A Wi-Fi Data-Driven Method for Non-Line-of-Sight Tracking in Smart Homes

Abstract

This invention discloses a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes, belonging to the technical field of Wi-Fi data-driven methods. This invention applies Wi-Fi signal channel change feature extraction technology, hyperbolic region tracking technology, and general-purpose scenario training dataset self-generation technology. By analyzing the characteristic changes of CSI signals in non-line-of-sight scenarios, a new theoretical model is proposed to model the mapping relationship between user behavior and signal features in non-line-of-sight scenarios. The model is learned using a neural network method, and finally, non-line-of-sight tracking functionality is achieved based on the designed and trained model. Compared with existing technologies, this invention can achieve higher tracking accuracy than pure model methods, and has greater flexibility and environmental adaptability, meeting the needs of diverse smart home scenarios.

Description

Technical Field

[0001] This invention relates to the field of Wi-Fi data-driven methods, and more specifically to a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes. Background Technology

[0002] In recent years, location information has become an indispensable part of modern society, with accurate location data playing a crucial role in various fields such as autonomous driving, ship positioning, and pedestrian navigation. To meet the demand for high-precision indoor location acquisition, research into indoor positioning technology has filled the technological gap in indoor environments compared to traditional positioning methods and actively promoted the development of the technology market. Currently, various technologies are being applied to indoor positioning research to achieve high-precision positioning, including acoustic positioning, ultra-wideband positioning, Bluetooth positioning, and Wi-Fi positioning. Although these positioning technologies can achieve considerable performance in experimental environments, except for Wi-Fi positioning, other technologies require the additional deployment of a large amount of dedicated hardware and equipment. In contrast, Wi-Fi-based indoor positioning has the technological advantages of widespread availability, low-cost deployment, and strong real-time performance. Furthermore, the IEEE 802.11bf protocol, themed around Wi-Fi sensing, is expected to be released in 2024, and one of the important applications supported by this standard is human body tracking and positioning. This means that in the future, Wi-Fi tracking will not only have the potential to be easily deployed in modern smart environments, but will also be able to network with each other based on communication protocol standards, forming a positioning network with wider coverage and stronger computing power, thereby greatly enhancing the social and commercial value of Wi-Fi tracking.

[0003] The rapid development of IoT technology has injected unprecedented intelligence and convenience into our lives. From smart homes to industrial automation, IoT is profoundly changing the way we live and work. Continuous innovation and miniaturization of sensors have made IoT devices increasingly compact and sophisticated, easily embedded in various devices and environments. This tight connectivity enables devices to communicate and collaborate in real time, greatly improving efficiency and intelligence. The past decade has witnessed the rapid development of Wi-Fi-based deviceless tracking. Because this technology does not require users to carry specific devices, it has promising applications in various scenarios. The basic idea behind deviceless tracking is to capture and analyze signals reflected from the human body from Channel State Information (CSI), and then extract motion-related features to track objects. To reveal the essence of deviceless sensing and tracking, the well-known Fresnel zone model was introduced. While the Fresnel zone model is useful in solving tracking problems in line-of-sight (LoS) scenarios, it remains insufficient in solving tracking challenges in non-line-of-sight (NLoS) scenarios.

[0004] In smart home applications, due to the complexity of indoor object placement, non-line-of-sight scenarios are more common than line-of-sight scenarios. With the development of IoT technology and the widespread adoption of smart home devices, many smart devices, such as mobile phones, smart speakers, smart lights, and TVs, are connected to a Wi-Fi router indoors. Therefore, we can obtain multiple Wi-Fi links for device tracking. Although there are many usable Wi-Fi devices in the bedroom, because the Wi-Fi router is located in the living room and there is no line-of-sight path between it and the devices in the bedroom, these non-router smart devices can only communicate with the Wi-Fi router. Figure 1 As shown, when we control a light using a smart speaker, the control command is first sent to the Wi-Fi router, then to the cloud, and finally transmitted by the router to the smart light. In this process, there is no direct Wi-Fi line-of-sight connection between the smart speaker and the light. Using Wi-Fi Direct technology to achieve direct communication between any two devices could increase the design cost of smart devices and lead to more complex communication protocols; moreover, this method may be incompatible with existing smart home setups. To address these issues, this invention proposes a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes. Summary of the Invention

[0005] The purpose of this invention is to provide a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes to solve the problems mentioned in the background art. This invention analyzes the characteristic changes of CSI signals in non-line-of-sight scenarios, proposes a new theoretical model to complete the modeling of the mapping relationship between user behavior and signal characteristics in non-line-of-sight scenarios, uses neural network methods to learn the model, and finally relies on the designed and trained model to realize the non-line-of-sight tracking function.

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

[0007] A Wi-Fi data-driven method for non-line-of-sight tracking in smart homes includes the following steps:

[0008] S1. Based on hyperbolic region theory, realize non-line-of-sight tracking: eliminate the common propagation path of the "transmitter-user" part, propose a new quantitative feature of differential reflection path length change rate, so as to establish the motion model relationship between "user-receiver", that is, the tracking model applicable to NLoS scenario - hyperbolic region model.

[0009] S2. Wi-Fi Channel Variation Characteristics Analysis: A complete channel variation characteristic estimation mechanism is designed, combining time-frequency analysis and CSI ratio model to achieve accurate estimation of Differential Path Length Change Rate (DPLCR).

[0010] S3. Design a data-driven framework to achieve predictive tracking: Provide sufficient datasets for the neural network prediction model by generating datasets on its own. With the help of the fitting effect of the neural network, learn the hyperbolic tracking theory in S1, and then use the trained neural network model to achieve fast and accurate trajectory tracking.

[0011] Preferably, S1 specifically includes the following:

[0012] S1.1 Assume there are two receivers that receive the Wi-Fi signal through diffuse reflection from the user's torso, and there is a common portion between the two signal transmission connections. When a person moves within the monitored area, the common portion of the two transmission connections will always undergo the same modification. By comparing the two trajectories and calculating the time derivative, the differential reflection path length change rate (DPLCR) is obtained as follows:

[0013] (1)

[0014] in, Indicates the first i The reflection path length variation rate (PLCR) of a link. Indicates the firsti Article and Section j Differential reflection path length variation rate (DPLCR) between links. Indicates the first i Article and Section j Rate of change of differential reflection path length between links T x Indicates the transmitter. H Indicates user, R i Indicates receiver i , Indicates the reflection path "transmitter-user-receiver" i The length of “” Indicates user and receiver i The path length between them; relying on the differential reflection path length change rate (DPLCR) to establish a "virtual" link between the receiver and the user, so as to reduce the adverse effects of redundant transmission channels on tracking sensing;

[0015] S1.2, Solve the velocity mapping based on the differential reflection path length change rate (DPLCR) obtained in S1.1: Assume there is one transmitter and... N The receiver, the first n Receiver coordinates This indicates that the user's current location coordinates are represented as follows: User speed is expressed as According to the projection theorem, the user and the first n The change in distance between receivers is represented as:

[0016] (2)

[0017] in, θ n Indicates the human body and the first n The connection between the receivers is x Projection angle on the axis This indicates the calculation of the distance between two points; substituting equation (2) into equation (1) yields the following expression for the differential reflection path length change rate (DPLCR):

[0018] (3)

[0019] in:

[0020] (4)

[0021] S1.3. By combining all candidate receiver devices in pairs, the following formula is obtained:

[0022] (5)

[0023] in, and They are represented as follows:

[0024] (6)

[0025] S1.4 Solving through iteration and The optimal velocity estimate for the corresponding position is obtained, and the specific calculation formula is as follows:

[0026] (7)

[0027] in, T Indicates transposed convolution;

[0028] S1.5. Using the optimal velocity estimate obtained in S1.4, determine the precise coordinates for the next moment. Repeat this process until a complete trajectory sequence is obtained. Then, derive the motion velocity from the acquired differential reflection path length change rate (DPLCR). Based on the above, the only input for the entire process is the user's initial position. The process of determining the coordinates is represented as follows:

[0029] (8)

[0030] in, This represents the user's position at time t. Δt The time interval represents the differential reflection path length change rate (DPLCR) sequence.

[0031] Preferably, S2 specifically includes the following:

[0032] S2.1 Use high-pass and low-pass filters to filter the signal and focus the signal features on the changes caused by the human body movement of interest;

[0033] S2.2. Use the ratio between antennas to eliminate phase shift, and use time-frequency analysis to help track the signal changes caused by human movement;

[0034] S2.3. Use short-time Fourier transform to process signal features and convert the position of the highest energy along the time distribution into the reflection path length change rate (PLCR). The specific conversion formula is as follows:

[0035] (9)

[0036] in, f Doppler This is the Doppler frequency shift of the signal, corresponding to the highest energy portion of the spectrum; This represents the reflection path length variation rate (PLCR) of the nth link. λ Indicates the wavelength of the transmitted signal;

[0037] S2.4 Select the part of the frequency spectrum with the highest power as the approximation of the Doppler frequency shift, and then calculate the difference between the reflection path length change rate (PLCR) according to the specified combination order to achieve accurate differential reflection path length change rate (DPLCR) estimation.

[0038] Preferably, S3 specifically includes the following:

[0039] S3.1. Introduce motion constraints into the simulation dataset. Based on the mathematical and physical relationship between velocity and position, derive the feasible region of the target position using velocity constraints. Specifically, the motion constraints refer to:

[0040] (10)

[0041] in, Indicates the simulated trajectory at t The coordinates of the current moment; Δt This indicates the simulation sampling time interval, which is consistent with the STFT sliding window time interval. Indicates the simulation speed; Indicates simulated acceleration; α ( t () indicates the angle between the velocity direction and the reference coordinate axis; β ( t () indicates the angle between the direction of acceleration and the reference coordinate axis;

[0042] S3.2. Introduce spatial constraints into the simulation dataset. These spatial constraints refer to the actual dimensions of the indoor space and the positions of objects. The specific formulas are as follows:

[0043] (11)

[0044] Among them, M allow Indicates the feasible indoor area;

[0045] S3.3, Randomly set the initial positions of a series of trajectories within the feasible area. =[ x e ( t 0), y e ( t 0)]; Generate a sufficient number of simulation trajectory sequences according to the motion constraints described in S3.1 to ensure coverage of the entire feasible region;

[0046] S3.4. Calculate the DPLCR sequence of the simulated device position using the trajectory coordinates of the simulated trajectory sequence described in S3.3, as a feature dataset. D e The corresponding trajectory sequences are used as the label dataset. T e The specific formula is as follows:

[0047] (12)

[0048] in, t Indicates time;

[0049] S3.5. Use an LSTM layer as the core layer of the network, utilize the LSTM neural network to learn the mapping relationship between user position and differential reflection path length change rate (DPLCR), and build a data-driven tracking framework model.

[0050] S3.6. Input the accurate differential reflection path length change rate (DPLCR) estimate obtained in S2 and the initial position information in S3.3 into the data-driven tracking framework model constructed in S3.5 to generate the final predicted trajectory. Its formula is expressed as follows:

[0051] (13)

[0052] in, This represents the trained data-driven tracking framework model; This represents the actual rate of change of differential reflection path length (DPLCR).

[0053] Compared with existing technologies, this invention proposes a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes, which has the following advantages:

[0054] (1) This invention implements a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes. Based on the proposed method, this invention implements the proposed Wi-Fi data-driven method using commercial equipment. This method can be quickly deployed in new environments and takes into account spatial constraints and obstacles in the environment when predicting trajectories. It can achieve higher tracking accuracy than pure model methods and has greater flexibility and environmental adaptability, which can meet the needs of diverse smart home scenarios.

[0055] (2) This invention discovers and proposes for the first time a theoretical tracking model applicable to NLoS scenarios. Unlike the Fresnel zone applied to LoS ​​scenarios, this model reveals the essential mechanism of signal propagation in NLoS scenarios, providing fundamental theoretical guidance for non-line-of-sight tracking, and proposes a new signal feature differential reflection path length change rate (DPLCR) to achieve quantization mapping of user speed.

[0056] (3) This invention uses a new parameter extraction method to obtain an accurate solution for the path length variation (PLCR) of the reflection path, thereby obtaining a more accurate differential path length variation (DPLCR). The accuracy of the differential path length variation (DPLCR) plays a crucial role in the final position tracking.

[0057] (4) This invention designs and develops a complete data-driven framework. It relies on the idea of ​​guiding neural networks with theory and uses the method of self-generated datasets to train the Siamese neural network, so that the neural network can achieve better tracking results than the pure model method in a shorter time.

[0058] (5) The present invention has conducted a large number of experiments in multiple environments to verify the performance of the system. A Wi-Fi data driving system matching the method was built using a commercial Intel 5300 network card, and decimeter-level tracking accuracy was achieved in the NLoS scenario. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of a typical non-line-of-sight smart home scenario mentioned in the background section of this invention;

[0060] Figure 2 This is a schematic diagram illustrating the principle of movement from the Fresnel region to the hyperbolic region mentioned in Embodiment 1 of the present invention;

[0061] Figure 3 This is a flowchart of the Wi-Fi channel change feature extraction process mentioned in Embodiment 1 of the present invention;

[0062] Figure 4 This is a schematic diagram comparing the Fresnel region and the hyperbolic region mentioned in Embodiment 1 of the present invention;

[0063] Figure 5 This is a schematic diagram illustrating the implementation process of the data-driven tracing framework mentioned in Embodiment 1 of the present invention. Detailed Implementation

[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0065] This invention focuses on the passive tracking problem in smart home scenarios, with the main function of enabling user tracking via routers in non-line-of-sight situations. The methods involved primarily include Wi-Fi signal channel variation feature extraction technology, hyperbolic region tracking technology, and general-purpose scenario training dataset self-generation technology.

[0066] Terminology Explanation:

[0067] Wireless sensing refers to the process of acquiring environmental information by receiving, analyzing, and interpreting wireless signals using wireless communication technology. This technology allows systems to sense and understand events, objects, or activities in the environment by monitoring surrounding wireless signals. These wireless signals can come from wireless communication devices such as Wi-Fi, Bluetooth, and RFID. In the field of wireless sensing, researchers typically use wireless sensing technology to implement a range of applications, such as wireless positioning, breath detection, and smart homes. By analyzing and utilizing the characteristics of wireless signals, the system can extract useful information, thereby enabling the perception and analysis of user behavior.

[0068] Wi-Fi wireless sensing is a subfield of wireless sensing that utilizes widely available Wi-Fi devices in the home for data collection and behavior sensing. Initially, Wi-Fi was primarily used for wireless network connections, but Wi-Fi signals themselves carry rich information and can be used for a wider range of applications, including environmental monitoring, object tracking, and location positioning. In Wi-Fi wireless sensing, characteristics of Wi-Fi signals, such as Received Signal Strength Indication (RSSI), Path Length Change Rate (PLCR), and phase, are used to sense changes in the surrounding environment. By analyzing the sensing-related components, the system can monitor and locate people, objects, or other physical environments without the need for additional sensor devices. The advantage of this technology lies in utilizing existing Wi-Fi infrastructure, eliminating the need to deploy additional sensor networks, thereby reducing costs and improving scalability. Wi-Fi wireless sensing is widely used in indoor positioning, smart buildings, and the Internet of Things (IoT), providing an efficient and convenient solution for achieving intelligent and adaptive systems.

[0069] Indoor positioning and tracking technology refers to the use of various technical means in indoor environments to determine the precise location of objects, people, or equipment in space and track them in real time during movement. Although satellite positioning systems such as the Global Positioning System (GPS) and the BeiDou Navigation Satellite System (BDS) can achieve good outdoor positioning and tracking, they suffer from lower positioning accuracy or are unusable in indoor environments. Researchers have begun to consider using commonly used indoor devices such as Wi-Fi and Bluetooth to achieve indoor tracking and improve tracking services.

[0070] Line-of-sight (LoS) and non-line-of-sight (NLoS) primarily refer to whether there are obstacles obstructing the path between devices. When the transmitter can directly transmit a signal to the receiver, the relationship between the devices is considered line-of-sight. Non-line-of-sight (NLoS) indicates that the straight-line propagation path between the transmitter and receiver is obstructed, and the wireless signal is reflected by other objects during its journey from the transmitter to the receiver. In smart home applications, due to the complexity of indoor object placement, NLoS scenarios are more common than line-of-sight scenarios.

[0071] The path length change rate (PLCR) refers to the rate at which the length of a signal changes as it travels through a reflected path during propagation. When a wireless signal propagates in indoor or urban environments, it may undergo multiple reflections, during which the length of the reflected path changes over time. The PLCR, which describes the rate of change of this length, is often used to describe the dynamics and complexity of the signal propagation environment. In tracking systems, the PLCR is crucial for accurately estimating the target's location.

[0072] The Fresnel zone model is a model used to describe the signal propagation characteristics in microwave or wireless communication systems, analyzing signal diffraction and multipath propagation effects. In wireless communication, signal propagation typically involves more than just straight-line propagation; it also includes phenomena such as reflection and diffraction. The Fresnel zone model considers these phenomena and divides the propagation path into a series of elliptical regions centered at the transmitter and receiver points.

[0073] Based on the above, the following describes a Wi-Fi data-driven method for non-line-of-sight tracking in smart homes, proposed in this invention, with specific examples.

[0074] Example 1:

[0075] To ensure adaptability to the testing environment, this invention selected a 6 m × 6 m open space and deployed the system using a commercial Intel 5300 network card. To meet the normal operating requirements of the system, this invention set up a scenario configuration of one transmitter and three receivers, with the relative positions of the transmitter and receivers all at NLoS. All devices were placed 0.8 m above the ground to simulate the typical height of smart devices in a smart home. The transmitter used a single antenna for packet transmission, while the receivers used three antennas for receiving and decoding CSI. The transmission and reception programs used the open-source CSITool code library. To reduce interference from surrounding Wi-Fi communication, this invention used the 5 GHz band for tracking. In the prototype system, the transmitter transmitted packets at a transmission frequency of 1000 Hz. After data acquisition, a laptop computer was used for algorithm processing to obtain the user trajectory. The system code was written in Matlab.

[0076] 1. Non-line-of-sight tracking based on hyperbolic region theory

[0077] A core innovation of this invention is the establishment of a tracking and perception model for NLoS scenarios. Most existing technical solutions are based on Fresnel zones, such as... Figure 2 As shown in (a), the rate of change of reflection path length (PLCR) is affected by the projection of the user's velocity onto the two reflection paths. As the reflection path length changes, the user can be positioned on a series of ellipses with the same focus during motion. By analyzing this process, the correlation between the three physical parameters—the rate of change of reflection path length (PLCR), the user's velocity, and the user's position—can be quantitatively established, as follows:

[0078] (1)

[0079] in, and Indicates user speed ( v x , v y The projection coefficient on the reflection path.

[0080] However, in NLoS scenarios, the signal must undergo one or more reflections before reaching the receiver. For example... Figure 2 As shown in (b), the reflection path length change rate (PLCR) is affected by both the movement of the path reflection point and the user's movement. Specifically, the velocity at point P... Influence T x -p The length change, the user's speed Influence H -R x The length change. The movement of the two endpoints makes it impossible to determine. P - H The direction and length of the segment are determined, therefore the projection method of formula (1) will no longer be applicable.

[0081] Based on the above analysis, it is clear that the reflection path before reaching the user has a significant impact on the original localization model (1). To achieve tracking in NLoS scenarios, removing or analyzing this impact is crucial. Assume there are two receivers that receive the Wi-Fi signal through diffuse reflection from the user's torso. There is a common component between the two signal transmission connections, such as... Figure 2 As shown in (c), when a person moves within the monitored area, the common portion of the two transmission connections will always undergo the same modification. By comparing the two trajectories and calculating the time derivative, the DPLCR can be obtained as follows:

[0082] (2)

[0083] in, Indicates the first i The reflection path length variation rate (PLCR) of a link. Indicates the first i Article and Section j Differential reflection path length variation rate (DPLCR) between links. Indicates the first i Article and Section j Rate of change of differential reflection path length between links T x Indicates the transmitter. H Indicates user, R i Indicates receiver i , Indicates the reflection path "transmitter-user-receiver" i The length of “” Indicates user and receiver i The path length between them; relying on the differential reflection path length variation rate (DPLCR), a "virtual" link is established between the receiver and the user to reduce the adverse effects of redundant transmission channels on tracking sensing. The following describes how to solve for the speed mapping based on the differential reflection path length variation rate (DPLCR):

[0084] Assume there is one transmitter and N The receiver, the first n Receiver coordinates This indicates that the user's current location coordinates are represented as follows: User speed is expressed as According to the projection theorem, the user and the first n The change in distance between receivers is represented as:

[0085] (3)

[0086] in, θ n Indicates the human body and the first n The connection between the receivers is x Projection angle on the axis This indicates the calculation of the distance between two points; substituting equation (3) into equation (2) yields the following expression for the differential reflection path length change rate (DPLCR):

[0087] (4)

[0088] in:

[0089] (5)

[0090] Then, by combining all the candidate receiver devices in pairs, we can obtain the following formula:

[0091] (6)

[0092] in, and They are represented as follows:

[0093] (7)

[0094] Solving through iteration and The optimal velocity estimate for the corresponding location can then be obtained, as follows:

[0095] (8)

[0096] in, T Indicates transposed convolution;

[0097] Finally, the precise coordinates for the next moment can be determined using the optimal velocity estimate. This process is repeated until a complete trajectory sequence is obtained. As mentioned above, the velocity can be derived from the acquired differential reflection path length change rate (DPLCR). Therefore, the only input required for the entire process is the user's initial position. The process of determining coordinates can be represented as follows:

[0098] (9)

[0099] in, This represents the user's position at time t. Δt The time interval represents the differential reflection path length change rate (DPLCR) sequence.

[0100] 2. Analysis of Wi-Fi Channel Variation Characteristics

[0101] Although a velocity mapping model for the NLoS scenario was established in the previous chapter, accurate location determination relies on precise analysis of channel variation characteristics. Due to various factors such as environmental interference and equipment disturbances, it is often difficult to directly obtain usable location-related features from the raw CSI signal. The CSI signal consists of both dynamic and static components; the goal of this invention is to analyze the wireless sensing portion, therefore focusing on the dynamic portion caused by human movement. To achieve this goal, this invention develops a feature extraction method, such as... Figure 3 As shown.

[0102] First, high-pass and low-pass filters are used to filter the signal, focusing the signal features on the changes caused by human motion. Due to phase shift noise, the original CSI signal cannot be used directly. This invention uses the ratio between antennas to eliminate phase shift. Significant power scattering occurs after the signal passes through the human body; therefore, time-frequency analysis can track the signal changes caused by human motion. This invention uses STFT processing and converts the position of highest energy along the time distribution into the reflection path length change rate (PLCR), as shown in the following formula:

[0103] (10)

[0104] in, f Doppler This is the Doppler frequency shift of the signal, corresponding to the highest energy portion of the spectrum; This represents the reflection path length variation rate (PLCR) of the nth link. λ The wavelength of the transmitted signal is represented. Since the human torso absorbs and scatters the majority of the signal power, this invention selects the portion of the frequency spectrum with the highest power as an approximation of the Doppler frequency shift. Then, by calculating the difference between the rates of change of reflection path length (PLCR) in a specified combination order, the differential rate of change of reflection path length (DPLCR) can be obtained.

[0105] 3. Data-driven framework for predictive tracking

[0106] While the aforementioned model theory and feature analysis have enabled the implementation of tracking systems in non-line-of-sight scenarios, relying solely on theoretical models for tracking has certain limitations. Firstly, typical model-based passive localization methods require large-scale traversal and computation to obtain relatively accurate velocity estimates. (See also...) Figure 4When multiple transceiver links exist in the environment, Fresnel-based line-of-sight tracking needs to consider the path length variation rate (PLCR) of each link, while hyperbolic-based non-line-of-sight tracking needs to consider the differential path length variation rate (DPLCR) of each link. This significantly increases the time cost of a single prediction and the overall time consumption. Secondly, pure model-based methods ignore the inherent physical characteristics of human motion, which makes them lack consistent robustness.

[0107] In contrast, data-driven frameworks composed of neural networks have lower execution time costs after training. Furthermore, neural networks can learn spatial boundary constraints and motion capability limitations contained in the dataset, such as human movement patterns, which makes the entire tracking system more robust. Based on these advantages, this invention proposes a more robust and faster system for non-line-of-sight tracking.

[0108] As is well known, the performance and effectiveness of neural network models largely depend on the quality of the training dataset. A high-quality dataset should have the following characteristics: (1) The dataset contains a sufficiently large amount of data to ensure that the neural network can fully learn the data distribution pattern. (2) The dataset is rich and diverse in content, which can ensure that the neural network can adapt to different scenarios and needs. (3) The dataset has a consistent annotation format to ensure that the neural network can understand the meaning of the data. However, it is regrettable that datasets in the field of Wi-Fi wireless sensing do not exhibit the above three characteristics. Existing open-source datasets have few samples and are only for specific scenarios, so the system will be completely unusable when the device location changes. Therefore, this invention designs a method for self-generating simulation datasets, incorporating human motion constraints and spatial constraints into the dataset.

[0109] This invention first introduces motion constraints into the dataset. For human motion in an indoor scene, the acceleration of the actual human body will be within a certain range, and the probability of the target turning during movement will tend to stabilize over time. Based on the mathematical and physical relationship between velocity and acceleration, it can be inferred that the simulated velocity of human motion will vary within a feasible region without sudden or abnormal changes. Based on the relationship between velocity and position, the feasible region of the target position can be derived using velocity constraints. The above motion constraints can be expressed as follows:

[0110] (11)

[0111] in, Indicates the simulated trajectory at t The coordinates of the current moment; Δt This indicates the simulation sampling time interval, which is consistent with the STFT sliding window time interval. Indicates the simulation speed; This represents the simulated acceleration. Since indoor environments typically do not involve strenuous activity, the magnitude of the acceleration is usually less than 2. m / s 2 The speed is usually less than 3 m / s Under simulation conditions, the probability of a steering event occurring is set to 0.05. α ( t () indicates the angle between the velocity direction and the reference coordinate axis; β ( t The angle () represents the angle between the direction of acceleration and the reference coordinate axis.

[0112] Then, this invention adds spatial constraints to the simulation dataset. These constraints mainly come from the actual dimensions of the indoor space and the positions of objects, as expressed by the following formula:

[0113] (12)

[0114] Among them, M allow This indicates a feasible indoor area, which can be flexibly adjusted when deploying in new environments. The present invention randomly sets the initial positions of a series of trajectories within this feasible area. =[ x e ( t 0), y e ( t 0). Then, a sufficient number of simulated trajectory sequences are generated according to formula (11) to ensure coverage of the entire feasible region. These trajectory coordinates and device positions are used to calculate the simulated differential reflection path length change rate (DPLCR) sequence, which is used as a feature dataset. D e Furthermore, considering that the above simulation process employs a supervised learning method, this invention uses the corresponding trajectory sequences as a label dataset. T e The training dataset can be represented as follows:

[0115] (13)

[0116] in, t The time is represented. After the above processing, the present invention successfully obtained the training dataset. Testing showed that the method proposed in this invention can generate 150,000 simulation data points within 7 seconds. Using these data as input to the training process, the corresponding solution function represented by the neural network model can be obtained. For a new scenario, model training can be completed within minutes without any additional manual labor.

[0117] After solving the dataset problem, the core part of the neural network, namely the neural network architecture, needs to be designed. The method of this invention adopts a "theoretical model guiding the neural network" approach, with the overall design philosophy being "emphasizing the theoretical model and downplaying the neural network," allowing the neural network to only play a fitting role. Considering the good predictive properties of LSTM for time series, this invention uses one LSTM layer as the core layer of the network, utilizing the LSTM neural network to learn the mapping relationship between the user's position and the differential reflection path length change rate (DPLCR). Finally, the differential reflection path length change rate (DPLCR) obtained from the above practical analysis and the initial position are input into the designed data-driven tracking framework to obtain the final predicted trajectory. It can be represented as follows:

[0118] (14)

[0119] in, This represents the trained data-driven tracking framework model; This represents the actual rate of change of differential reflection path length (DPLCR). The implementation of the complete data-driven framework is as follows: Figure 5 As shown.

[0120] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A Wi-Fi data-driven method for non-line-of-sight tracking in smart homes, characterized in that, Includes the following steps: S1. Based on the hyperbolic region theory, non-line-of-sight tracking is realized: the common propagation path of the "transmitter-user" part is eliminated, and a new quantitative feature of the differential reflection path length change rate is proposed to establish the motion model relationship between "user-receiver", that is, the tracking model applicable to NLoS scenario - the hyperbolic region model. S2. Wi-Fi Channel Variation Characteristics Analysis: A complete channel variation characteristic estimation mechanism is designed, combining time-frequency analysis and CSI ratio model to achieve accurate estimation of differential reflection path length variation rate; S3. Design a data-driven framework to achieve predictive tracking: Provide a dataset for the neural network prediction model by generating a dataset on its own. With the help of the fitting effect of the neural network, learn the hyperbolic tracking theory in S1, and then use the trained neural network model to achieve fast and accurate trajectory tracking.

2. The Wi-Fi data-driven method for non-line-of-sight tracking in smart homes according to claim 1, characterized in that, S1 specifically includes the following: S1.1 Assume there are two receivers that receive the Wi-Fi signal through diffuse reflection from the user's torso, and there is a common part between the two signal transmission connections. When a person moves within the monitored area, the common part of the two transmission connections will always undergo the same modification. By comparing the two trajectories and calculating the time derivative, the rate of change of the differential reflection path length is obtained as follows: (1) in, Indicates the first i The rate of change of the reflection path length of the link Indicates the first i Article and Section j Rate of change of differential reflection path length between links T x Indicates the transmitter. H Indicates user, R i Indicates receiver i , Indicates the reflection path "transmitter-user-receiver" i The length of “” Indicates user and receiver i The path length between them; relying on the differential reflection path length change rate, a "virtual" link is established between the receiver and the user to reduce the adverse effects of redundant transmission channels on tracking and sensing; S1.2, Solve the velocity mapping based on the differential reflection path length change rate obtained in S1.1: Assume there is one transmitter and... N The receiver, the first n Receiver coordinates This indicates that the user's current location coordinates are represented as follows: User speed is expressed as According to the projection theorem, the user and the first n The change in distance between receivers is represented as: (2) in, θ n Indicates the human body and the first n The connection between the receivers is x Projection angle on the axis This indicates the calculation of the distance between two points; substituting equation (2) into equation (1) yields the following expression for the rate of change of the differential reflection path length: (3) in: (4) S1.

3. By combining all candidate receiver devices in pairs, the following formula is obtained: (5) in, and They are represented as follows: (6) S1.4 Solving through iteration and The optimal velocity estimate for the corresponding position is obtained, and the specific calculation formula is as follows: .(7) in, T Indicates transposed convolution; S1.

5. Using the optimal velocity estimate obtained in S1.4, determine the precise coordinates for the next moment. Repeat this process until a complete trajectory sequence is obtained. Then, derive the motion velocity from the obtained differential reflection path length change rate. Based on the above, the only input for the entire process is the user's initial position. The process of determining the coordinates is represented as follows: (8) in, This represents the user's position at time t. Δt This represents the time interval of the differential reflection path length change rate sequence.

3. The Wi-Fi data-driven method for non-line-of-sight tracking in smart homes according to claim 1, characterized in that, S2 specifically includes the following: S2.1 Use high-pass and low-pass filters to filter the signal and focus the signal features on the changes caused by the human body movement of interest; S2.

2. Use the ratio between antennas to eliminate phase shift, and use time-frequency analysis to help track the signal changes caused by human movement; S2.

3. Use short-time Fourier transform to process signal features and convert the highest energy position along the time distribution into PLCR. The specific conversion formula is as follows: （9） in, f Doppler This is the Doppler frequency shift of the signal, corresponding to the highest energy portion of the spectrum; This represents the rate of change of the reflection path length of the nth link; λ Indicates the wavelength of the transmitted signal; S2.4 Select the part with the highest power in the time spectrum as an approximation of the Doppler frequency shift, and then calculate the difference between the rates of change of the reflection path length according to the specified combination order to achieve accurate differential reflection path length rate of change estimation.

4. The Wi-Fi data-driven method for non-line-of-sight tracking in smart homes according to claim 1, characterized in that, S3 specifically includes the following: S3.

1. Introduce motion constraints into the simulation dataset. Based on the mathematical and physical relationship between velocity and position, derive the feasible region of the target position using velocity constraints. Specifically, the motion constraints refer to: （10） in, Indicates the simulation trajectory at t The coordinates of the current moment; Δt This indicates the simulation sampling time interval, which is consistent with the STFT sliding window time interval. Indicates the simulation speed; Indicates simulated acceleration; α ( t () indicates the angle between the velocity direction and the reference coordinate axis; β ( t () indicates the angle between the direction of acceleration and the reference coordinate axis; S3.

2. Introduce spatial constraints into the simulation dataset. These spatial constraints refer to the actual dimensions of the indoor space and the positions of objects. The specific formulas are as follows: （11） Among them, M allow Indicates the feasible indoor area; S3.3, Randomly set the initial positions of a series of trajectories within the feasible area. =[ x e ( t 0), y e ( t 0)]; Generate a simulation trajectory sequence according to the motion constraints described in S3.1 to ensure coverage of the entire feasible region; S3.

4. Using the trajectory coordinates of the simulated trajectory sequence described in S3.3, calculate the differential reflection path length change rate sequence of the simulated device position, as a feature dataset. D e The corresponding trajectory sequences are used as the label dataset. T e The specific formula is as follows: （12） in, t Indicates time; S3.

5. Use an LSTM layer as the core layer of the network, utilize the LSTM neural network to learn the mapping relationship between user position and differential reflection path length change rate, and build a data-driven tracking framework model. S3.

6. Input the accurate differential reflection path length change rate estimate obtained in S2 and the initial position information in S3.3 into the data-driven tracking framework model constructed in S3.5 to generate the final predicted trajectory. Its formula is expressed as follows: （13） in, This represents the trained data-driven tracking framework model; This represents the rate of change of the actual acquired differential reflection path length.

Images