension. Based on the electromagnetic wave propagation physical model, a linear mapping relationship is established between the phase difference change and the human body displacement increment. The displacement increment of each antenna-subcarrier pair is calculated by combining the WiFi signal wavelength. Weighted fusion is then performed based on the real-time signal-to-noise ratio (SNR) of each subcarrier and antenna pair to output the global displacement increment sequence and Doppler frequency shift characteristics. Specifically, S2 includes the following:

[0077] S21. Antenna Pair Selection and Phase Difference Calculation: For a receiver array with N antennas, prioritize the adjacent antenna pairs with the smallest physical spacing. Calculate the phase difference of each antenna pair on the same subcarrier k using the following formula:

[0078] ;

[0079] in, Let be the phase difference between the i-th and j-th antennas at time t under the k-th subcarrier. , These represent the receiving phases of the two antennas, respectively.

[0080] S22. Use subcarrier moving average (window size = 5) to smooth the phase difference:

[0081] ;

[0082] in, Let (i, j) be the smoothed phase difference of antenna pair (i, j) on the k-th subcarrier at time t. At time t, the original phase difference of antenna pair (i, j) on the m-th subcarrier.

[0083] S23. Displacement Increment Calculation and Weighted Average: Based on the principle of electromagnetic wave propagation, a linear relationship is established between the change in phase difference and the displacement increment.

[0084] ;

[0085] in, Let be the displacement increment calculated at time t based on the k-th subcarrier and antenna pair (i, j). - Let λ represent the change in phase difference, and λ be the wavelength of the WiFi signal, which in this embodiment is equal to 0.0579m (5GHz WiFi wavelength). Outliers (|Δd|>1cm) are replaced.

[0086] S24. Weighted Fusion: Calculate the SNR of each subcarrier-antenna pair element, set an SNR threshold (10dB), assign weights to elements with SNR values ​​higher than the threshold according to their SNR values ​​and normalize them, and finally obtain the global displacement increment Δd(t) by weighted summation:

[0087] Calculate weights based on SNR:

[0088] ;

[0089] in, Let be the signal-to-noise ratio of the antenna pair (i, j) and the k-th subcarrier at time t. For time point t, the normalized weights for antenna pair (i, j) and the k-th subcarrier are... The sum of the SNR of all valid antenna pairs and subcarriers at the current time t (the weight is set to 0 when SNR < 10dB).

[0090] The global displacement increment is:

[0091] ;

[0092] in, Let be the global displacement increment at time t. For time point t, the normalized weights for antenna pair (i, j) and the k-th subcarrier are... Let be the individual displacement increment at time t, based on antenna pair (i, j) and the k-th subcarrier;

[0093] The cumulative displacement increment is:

[0094] .

[0095] Where d(t) is the cumulative displacement at time t. Let n be the global displacement increment at time point n.

[0096] Furthermore, such as Figure 2 As shown, the network architecture of the spatiotemporal Transformer model in S3 is as follows:

[0097] Temporal window attention layer: window size = 20 frames, capturing local micro-movements;

[0098] Spatial Antenna Attention Layer: Constructs spatial embedding on an antenna pair basis and models multi-antenna interactions;

[0099] Multi-scale feature fusion layer: It adopts 2 / 4 / 8 times subcarrier downsampling, with convolution kernels of 2, 4 and 8 respectively, and stride corresponding to each other. Then, it is upsampled back to the original scale by bilinear interpolation, and features are fused by 1×1 convolution.

[0100] The output layer is a fully connected decoder that regresses the 3D coordinates [x, y, z] of each frame. Training hyperparameters: learning rate 1e-4, batch size=32, AdamW optimizer, StepLR decay (step=20, gamma=0.5).

[0101] The specific operation process of S3 is as follows:

[0102] S31. Input Feature Engineering: Extract four-dimensional features: CSI amplitude, phase difference, displacement increment, and Doppler frequency shift. Doppler frequency shift is calculated using phase center difference.

[0103] ;

[0104] in, Let be the phase value at time point t+1, on the k-th subcarrier, and on the i-th antenna. Let be the phase value at time point t-1, on the k-th subcarrier, and on the i-th antenna.

[0105] S32. Perform Z-score normalization on the four-dimensional features to construct the input tensor [T×114×3×4] (T=512 frames).

[0106] Furthermore, a self-supervised trajectory constraint learning mechanism based on physical consistency is designed in S4, such as... Figure 3 As shown, without the need for optical annotation data, the initial trajectory output in step S3 is optimized by constructing and jointly optimizing three physical driving loss functions—phase conservation loss, trajectory smoothing loss, and Doppler consistency loss—to ensure that the predicted trajectory conforms to the basic physical laws of wireless signal propagation and object motion.

[0107] The phase conservation loss function is:

[0108] ;

[0109] in, Let i be the predicted displacement increment for the i-th frame. Let λ be the cumulative sum of predicted displacement increments from frame 1 to frame t, where λ is the wavelength of the WiFi signal. Let be the CSI phase measurement value of the t-th frame. The CSI phase measurement value for the first frame;

[0110] The trajectory smoothing loss function is:

[0111] ;

[0112] Where d_t is the predicted displacement increment of frame t, d{t-1} is the predicted displacement increment of frame t-1, and T is the total number of frames in a training batch.

[0113] The Doppler consistency loss function is:

[0114] ;

[0115] Where d_t is the displacement increment sequence, Dopper is the Doppler frequency shift sequence, and FFT() is the Fast Fourier Transform, which converts the time-domain signal into a frequency-domain signal.

[0116] The self-supervised learning mechanism is a weighted sum of the phase conservation loss function, the trajectory smoothing loss function, and the Doppler consistency loss function:

[0117] .

[0118] Training strategy: It is divided into a pre-training stage (no label, 50 epochs, only L_phase + L_smooth) and a fine-tuning stage (adding a small amount of labeled data, 30 epochs, full loss), combined with early stopping (Patience=10) and data augmentation (temporal pruning, noise injection, phase shift).

[0119] Furthermore, S5 integrates Extended Kalman Filtering (EKF) with Bayesian multipath fusion theory. Through Bayesian inference, it filters high-confidence observation paths from direct and multiple reflection paths. Based on CSI signal quality and path confidence, it adaptively adjusts the process noise and observation noise covariance of the EKF, outputting the final smooth three-dimensional trajectory. Specifically, as... Figure 4 As shown, S5 includes the following:

[0120] S51, The state vector of the Extended Kalman Filter (EKF) is defined as follows:

[0121] ;

[0122] The state transition matrix A is constructed based on a uniform motion model (sampling interval Δt = 1ms).

[0123] S52, Bayesian multipath fusion: The path is divided into direct path (LOS), reflection path 1 (NLOS1), and reflection path 2 (NLOS2), with prior probabilities of 0.6, 0.3, and 0.1, respectively. The posterior probability is calculated using the Gaussian likelihood function, and the path with a posterior probability > 0.7 is selected as the effective observation.

[0124] S53, Noise Covariance Adaptation: The process noise covariance Q is negatively correlated with SNR, and the observation noise covariance R is negatively correlated with the effective path posterior probability P_max.

[0125] S54. The filtered output is the final 3D trajectory with an end-to-end delay of <40ms, meeting the requirements for real-time interaction.

[0126] S55. Trajectory Display and Output: The system displays 3D trajectories in real time using the Open3D library, supporting trajectory playback and key point annotation. It also features trajectory export functionality (format: CSV / PLY) for easy medical data analysis or interactive command recognition.

[0127] On the other hand, a micro-motion 3D perception system based on WiFi channel state information is provided, including a CSI acquisition and synchronization module, a phase difference domain displacement calculation module, an initial trajectory calculation module, a trajectory optimization module, and a correction module; wherein...

[0128] The CSI acquisition and synchronization module is used to acquire CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix.

[0129] The phase difference domain displacement calculation module is used to obtain the phase difference between adjacent antenna pairs from the phase matrix, and after denoising the phase difference, it constructs a linear mapping model between the displacement increment and the phase difference change, and outputs the global displacement increment and Doppler frequency shift characteristics.

[0130] The initial trajectory calculation module is used to design the spatiotemporal Transformer model. It normalizes the CSI amplitude, phase difference, displacement increment and Doppler frequency shift features into input tensors and outputs the initial three-dimensional trajectory through a temporal-spatial joint attention mechanism.

[0131] The trajectory optimization module is used to optimize the initial 3D trajectory by performing label-free training using a self-supervised learning mechanism based on physical consistency constraints.

[0132] The correction module is used to construct a dynamic path filter, which combines extended Kalman filtering and Bayesian multipath fusion technology to perform real-time filtering and correction on the optimized trajectory and output the final three-dimensional trajectory.

[0133] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention.

[0134] In conclusion, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A micro-motion three-dimensional perception method based on WiFi channel state information, characterized in that, The specific steps include the following: Collect CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix; The phase difference between adjacent antenna pairs is obtained from the phase matrix. After denoising the phase difference, a linear mapping model between the displacement increment and the phase difference change is constructed, and the global displacement increment is output. The global displacement increment specifically includes: establishing a linear mapping relationship between the phase difference change and the individual displacement increment based on the electromagnetic wave propagation physical model; calculating the displacement increment of each antenna-subcarrier pair by combining the WiFi signal wavelength; and performing weighted fusion based on the real-time signal-to-noise ratio of each subcarrier and antenna pair to output the global displacement increment. A spatiotemporal Transformer model is designed, which normalizes the CSI amplitude matrix, phase difference, displacement increment, and Doppler frequency shift features into input tensors. An initial three-dimensional trajectory is output through a temporal-spatial joint attention mechanism. The Doppler frequency shift features are calculated by differentially analyzing the phase center of each pair of antennas and subcarriers at different time points. The initial 3D trajectory is optimized by using a self-supervised learning mechanism based on physical consistency constraints for unlabeled training. A dynamic path filter is constructed, and the optimized trajectory is filtered and corrected in real time by combining extended Kalman filtering and Bayesian multipath fusion technology, and the final three-dimensional trajectory is output.

2. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The CSI data is collected at a sampling rate of no less than 1000Hz using a hardware platform consisting of a commercial WiFi network card that supports CSI extraction and an edge computing device.

3. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The specific steps for preprocessing the CSI data are as follows: The acquired CSI data undergoes carrier frequency offset compensation and sampling time offset compensation, and the Dynamic Time Warping (DTW) algorithm is used to achieve time synchronization of multi-antenna data streams; the compensation formula is: ； in, The compensated phase, For the original phase, , This is the sampling time offset. Here, t represents the carrier frequency offset, t is the sampling time in the time dimension, and k is the subcarrier index, identifying a specific subcarrier number in the Wi-Fi signal. This refers to hardware phase offset.

4. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The instantaneous phase difference between adjacent antenna pairs under the same subcarrier is obtained from the phase matrix. The formula for calculating the phase difference is: ； in, Let be the phase difference between the i-th and j-th antennas at time t under the k-th subcarrier. , These represent the receiving phases of the two antennas, respectively. The phase difference is smoothed using the subcarrier sliding window averaging method.

5. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The global displacement increment is calculated as follows: Calculate the signal-to-noise ratio of each subcarrier-antenna pair element; Set a signal-to-noise ratio threshold, and assign weights to cells that exceed the threshold according to their signal-to-noise ratio values ​​and normalize them; Finally, the global displacement increment is obtained by weighted summation of the individual displacement increments of each subcarrier-antenna pair element according to their respective weights.

6. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The spatiotemporal Transformer model adopts an encoder-decoder structure. The encoder sequentially includes a temporal window attention layer, a spatial antenna attention layer, and a multi-scale feature fusion layer. Residual connections and layer normalization ensure training stability. The decoder is a multilayer perceptron (MLP) that maps the high-dimensional features output by the encoder into a continuous sequence of three-dimensional spatial coordinates, i.e., the initial three-dimensional trajectory.

7. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, The physical consistency constraints include phase conservation, trajectory smoothing, and Doppler matching; wherein, the phase conservation loss function is: ； in, Let i be the predicted displacement increment for the i-th frame. This is the cumulative sum of the predicted displacement increments from frame 1 to frame t. The wavelength of the WiFi signal. Let be the CSI phase measurement value of the t-th frame. The CSI phase measurement value for the first frame; The trajectory smoothing loss function is: ； in, Let be the predicted displacement increment for frame t. for The predicted displacement increment of a frame, where T is the total number of frames in a training batch; The Doppler consistency loss function is: ； in, For displacement increment sequences, Dopper is a Doppler frequency shift sequence. Fast Fourier Transform converts the time-domain signal into a frequency-domain signal; The self-supervised learning mechanism is a weighted sum of the phase conservation loss function, the trajectory smoothing loss function, and the Doppler consistency loss function: 。 8. The micro-motion three-dimensional perception method based on WiFi channel state information according to claim 1, characterized in that, High-confidence observation paths are selected from direct and multiple reflection paths using Bayesian inference. The process noise and observation noise covariance of the extended Kalman filter are adaptively adjusted based on CSI signal quality and path confidence, and the final smooth three-dimensional trajectory is output.

9. A micro-motion three-dimensional sensing system based on WiFi channel state information, characterized in that, It includes a CSI acquisition and synchronization module, a phase difference domain displacement calculation module, an initial trajectory calculation module, a trajectory optimization module, and a correction module; among which, The CSI acquisition and synchronization module is used to acquire CSI data, perform preprocessing operations on the CSI data, and output the CSI amplitude matrix and phase matrix. The phase difference domain displacement calculation module is used to obtain the phase difference between adjacent antenna pairs from the phase matrix, denoise the phase difference, construct a linear mapping model between the displacement increment and the phase difference change, and output the global displacement increment. The global displacement increment specifically includes: establishing a linear mapping relationship between the phase difference change and the individual displacement increment based on the electromagnetic wave propagation physical model, calculating the displacement increment of each antenna-subcarrier pair in combination with the WiFi signal wavelength, and performing weighted fusion based on the real-time signal-to-noise ratio of each subcarrier and antenna pair to output the global displacement increment. The initial trajectory calculation module is used to design a spatiotemporal Transformer model, which standardizes the CSI amplitude, phase difference, displacement increment, and Doppler frequency shift features into input tensors, and outputs the initial three-dimensional trajectory through a temporal-spatial joint attention mechanism; the Doppler frequency shift features are obtained by calculating the phase center difference of each pair of antennas and subcarriers at different time points. The trajectory optimization module is used to optimize the initial three-dimensional trajectory by performing label-free training using a self-supervised learning mechanism based on physical consistency constraints. The correction module is used to construct a dynamic path filter, which combines extended Kalman filtering and Bayesian multipath fusion technology to perform real-time filtering and correction on the optimized trajectory and output the final three-dimensional trajectory.

Images