Long-time passive human body tracking method and system based on commercial Wi-Fi equipment

Abstract

The invention discloses a long-time passive human body tracking method and system based on commercial Wi-Fi equipment, and relates to the technical field of Wi-Fi passive tracking in an indoor environment. According to the method, a theoretical model method is taken as a theoretical basis, a neural network is taken as a fitting tool, and a cross-link behavior of a user is taken as an opportunity, so that inaccurate initial position estimation is eliminated. In the implementation process, an analogue simulation data set is generated, the neural network is trained by using the data set, and meanwhile, kinematics constraints and user cross-link physical restrictions are embedded into the neural network. In addition, the invention further provides a confidence-based method, and the current track is dynamically corrected by using historical track information, so that the tracking accuracy and reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of passive Wi-Fi tracking technology in indoor environments, and particularly to a method and system for long-term passive human body tracking based on commercial Wi-Fi devices. Background Technology

[0002] With the rapid development of IoT technology, the informatization and networking levels of various industries have significantly improved. Smart devices and their applications have deeply integrated into people's daily lives, becoming an indispensable part. In the home setting, the number of connected devices continues to grow. These devices, through interconnection, provide strong support for home automation and intelligence, driving the smart home field into an unprecedented stage of transformation. Compared to traditional device sensing methods, wireless sensing technology stands out with its unique advantages. This technology senses the presence and behavior of users by analyzing wireless signals in the environment, providing users with a natural and unrestricted human-computer interaction experience. Among many wireless communication technologies, Wi-Fi occupies a key position in the smart home field due to its low cost, convenient deployment, and wide coverage. It can not only realize traditional data communication but also sense the surrounding physical environment based on the transmission and reflection laws of signals, providing more possibilities for the intelligent development of smart homes.

[0003] Wi-Fi passive tracking technology primarily relies on analyzing the wireless signals reflected from a target. By deeply analyzing Channel State Information (CSI), various key parameters can be extracted. These parameters can be interpreted from either the time or frequency domain, including the Angle of Arrival (AoA), Time of Flight (ToF), and Doppler Frequency Shift (DFS). However, in practical applications, commercial Wi-Fi devices are often affected by hardware precision limitations and wireless signal characteristics. For example, initial phase shift errors often make AoA measurement extremely complex; while sampling time offset and packet detection delays can significantly interfere with the accuracy of ToF measurements. Faced with numerous challenges in the commercial Wi-Fi environment, such as hardware precision limitations and signal interference, DFS, with its advantages of accuracy and ease of acquisition, has gradually become the preferred signal feature for tracking technology. Theoretically, Doppler frequency shift directly reflects the rate of change of the reflection path length, a physical characteristic that makes it more adaptable and feasible in practical applications.

[0004] For example, the Widar positioning system proposed in 2017 and the fewSens positioning system proposed in 2024 both achieve human body tracking by extracting Direct-Speed ​​Flow (DFS) from the received signal. The basic principle of this method is to extract the frequency changes of the received signal to obtain the user's velocity, and then integrate the velocity based on the initial position to obtain the user's trajectory. However, this Doppler velocity tracking method suffers from the problem of accumulated error, which significantly limits the practical application of Wi-Fi-based tracking systems and their sustained effectiveness in real-world scenarios. We conducted preliminary experiments to verify this view, arranging volunteers to walk six laps along a square path in a real-world environment and using this model-based tracking method for user tracking. The results show that relatively accurate trajectories can be obtained in the initial few laps, but the error accumulates over time, reaching several meters in the later stages, causing the tracking system to essentially fail. Furthermore, erroneous signals caused when users cross device links also lead to incorrect trajectory predictions. To address these issues, this invention proposes a long-term passive human body tracking method and system based on commercial Wi-Fi devices. Summary of the Invention

[0005] The purpose of this invention is to propose a long-term passive human body tracking method and system based on commercial Wi-Fi devices to solve the problems mentioned in the background art. This invention eliminates the error accumulated by velocity integration and the trajectory error caused by the user crossing the link by periodically updating the initial position, and predicts a more accurate and reliable trajectory by learning motion constraints and cross-link constraints through neural networks, which greatly shortens the running time of the original passive positioning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] Long-term passive human body tracking methods based on commercial Wi-Fi devices include:

[0008] S1. Feature Extraction:

[0009] S1.1 Eliminate time asynchrony and random offset caused by hardware devices by using CSI ratio; use short time Fourier transform (STFT) technology to process CSI signal, extract frequency offset DFS caused by path length change, and then calculate user speed information;

[0010] S1.2. To address the phase change caused by signal abrupt changes, CI is used to detect the cross-link behavior of the target; the C-STD method for correcting phase jumps is applied to resolve the potential interference of phase abrupt changes on detection accuracy.

[0011] S2. Tracking and initial position positioning:

[0012] S2.1 Tracking: Using a knowledge-driven approach, a large-scale neural network training dataset is generated by simulating real human movement trajectories;

[0013] Considering the limitations of human kinematics, the simulation principle of "acceleration-velocity-position sequence" is used to simulate all walking trajectories of users in the positioning environment, so as to achieve the goal of covering the global positioning area;

[0014] The initial position and CI are used as input features to further optimize the prediction model of the neural network;

[0015] S2.2 Initial Position Localization: The neural network is combined with the theoretical model, and the initial position is obtained by reverse engineering using the analytical capability of the model; the user trajectory is obtained by using all candidate points within the feasible region, and the corresponding PLCR and CI are obtained by using the Fresnel zone theory and the reverse engineering of the model. The optimal initial position is obtained through feature mapping.

[0016] S3, Trajectory Correction:

[0017] A method that integrates current data and historical data is proposed to optimize trajectory correction. A confidence evaluation mechanism is constructed, and a model-based fitter is designed to evaluate the reliability of each historical data segment and assign corresponding weights to it according to its confidence.

[0018] By aggregating data trajectories with confidence levels higher than the current data segment, a more accurate and reliable trajectory is constructed.

[0019] Preferably, the formula for the operation described in S1.1 is expressed as follows:

[0020]

[0021] In the formula, f D λ represents frequency offset (DFS); λ represents wavelength; and r represents path length change rate (PLCR).

[0022] Preferably, the formula for the operation described in S1.2 is expressed as follows:

[0023]

[0024] In the formula, N represents the total number of data sampling points; n represents the nth sampling point; e represents the base of the natural logarithm; j represents the imaginary unit; x n This represents the phase value of the nth sampling point; This represents the phase mean of all sampled data points.

[0025] Preferably, human kinematics constraints are incorporated into the simulation of the walking trajectory, and the formula for these human kinematic constraints is expressed as:

[0026]

[0027] In the formula, These represent the acceleration and velocity of the human body during the simulation, respectively.

[0028] The generated simulation trajectory is constrained by the feasible region boundary, and its formula is expressed as:

[0029]

[0030] In the formula, x and y represent the x and y coordinates of a location point in the user's location sequence, respectively; x lower ,y lower Let x represent the minimum values ​​of the x and y coordinates within the feasible region, respectively; upper ,y upper These represent the maximum values ​​of the horizontal and vertical coordinates within the feasible region, respectively.

[0031] Preferably, the simulation process described in S2.1 is repeated to obtain multiple simulation trajectories as the training dataset for the subsequent neural network; features are extracted from the obtained simulation trajectories to obtain the corresponding position sequences P. d (i) with PLCR sequence R d (i) and CI sequence CI d (i), where d represents the d-th simulation trajectory and i is the i-th position point of a single trajectory;

[0032] Define the feature set and the label set:

[0033] Set feature =[R d (i),CI d (i),P init (i)]

[0034] Set label =[P(i,j)]

[0035] In the formula, P init (i) represents the initial position point; P(i,j) represents the set of position points on the trajectory;

[0036] Assuming there are N receivers and the output trajectory length is T, the input of the neural network is a (2N+2)×T dimensional matrix; the first N columns represent PLCR, the (N+1)th to 2Nth columns represent CI, and the last two columns represent the x and y coordinates of the initial position.

[0037] Preferably, S3 specifically includes the following:

[0038] Assuming all data are D(i), i∈(0,n), and the current data is D(i'), i'∈(m,n), T is obtained based on network prediction. current Assuming there are five historical windows, each with a length of winlen and a window step size of l, the historical trajectories obtained from the historical data are T. his (0, winlen), T his (l,winlen+l),T his (2*l, winlen+2*l), T his (3*l, winlen+3*l), T his (4*l, winlen+4*l); where winlen+4*l=m;

[0039] A model-based fitter is used to solve for the confidence scores of the five historical trajectories and the current trajectory. The inverse process of the model method is applied to obtain the corresponding PLCR features of the five historical window trajectories and the current trajectory. The confidence score of each trajectory is obtained by comparing the measured and observed PLCR values. The smaller the error between the observed and calculated PLCR, the higher the confidence score of the trajectory. Data with confidence scores lower than the current trajectory in the historical trajectories are discarded to avoid introducing additional errors. The remaining trajectories are weighted according to their confidence scores, and the final trajectory is obtained by summing the weights. The specific formula is as follows:

[0040]

[0041] In the formula, ω k The weight of each trajectory segment is calculated based on the confidence level of the trajectory; T k This represents the k-th trajectory.

[0042] The long-term passive human body tracking system based on commercial Wi-Fi devices, applicable to the aforementioned long-term passive human body tracking method based on commercial Wi-Fi devices, includes a feature extraction module, a tracking and initial position localization module, and a trajectory correction module, wherein:

[0043] Feature extraction module:

[0044] The CSI signal is processed to eliminate time asynchrony and random offset caused by hardware devices; the CSI signal is processed using short-time Fourier transform (STFT) technology to extract the frequency offset (DFS) caused by changes in path length, and then the user's speed information is calculated.

[0045] To address phase changes caused by signal abrupt changes, cross-link behavior of the target is detected using CI; the C-STD method for correcting phase jumps is applied to resolve the potential interference of phase abrupt changes on detection accuracy.

[0046] Tracking and initial position positioning module:

[0047] Tracking submodule: Employs a knowledge-driven approach to generate a large-scale neural network training dataset by simulating real human movement trajectories;

[0048] Considering the limitations of human kinematics, the simulation principle of "acceleration-velocity-position sequence" is used to simulate all walking trajectories of users in the positioning environment, so as to achieve the goal of covering the global positioning area;

[0049] The initial position and CI are used as input features to further optimize the prediction model of the neural network;

[0050] Initial position localization submodule: Combines neural network with theoretical model, uses the analytical capability of model to back-calculate initial position; uses all candidate points within feasible region to obtain user trajectory, combines Fresnel zone theory, uses back-calculation of model to obtain corresponding PLCR and CI, and obtains optimal initial position through feature mapping;

[0051] Trajectory correction module:

[0052] A method that integrates current data and historical data is proposed to optimize trajectory correction. A confidence evaluation mechanism is constructed, and a model-based fitter is designed to evaluate the reliability of each historical data segment and assign corresponding weights to it according to its confidence.

[0053] By aggregating data trajectories with confidence levels higher than the current data segment, a more accurate and reliable trajectory is constructed.

[0054] The present invention further protects a computer device, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the instruction, program, code set or instruction set being loaded and executed by the processor to implement the long-term passive human body tracking method based on commercial Wi-Fi devices as described above.

[0055] The present invention further protects a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the instruction, program, code set, or instruction set is loaded and executed by a processor to implement the above-described long-term passive human body tracking method based on commercial Wi-Fi devices.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) Compared with the traditional DFS-based tracking method, which accumulates errors over a long period of time, the method of periodically updating the initial position proposed in this invention can eliminate the accumulated errors caused by time accumulation and can also eliminate erroneous trajectories caused by cross-links. In addition, the method of predicting trajectories through neural networks can also embed human kinematic constraints, which can shorten the trajectory prediction time while ensuring the reliability and stability of the trajectory.

[0058] (2) The present invention also proposes a method for correcting the trajectory based on historical information. By jointly utilizing historical trajectory data and designing a confidence index, more accurate historical information can be used to optimize the current tracking trajectory, especially when the current trajectory is subject to large noise interference, so that even a lightweight network architecture can achieve accurate tracking. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the embodiments are now briefly described. Obviously, the drawings in the following description are merely illustrative of some embodiments of the present invention. For those skilled in the art, other forms of drawings can be constructed based on these drawings without creative effort.

[0060] Figure 1 This is a flowchart of the long-term passive human body tracking method based on commercial Wi-Fi devices mentioned in Embodiment 1 of the present invention;

[0061] Figure 2 This is a schematic diagram of the Fresnel zone cut by human movement mentioned in Embodiment 2 of the present invention;

[0062] Figure 3 This is a schematic diagram illustrating the determination of speed uniqueness across multiple links as mentioned in Embodiment 2 of the present invention;

[0063] Figure 4 This is a diagram illustrating the passive positioning effect of the method mentioned in Embodiment 2 of the present invention. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0065] First, the technical terms mentioned in this invention will be explained and clarified:

[0066] (1) Wi-Fi passive tracking is an innovative method that uses Wi-Fi signals to determine a user's location. Unlike traditional active positioning methods, it does not require users to carry intrusive devices such as mobile phones or sensors. The core of this technology lies in the in-depth analysis of Wi-Fi signals reflected by the human body to accurately complete the positioning task.

[0067] (2) Channel state information describes the propagation process of wireless signals between the transmitter and the receiver. This information comprehensively reflects the combined effects of various factors such as distance attenuation, power loss, and multipath scattering that the signal is subjected to during transmission, and includes key parameters such as the amplitude, phase, and frequency of the signal.

[0068] (3) Recurrent neural networks are a type of recurrent neural network that processes sequence information. They introduce a time dimension on the basis of ordinary neural network models, and consider the connection of information between nodes at different time periods, thus having the ability to handle sequence problems of variable length.

[0069] (4) Fresnel zones describe multiple ellipsoidal regions formed around the transmitter and receiver as foci during the propagation of electromagnetic wave signals. The boundaries of these elliptical regions are defined by specific differences in the length of the signal propagation path. As the ordinal number of the Fresnel zone increases, the distance from each point on the ellipsoidal boundary to the two foci gradually increases. When a user moves through these Fresnel zones, the user's velocity will generate an effective velocity component in the normal direction of the Fresnel zone, resulting in a corresponding change in the characteristics of the observed signal. This is the theoretical basis upon which this positioning system relies.

[0070] (5) The Path Length Change Rate (PLCR) describes the rate of change of the signal path length after reflection from the human body. It reflects the dynamic changes in the path and is the cause of Doppler frequency shift. Combining Fresnel zone theory, the normal velocity for a specific link can be obtained by analyzing the PLCR. By synthesizing the normal velocities of multiple links, the user's true velocity magnitude and direction can be obtained.

[0071] (6) The Cross Indicator (CI) describes the signal changes caused by the human body blocking dynamically reflected signals and direct path signals when crossing the equipment link formed by the transmitter and receiver. It is the cause of signal phase abrupt changes. By analyzing this phase change, it is possible to determine when a user has engaged in cross-link behavior, thereby limiting the user's position to that link at that moment.

[0072] This invention addresses the limitations of existing theoretical model methods by proposing a periodic automatic re-estimation of the initial position. This method fundamentally eliminates the dependence on the initial position, effectively solving the problem of error accumulation. Specifically, this invention uses theoretical model methods as its foundation, leverages neural networks as a fitting tool, and considers the user's cross-link behavior as an opportunity to eliminate inaccurate initial position estimates. In implementation, this invention generates a simulation dataset and uses this dataset to train the neural network, while embedding kinematic constraints and the physical limitations of the user's cross-link behavior into the neural network. Furthermore, this invention proposes a confidence-based method that dynamically corrects the current trajectory using historical trajectory information, thereby significantly improving the accuracy and reliability of tracking.

[0073] The long-term passive human body tracking method based on commercial Wi-Fi devices proposed in this invention will be described below with reference to the relevant accompanying drawings and specific examples.

[0074] Example 1:

[0075] This invention proposes a long-term passive human body tracking method based on commercial Wi-Fi devices. Please refer to [link / reference]. Figure 1 This invention mainly includes three stages: feature extraction, tracking and initial position localization, and trajectory correction, as detailed below:

[0076] (1) Feature extraction stage:

[0077] The goal of this stage is to extract core features from the raw wireless signal that accurately reflect the target's motion state. The specific steps are as follows: First, time asynchrony and random offsets caused by hardware devices are eliminated using CSI ratio. Then, STFT technology is used to process the CSI signal to extract the frequency offset (DFS) caused by changes in path length. The user's speed information is then calculated. Subsequently, for phase changes caused by signal abrupt changes, the target's cross-link behavior is detected using CI. To further improve detection accuracy, the C-STD method for correcting phase jumps is applied, effectively solving the potential interference of phase jumps on detection accuracy.

[0078] (2) Tracking and initial location positioning stage:

[0079] Tracking: This phase proposes a novel knowledge- and data-driven dual-engine architecture. Specifically, a knowledge-driven approach is first employed to generate a large-scale neural network training dataset by simulating real human motion trajectories. Simultaneously, considering the limitations of human kinematics, the simulation principle of "acceleration-velocity-position sequence" is used to simulate the walking trajectories that users may exhibit in the positioning environment, achieving the goal of covering the global positioning area. Based on this, the initial position and CI are used as input features to further optimize the neural network's prediction model.

[0080] Initial Position Localization: To eliminate the system's dependence on the initial position, this invention combines neural networks with theoretical models, utilizing the model's analytical capabilities to inversely derive the initial position. Specifically, the user trajectory obtained from all candidate points within the feasible region is used, combined with Fresnel zone theory, to obtain the corresponding PLCR and CI through inverse modeling. The optimal initial position is then obtained through feature mapping.

[0081] (3) Trajectory Correction Stage: This stage employs a method that integrates current and historical data to optimize trajectory correction. This invention proposes a confidence assessment mechanism. Specifically, a model-based fitter is proposed to evaluate the reliability of each historical data segment and assign corresponding weights based on its confidence level. Finally, by aggregating data trajectories with confidence levels higher than the current data segment, a more accurate and reliable trajectory is constructed.

[0082] Example 2:

[0083] Based on Embodiment 1, but with some differences, the long-term passive human body tracking method based on commercial Wi-Fi devices proposed in this invention will be described below in conjunction with relevant accompanying drawings and the best implementation method. The specific content is as follows.

[0084] One transmitter and three receivers were deployed in a spacious indoor environment measuring 6.4 x 6.4 meters. The cables connecting the transmitters and receivers formed a 90-degree angle, creating a square tracking area. All transmitters and receivers were pre-installed with Ubuntu 14.04.3 and equipped with Intel 5300 series wireless network cards. Each transmitter had a single antenna, while each receiver had three antennas arranged linearly, with a spacing of 2.5 cm between the antennas. The signal frequency range was 5.31–5.33 GHz. Both transmitters and receivers were placed on high stools 0.8 meters high. All transmitters and receivers were remotely connected to a Lenovo Xiaoxin Pro 13 laptop via SSH, allowing remote control of the access point's transmission and reception via commands from the host computer. After data collection, code written in Matlab was executed on a desktop computer.

[0085] The proposed method specifically includes the following:

[0086] (1) Feature extraction stage:

[0087] 1.1) First, the received signal is processed using CSI ratio to eliminate time asynchrony and random offset caused by hardware. Then, a sliding window is moved across the original signal, and a Fourier transform is performed on each time segment to extract its frequency domain information. The component with the highest energy in the time spectrum is determined as the DFS, and then the PLCR is calculated according to the formula:

[0088]

[0089] In the formula, f D λ represents frequency offset (DFS); λ represents wavelength; and r represents path length change rate (PLCR).

[0090] 1.2) Next, calculate the phase change of the signal received by the receiver, and use C-STD to calculate the periodic standard deviation of the phase to eliminate the effect of phase jump:

[0091]

[0092] In the formula, N represents the total number of data sampling points; n represents the nth sampling point; e represents the base of the natural logarithm; j represents the imaginary unit; x n This represents the phase value of the nth sampling point; This represents the phase mean of all sampled data points.

[0093] (2) Tracking and initial location positioning stage:

[0094] 2.1) First, using the coordinate information of the device recorded in the test environment, the user-feasible domain for the simulation process is determined. A trajectory random generator is used to generate simulated human motion trajectories in a real-world scenario, covering all possible walking scenarios for the user within the simulation environment. During the simulation, this invention incorporates human kinematic constraints on the simulation trajectory; the specific kinematic constraint formulas are as follows:

[0095]

[0096] In the formula, Let represent the acceleration and velocity of the human body during the simulation, respectively; Δt = 0.1s, and the time length of a single trajectory is 10s. Simultaneously, the generated trajectory is constrained by the feasible region boundary.

[0097]

[0098] In the formula, x and y represent the x and y coordinates of a location point in the user's location sequence, respectively; x lower ,y lowerLet x represent the minimum values ​​of the x and y coordinates within the feasible region, respectively; upper ,y upper These represent the maximum values ​​of the horizontal and vertical coordinates within the feasible region, respectively. Simulated trajectories generated by the random trajectory simulator that exceed the feasible region are discarded. The above simulation process is repeated to obtain over 100,000 usable simulated trajectories as the training dataset for the subsequent neural network.

[0099] 2.2) In this stage, the present invention performs feature extraction on the 100,000 trajectory sequences generated in the simulation in 2.1), which yields the corresponding position sequences P. d (i) with PLCR sequence R d (i) and CI sequence CI d (i), where d represents the d-th simulated trajectory and i is the i-th position point of a single trajectory.

[0100] Now we can define the feature set and the label set:

[0101] Set feature =[R d (i),CI d (i),P init (i)]

[0102] Set label =[P(i,j)]

[0103] In the formula, P init (i) represents the initial position point; P(i,j) represents the set of position points on the trajectory;

[0104] In short, assuming there are N receivers and the output trajectory length is T, the input to the neural network is a (2N+2)×T dimensional matrix. The first N columns represent PLCR, the (N+1)th to 2Nth columns represent CI, and the last two columns indicate the x and y coordinates of the initial position.

[0105] 2.3) In this stage, one transmitter and three receivers are configured according to the established experimental setup. The host computer remotely controls the four devices via the SSH protocol. The transmitter is set to inject mode with a packet interval of 1ms (i.e., a packet rate of 1000 packets / s), and its signal transmission is controlled via command line. Both receivers are configured to monitor mode to ensure simultaneous detection of the channel state information of the transmitted signals. The received signals undergo feature extraction before being concatenated with the initial position to obtain the neural network input shown in 2.2). The predicted trajectory is then obtained through neural network prediction.

[0106] 2.4) The feasible region is divided into sufficiently small units as candidate points for the initial position. Each position is expanded and combined with the PLCR and CI sequences extracted from the received signal to predict a trajectory using a neural network. Trajectory clusters are obtained by traversing all candidate points in the feasible region. The calculated PLCR and CI sequences are obtained by applying the inverse process of the model method to each trajectory in the trajectory cluster. The error is estimated by comparing the calculated feature sequences with the feature sequences extracted from the real signal. The trajectory sequence corresponding to the feature sequence with the smallest error is the optimal trajectory, and the corresponding initial position is the final result.

[0107] (3) Trajectory correction stage:

[0108] 3.1) In this stage, the present invention will predict the corresponding current trajectory T based on the current data using a neural network. current Assume all data are D(i), i∈(0,n), and the current data is D(i'), i'∈(m,n). Based on network prediction, T is obtained. current Assuming there are five historical windows, each with a length of winlen and a window step size of l, the historical trajectories obtained from the historical data are T. his (0, winlen), T his (l,winlen+l),T his (2*l, winlen+2*l), T his (3*l, winlen+3*l), T his (4*l, winlen+4*l); where winlen+4*l=m; the confidence scores of the five historical trajectories and the current trajectory are solved using a model-based fitter. Specifically, by applying the inverse process of the model method, the corresponding PLCR features of the five historical window trajectories and the current trajectory can be obtained. The confidence score of each trajectory is obtained by comparing the measured and observed values ​​of these PLCRs and calculating the error. The smaller the error between the observed and calculated PLCRs, the higher the confidence score of the trajectory, meaning that the trajectory is more reliable. Data with a confidence score lower than the current trajectory in the historical trajectories are discarded to avoid introducing additional errors. The remaining trajectories are weighted according to their confidence scores, and the final trajectory is obtained by summing the weights.

[0109]

[0110] In the formula, ω k The weight of each trajectory segment is calculated based on the confidence level of the trajectory; T k This represents the k-th trajectory.

[0111] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A long-term passive human body tracking method based on commercial Wi-Fi devices, characterized in that, include: S1. Feature Extraction: S1.1 Eliminate time asynchrony and random offset caused by hardware devices by using CSI ratio; The CSI signal is processed using short-time Fourier transform technology to extract the frequency shift (DFS) caused by changes in path length, and then the user's speed information is calculated. S1.

2. To detect the cross-link behavior of the target by CI in response to phase changes caused by signal abrupt changes; The C-STD method for correcting phase jumps is applied to address the potential interference of phase abrupt changes on detection accuracy. S2. Tracking and initial position positioning: S2.1 Tracking: Using a knowledge-driven approach, a large-scale neural network training dataset is generated by simulating real human movement trajectories; Considering the limitations of human kinematics, the simulation principle of "acceleration-velocity-position sequence" is used to simulate all walking trajectories of users in the positioning environment, so as to achieve the goal of covering the global positioning area; The initial position and CI are used as input features to further optimize the prediction model of the neural network; S2.2 Initial Position Localization: The neural network is combined with the theoretical model, and the initial position is obtained by reverse engineering using the analytical capability of the model; the user trajectory is obtained by using all candidate points within the feasible region, and the corresponding PLCR and CI are obtained by using the Fresnel zone theory and the reverse engineering of the model. The optimal initial position is obtained through feature mapping. S3, Trajectory Correction: A method that integrates current data and historical data is proposed to optimize trajectory correction. A confidence evaluation mechanism is constructed, and a model-based fitter is designed to evaluate the reliability of each historical data segment and assign corresponding weights to it according to its confidence. By aggregating data trajectories with confidence levels higher than the current data segment, a more accurate and reliable trajectory is constructed.

2. The long-term passive human body tracking method based on commercial Wi-Fi devices according to claim 1, characterized in that, The formula for the operation described in S1.1 is as follows: In the formula, f D λ represents frequency offset (DFS); λ represents wavelength; and r represents path length change rate (PLCR).

3. The long-term passive human body tracking method based on commercial Wi-Fi devices according to claim 2, characterized in that, The formula for the operation described in S1.2 is as follows: In the formula, N represents the total number of data sampling points; n represents the nth sampling point; e represents the base of the natural logarithm; j represents the imaginary unit; x n This represents the phase value of the nth sampling point; This represents the phase mean of all sampled data points.

4. The long-term passive human body tracking method based on commercial Wi-Fi devices according to claim 3, characterized in that, Human kinematics constraints are incorporated into the simulation of the walking trajectory. The formula for these human kinematic constraints is as follows: In the formula, These represent the acceleration and velocity of the human body during the simulation, respectively. The generated simulation trajectory is constrained by the feasible region boundary, and its formula is expressed as: In the formula, x and y represent the x and y coordinates of a location point in the user's location sequence, respectively; x lower ,y lower Let x represent the minimum values ​​of the x and y coordinates within the feasible region, respectively; upper ,y upper These represent the maximum values ​​of the horizontal and vertical coordinates within the feasible region, respectively.

5. The long-term passive human body tracking method based on commercial Wi-Fi devices according to claim 4, characterized in that, Repeat the simulation process described in S2.1 to obtain multiple simulation trajectories as the training dataset for the subsequent neural network; extract features from the obtained simulation trajectories to obtain the corresponding position sequences P. d (i) with PLCR sequence R d (i) and CI sequence CI d (i), where d represents the d-th simulation trajectory and i is the i-th position point of a single trajectory; Define the feature set and the label set: Set feature =[R d (i),CI d (i),P init (and)] Set label =[P(i,j)] In the formula, P init (i) represents the initial position point; P(i,j) represents the set of trajectory location points; Assuming there are N receivers and the output trajectory length is T, the input of the neural network is a (2N+2)×T dimensional matrix; the first N columns represent PLCR, the (N+1)th to 2Nth columns represent CI, and the last two columns represent the x and y coordinates of the initial position.

6. The long-term passive human body tracking method based on commercial Wi-Fi devices according to claim 5, characterized in that, S3 specifically includes the following: Assuming all data are D(i), i∈(0,n), and the current data is D(i'), i'∈(m,n), T is obtained based on network prediction. current Assuming there are five historical windows, each with a length of winlen and a window step size of l, the historical trajectories obtained from the historical data are T. his (0, winlen), T his (l,winlen+l),T his (2*l, winlen+2*l), T his (3*l, winlen+3*l), T his (4*l, winlen+4*l); where winlen+4*l=m; A model-based fitter is used to solve for the confidence scores of the five historical trajectories and the current trajectory. The inverse process of the model method is applied to obtain the corresponding PLCR features of the five historical window trajectories and the current trajectory. The confidence score of each trajectory is obtained by comparing the measured and observed PLCR values. The smaller the error between the observed and calculated PLCR, the higher the confidence score of the trajectory. Data with confidence scores lower than the current trajectory in the historical trajectories are discarded to avoid introducing additional errors. The remaining trajectories are weighted according to their confidence scores, and the final trajectory is obtained by summing the weights. The specific formula is as follows: In the formula, ω k The weight of each trajectory segment is calculated based on the confidence level of the trajectory; T k This represents the k-th trajectory.

7. A long-term passive human body tracking system based on commercial Wi-Fi devices, applicable to the long-term passive human body tracking method based on commercial Wi-Fi devices as described in any one of claims 1-6, characterized in that, It includes a feature extraction module, a tracking and initial position localization module, and a trajectory correction module, wherein: Feature extraction module: The CSI signal is processed to eliminate time asynchrony and random offset caused by hardware devices; the CSI signal is processed using short-time Fourier transform (STFT) technology to extract the frequency offset (DFS) caused by changes in path length, and then the user's speed information is calculated. To address phase changes caused by signal abrupt changes, cross-link behavior of the target is detected using CI; the C-STD method for correcting phase jumps is applied to resolve the potential interference of phase abrupt changes on detection accuracy. Tracking and initial position positioning module: Tracking submodule: Employs a knowledge-driven approach to generate a large-scale neural network training dataset by simulating real human movement trajectories; Considering the limitations of human kinematics, the simulation principle of "acceleration-velocity-position sequence" is used to simulate all walking trajectories of users in the positioning environment, so as to achieve the goal of covering the global positioning area; The initial position and CI are used as input features to further optimize the prediction model of the neural network; Initial position localization submodule: Combines neural network with theoretical model, uses the analytical capability of model to back-calculate initial position; uses all candidate points within feasible region to obtain user trajectory, combines Fresnel zone theory, uses back-calculation of model to obtain corresponding PLCR and CI, and obtains optimal initial position through feature mapping; Trajectory correction module: A method that integrates current data and historical data is proposed to optimize trajectory correction. A confidence evaluation mechanism is constructed, and a model-based fitter is designed to evaluate the reliability of each historical data segment and assign corresponding weights to it according to its confidence. By aggregating data trajectories with confidence levels higher than the current data segment, a more accurate and reliable trajectory is constructed.

8. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the instruction, program, code set, or instruction set is loaded and executed by the processor to implement the long-term passive human body tracking method based on a commercial Wi-Fi device as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, which is loaded and executed by a processor to implement the long-term passive human body tracking method based on a commercial Wi-Fi device as described in any one of claims 1-6.

Images