Human body action sensing method and system based on multi-node collaboration

Abstract

The invention relates to a human body action sensing method and system based on multi-node collaboration, and the method comprises the steps: collecting CSI data transmitted by each sensing node, estimating the angle information of different sensing nodes, and calculating the position of a human body through a triangulation positioning algorithm; screening a plurality of optimal sensing nodes according to the distance from the sensing nodes to the human body position and the signal quality; extracting a time domain feature, a frequency domain feature and a time-frequency domain feature from CSI data returned by the optimal sensing nodes, and constructing a multi-dimensional feature vector of each optimal sensing node; utilizing an action recognition model to generate action classification results of the corresponding nodes according to the input multi-dimensional feature vectors; and fusing the action classification results of all the optimal sensing nodes into an action recognition result, and dynamically adjusting the fusion weight according to the signal quality. The method can improve the accuracy and efficiency of human body activity recognition in the environment.

Description

Technical Field

[0001] This invention relates to the field of human motion perception technology, and in particular to a human motion perception method and system based on multi-node collaboration. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and intelligent sensing technologies, wireless sensing technology based on WiFi Channel State Information (CSI) has received widespread attention. The CSI received by a WiFi receiver contains the amplitude and phase values ​​of the wireless signal during propagation. When there is human activity in the environment, it affects the signal received by the WiFi, thus enabling non-contact human activity identification by analyzing the CSI data. This technology has significant application value in fields such as smart homes, health monitoring, and security surveillance.

[0003] However, existing WiFi CSI-based sensing technologies are mainly divided into single-node sensing and multi-node sensing. Single-node sensing has obvious limitations: on the one hand, a single WiFi node is easily affected by environmental interference, signal obstruction, multipath effects, etc., resulting in unstable sensing accuracy and poor performance in complex environments; on the other hand, the person being sensed needs to be in a limited area of ​​physical connection between the node and the transmitter, which also limits the spatial range of accurate sensing. Existing multi-node systems lack a unified device scheduling strategy. CN117643468A discloses a wireless fall detection method based on multi-source information fusion. Although this method combines IMU and CSI features, each node works independently and lacks a unified device selection and scheduling mechanism; CN117528768A proposes an indoor positioning method that fuses RSSI and CSI feature parameters. This scheme mainly focuses on positioning accuracy, but does not adequately consider the device selection and collaborative sensing issues among multiple nodes; CN115242327B proposes a CSI action recognition method based on multi-channel fusion, which uses an attention mechanism for feature fusion, but lacks intelligent selection strategies for different device locations and collaborative sensing methods.

[0004] Furthermore, existing methods typically lack effective multi-device collaborative positioning mechanisms, failing to guarantee spatial consistency of data across devices. Existing device selection methods lack intelligent scheduling strategies tailored to the characteristics of different devices, resulting in inconsistent performance in multi-device environments. More importantly, existing technologies generally lack systematic device selection methods, failing to guarantee the sensing accuracy and computational efficiency of multi-device systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a human motion perception method and system based on multi-node collaboration, which can improve the accuracy and efficiency of human activity recognition in the environment.

[0006] The technical solution adopted by this invention to solve its technical problem is: providing a human motion perception method based on multi-node collaboration, applied to a human motion perception system including a transmitting node and a sensing node, wherein the sensing node receives the detection signal transmitted by the transmitting node, extracts CSI data from it, and then transmits it, including:

[0007] Collect CSI data emitted by each sensing node, then estimate the angle information of different sensing nodes, and use the triangulation algorithm to calculate the human body position;

[0008] Based on the distance from the sensing node to the human body location and the signal quality, several optimal sensing nodes are selected.

[0009] Extract time-domain features, frequency-domain features, and time-frequency-domain features from the CSI data returned by the optimal sensing node, and construct a multi-dimensional feature vector for each optimal sensing node;

[0010] An action recognition model is used to generate action classification results for corresponding nodes based on the input multidimensional feature vector.

[0011] The action classification results of all the best perception nodes are fused into action recognition results, and the fusion weights are dynamically adjusted according to the signal quality.

[0012] Furthermore, the step of selecting several optimal sensing nodes based on the distance from the sensing node to the human body location and signal quality includes:

[0013] The sensing node closest to the human body location is selected as the primary sensing node. If the closest sensing node is not unique, the sensing node with the best signal quality is selected as the primary sensing node.

[0014] Select the next closest sensing node as an auxiliary sensing node or put it into standby mode;

[0015] The selected primary and secondary sensing nodes are designated as the optimal sensing nodes.

[0016] Furthermore, the signal quality is evaluated based on signal strength, stability, and noise level.

[0017] Furthermore, the signal quality is calculated using the following formula:

[0018]

[0019] in, Indicates signal quality, Indicates signal strength. Indicates signal stability. Indicates noise level, , , These are the weighting coefficients.

[0020] Furthermore, the calculation of human body position using the triangulation algorithm is achieved by constructing virtual sensing nodes and using a multi-node collaborative method based on the CSI data of the sensing nodes and virtual sensing nodes.

[0021] Furthermore, it also includes a step of periodically checking the data consistency between various sensing nodes using cross-validation methods, and performing calibration when the data deviation exceeds a set threshold.

[0022] Furthermore, for scenarios with prominent temporal or spatial features, action recognition models based on LSTM or CNN are selected respectively.

[0023] Furthermore, the fusion of the action classification results of all the best perception nodes into the action recognition result is achieved by using a weighted voting or Bayesian fusion method.

[0024] Furthermore, the terminal nodes are evenly distributed in the target area.

[0025] The present invention also provides a human motion perception system based on multi-node collaboration, comprising:

[0026] Transmitting nodes are used to transmit detection signals;

[0027] The sensing node is used to receive the detection signal transmitted by the transmitting module, extract the CSI data from it, and then transmit it.

[0028] The positioning module is used to collect CSI data emitted by each sensing node, thereby estimating the angle information of different sensing nodes, and using the triangulation algorithm to calculate the human body position.

[0029] The node selection module is used to filter several optimal sensing nodes based on the distance of the sensing node from the human body location and the signal quality.

[0030] The action recognition module is used to extract time-domain features, frequency-domain features, and time-frequency-domain features from the CSI data returned by the optimal sensing nodes, and construct a multi-dimensional feature vector for each optimal sensing node; it uses the action recognition model to generate the action classification result of the corresponding node based on the input multi-dimensional feature vector; it merges the action classification results of all optimal sensing nodes into an action recognition result, and dynamically adjusts the fusion weights according to the signal quality.

[0031] Beneficial effects

[0032] By adopting the above-mentioned technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0033] This invention employs an intelligent node selection strategy that prioritizes distance over signal quality. Through Euclidean distance calculation and comprehensive signal quality evaluation, it dynamically selects the optimal node combination. Compared to traditional fixed equipment or polling methods, it achieves an optimal balance between sensing accuracy and computational efficiency, avoiding unnecessary waste of computational resources. Furthermore, it provides data fusion functionality through the collaborative work of master and auxiliary nodes.

[0034] This invention effectively solves the problem of insufficient positioning accuracy in complex indoor environments by using multi-node collaborative positioning and pseudo-point methods. The system uses the MUSIC / ESPRIT algorithm for angle estimation, combined with triangulation calculation and pseudo-point optimization, which has higher positioning accuracy and better environmental adaptability compared with traditional single-device positioning methods, and can effectively cope with complex environmental factors such as multipath effects and signal blockage.

[0035] This invention utilizes real-time data acquisition and processing based on WiFi CSI to meet the needs of real-time sensing applications. It also supports dynamic node deployment, adapting to scenarios involving moving targets. The system features short response time, high processing efficiency, and supports continuous monitoring and real-time feedback, meeting the needs of real-time applications such as smart homes, health monitoring, and security surveillance.

[0036] This invention organically integrates multi-node collaborative localization, intelligent node selection, and action recognition technologies to form a complete sensing solution. The system integrates multiple advanced technologies such as multi-dimensional feature extraction, deep learning models, and data fusion, exhibiting high technical integration and strong practicality. It provides a complete technical solution and implementation path for WiFi CSI-based wireless sensing technology. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a scenario according to the first embodiment of the present invention;

[0038] Figure 2 This is a flowchart illustrating the first embodiment of the present invention;

[0039] Figure 3 This is a system architecture diagram of the first embodiment of the present invention;

[0040] Figure 4 This is a flowchart of the multi-node localization algorithm according to the first embodiment of the present invention;

[0041] Figure 5 This is a flowchart of the intelligent node selection strategy according to the first embodiment of the present invention;

[0042] Figure 6 This is a flowchart of the action recognition algorithm according to the first embodiment of the present invention;

[0043] Figure 7 This is a system calibration flowchart of the first embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0045] The first embodiment of the present invention relates to a multi-node collaborative sensing method based on WiFi CSI, which combines multi-node collaborative positioning with intelligent node selection. First, the human body position is determined through multi-node collaborative positioning, and then the optimal node is selected for action recognition according to a strategy of prioritizing distance and then signal quality, thereby improving computational efficiency while ensuring recognition accuracy.

[0046] Deployment scenarios such as Figure 1 As shown, a top-mounted WiFi gateway architecture is adopted, with multiple Intel 5300 nodes working collaboratively to achieve precise localization and motion recognition of human targets. The specific workflow is as follows: Figure 2 As shown: Multiple Intel 5300 nodes connect to a WiFi gateway via a WiFi network. Each node collects detection signals sent by the WiFi gateway router, extracts CSI data, and transmits it back to the gateway's central processing unit. The central processing unit first uses a multi-Intel 5300 node collaborative localization algorithm to determine the location of the human target, then calculates the distance from each node to the target human, and preferentially selects the closest Intel 5300 node as the primary sensing terminal. When multiple Intel 5300 nodes are close to each other, the system selects the optimal node based on the signal quality evaluation results of each Intel 5300 node. After selecting a node, the system uses the CSI data collected and transmitted by that node for human motion recognition, while simultaneously fusing CSI data collected and transmitted by auxiliary sensing terminals to improve recognition accuracy. In addition, the system includes an automatic calibration mechanism, periodically performing inter-node calibration and data consistency calibration to ensure long-term operational stability.

[0047] The following describes a human motion recognition method based on multi-node intelligent scheduling and collaborative perception using specific embodiments and in conjunction with the accompanying drawings.

[0048] like Figure 3As shown, a multi-node collaborative sensing architecture is adopted, including a WiFi gateway and multiple Intel 5300 nodes. The WiFi gateway, comprising a router and a central processing unit, is mounted in the center of the ceiling and is responsible for continuously sending WiFi detection signals, network management, receiving signals transmitted back from nodes, multi-node positioning, intelligent node selection, and action recognition. The Intel 5300 nodes are evenly deployed throughout the building space and are responsible for continuously receiving WiFi detection signals from the gateway, extracting CSI data, and transmitting it back to the gateway in real time.

[0049] During system initialization, the WiFi router starts up and establishes the network. Multiple Intel 5300 nodes connect and register via the WiFi network, registering their device IDs, hardware information, operating channels, and location information. The central processing unit configures the network based on the node information, ensuring that all nodes work collaboratively on the same channel and establishing a node topology.

[0050] The flowchart of the multi-node localization algorithm is as follows: Figure 4 As shown, by aggregating the WiFi CSI data transmitted back from each Intel 5300 node, and performing preprocessing and filtering, the signal arrival angle is calculated using the MUSIC algorithm or ESPRIT algorithm to establish a mapping relationship between the angle information of the received signal of each node and its spatial location.

[0051] The following explanation uses the MUSIC algorithm as an example: The MUSIC algorithm is a high-resolution spectral estimation technique based on array signal processing, suitable for angle estimation in multipath environments. Its core idea is to decompose the covariance matrix of the received signal into a signal subspace and a noise subspace, and then use the orthogonality of the two to construct a spatial spectral function, with the spectral peak position corresponding to the signal's angle of arrival.

[0052] The spatial spectrum function of the MUSIC algorithm is:

[0053] (1)

[0054] in, It is a direction vector. For the noise subspace, The angle of arrival. In equation (1), the denominator is the direction vector. With noise subspace Orthogonality measure. When When the actual angle is reached, Orthogonal to the signal subspace and correlated with the noise subspace, the denominator becomes extremely small, resulting in a sharp peak in the spectral function. The signal's angle of arrival can be determined by searching for the peak position of the spectral function. .

[0055] The ESPRIT algorithm is similar to MUSIC, but it estimates angles directly through the translation invariance of the array, eliminating the need to search for spectral functions and thus offering higher computational efficiency. The two algorithms can be chosen based on specific requirements.

[0056] A triangulation algorithm is employed, utilizing the angle information from multiple Intel 5300 nodes and combining it with the known coordinates of the nodes to calculate the three-dimensional coordinates of a human target. In indoor environments, multipath effects and non-line-of-sight conditions can lead to angle estimation errors, thereby reducing positioning accuracy. Positioning accuracy can be improved through multi-node collaboration and pseudo-point methods.

[0057] The pseudo-point method introduces virtual auxiliary nodes (pseudo-points) in addition to the actual physical nodes. Based on the known spatial distribution of nodes and signal propagation models, one or more virtual observation points are constructed in geometric space to enhance the stability and redundancy of angle estimation. The location of the pseudo-points can be determined through interpolation or optimization algorithms, such as Kriging interpolation or least squares fitting. Spatial consistency is ensured by maintaining the spatial consistency of data across all nodes within the cell.

[0058] Calculate the Euclidean distance from each Intel 5300 node to the target human body, establish a distance-node mapping relationship, and provide a distance reference for node selection. The distance calculation formula is:

[0059] (2)

[0060] in, Let be the distance from the i-th INTEL 5300 node to the target human body. Let be the three-dimensional coordinates of the position of the i-th node when it is deployed. The three-dimensional coordinates of the target human body.

[0061] A signal quality assessment system is established to evaluate node signal quality from three dimensions: signal strength, stability, and noise level. The formula for calculating the comprehensive signal quality score is as follows:

[0062] (3)

[0063] in, Score the signal quality of the i-th node. For signal strength, For signal stability, Let α be the noise level, β be the weighting coefficients, and α + β + γ = 1.

[0064] Then, based on the human body's location and signal quality, the optimal node combination is dynamically selected, employing a selection strategy that prioritizes distance over signal quality: the primary sensing Intel 5300 node is selected based on distance, and when distances are similar, signal quality is considered; the next closest Intel 5300 node is selected to provide data fusion and backup. The intelligent node selection flowchart is as follows: Figure 5 As shown.

[0065] like Figure 6 As shown, time-domain, frequency-domain, and time-frequency-domain features are extracted from the CSI data of selected nodes to construct a multi-dimensional feature vector. An LSTM or CNN model is used for action classification, and a weighted voting or Bayesian fusion method is employed to fuse the prediction results from the master and auxiliary devices. The fusion weights are dynamically adjusted based on the node signal quality.

[0066] The weighted fusion formula is:

[0067] (4)

[0068] in, For the final fusion result, The recognition result for the nth node is... Let be the fusion weight of the nth node, and ∑ = 1.

[0069] The system calibration procedure is as follows: Figure 7 As shown, periodic mutual calibration between devices is performed, and data consistency is checked using cross-validation. The calibration procedure is automatically initiated when data deviation exceeds a preset threshold. The formula for calculating data deviation is:

[0070] (5)

[0071] in, The data deviation between node i and node j These are the measured values ​​for nodes i and j, respectively. For reference only. When > At that time, the calibration procedure is initiated, in which... The preset threshold is used. Systematic deviations between nodes are eliminated by using a standard signal source or reference node, and the time and positioning deviations of each node are periodically detected and corrected.

[0072] The second embodiment of the present invention relates to a multi-node collaborative sensing system based on WiFi CSI, including a multi-node positioning module, an intelligent node selection module, and an action recognition module.

[0073] 1) Multi-node positioning module

[0074] The multi-node positioning module utilizes data collected and transmitted simultaneously by multiple Intel 5300 nodes to work collaboratively, achieving human target localization by analyzing WiFi CSI data. This module comprises five functional units: CSI data acquisition, angle estimation, positioning calculation, accuracy optimization, and spatial consistency assurance.

[0075] The CSI data acquisition unit is responsible for collecting the WiFi CSI data transmitted back from each Intel 5300 node, and performing preprocessing and filtering to ensure data quality and temporal consistency between nodes. The angle estimation unit uses the MUSIC or ESPRIT algorithm to calculate the signal angle of arrival, establishing a mapping relationship between the angle information of the received signals from each Intel 5300 node and their spatial location. The spatial spectrum function of the MUSIC algorithm is:

[0076] (1)

[0077] in, It is a direction vector. For the noise subspace, To reach the angle.

[0078] The positioning calculation unit employs a triangulation algorithm, utilizing angle information from multiple Intel 5300 nodes to calculate the coordinates of the human target. The positioning accuracy optimization unit enhances positioning accuracy through multi-node collaboration and a pseudo-point method. The pseudo-point method involves introducing virtual auxiliary nodes (i.e., pseudo-points) in addition to the actual physical nodes. Based on the known spatial distribution and signal propagation model of the nodes, one or more virtual observation points are constructed in geometric space to enhance the stability and redundancy of angle estimation. The positions of the pseudo-points can be determined through interpolation or optimization algorithms, such as Kriging interpolation or least-squares fitting. The spatial consistency guarantee unit maintains the spatial consistency of the data across all nodes.

[0079] 2) Intelligent Node Selection Module

[0080] The intelligent node selection module dynamically selects the optimal node combination based on the human's location and signal quality, employing a selection strategy that prioritizes distance over signal quality. This module comprises five functional units: distance calculation, signal quality assessment, primary node selection, secondary node selection, and dynamic adjustment.

[0081] The distance calculation unit calculates the Euclidean distance from each Intel 5300 node to the target human body, establishing a distance-device mapping relationship to provide a distance reference for node selection. The distance calculation formula is:

[0082] (2)

[0083] in, Let be the distance from the i-th INTEL 5300 node to the target human body. Let be the three-dimensional coordinates of the position of the i-th node when it is deployed. The target human body has three-dimensional coordinates. The signal quality assessment unit evaluates the signal quality of nodes, including three dimensions: signal strength, stability, and noise level, establishing a signal quality assessment system. The comprehensive signal quality score calculation formula is as follows:

[0084] (3)

[0085] in, Score the signal quality of the i-th node. For signal strength, For signal stability, Let α be the noise level, β be the weighting coefficients, and α + β + γ = 1.

[0086] The primary node selection unit selects the main sensing terminal based on distance priority, and selects based on signal quality when the distance is close. The auxiliary node selection unit selects the next closest device to provide data fusion. The dynamic adjustment unit adjusts the selection strategy according to environmental changes.

[0087] 3) Action recognition module

[0088] The action recognition module uses CSI data from selected nodes to recognize human actions, and combines this with a deep learning model for feature extraction and classification. This module comprises five functional units: multi-dimensional feature extraction, deep learning model, data fusion, recognition optimization, and real-time processing.

[0089] The multi-dimensional feature extraction unit extracts time-domain, frequency-domain, and time-frequency-domain features from CSI data to construct multi-dimensional feature vectors. The deep learning model unit uses an LSTM or CNN model for action classification, including an input layer, a feature extraction layer, an attention mechanism, and a classification output layer. The data fusion unit fuses data from master and auxiliary nodes using weighted voting or Bayesian fusion methods, dynamically adjusting the fusion weights based on the node signal quality. The weighted fusion formula is:

[0090] (4)

[0091] in, For the final fusion result, The recognition result for the nth node is... Let be the fusion weight of the nth node, and ∑ = 1

[0092] The recognition and optimization unit combines multi-node spatial information and time-frequency domain features to achieve action recognition; the real-time processing unit performs real-time data acquisition and processing based on WiFi CSI.

[0093] 4) System calibration method

[0094] The system calibration method employs an automatic calibration mechanism to periodically calibrate and optimize the system. This method comprises five functional units: inter-node calibration, data time consistency calibration, positioning accuracy calibration, dynamic calibration adjustment, and calibration effect evaluation.

[0095] The inter-node calibration unit periodically performs mutual calibration between nodes, using cross-validation to check data consistency. The calibration procedure is automatically initiated when data deviation exceeds a preset threshold. The data deviation calculation formula is:

[0096] (5)

[0097] in, The data deviation between node i and node j These are the measured values ​​for nodes i and j, respectively. For reference only. When > At that time, the calibration procedure is initiated, in which... The preset threshold is used. The data consistency calibration unit eliminates systematic deviations between nodes through a standard signal source or reference node. The positioning accuracy calibration unit periodically detects and corrects the positioning deviations of each node. The dynamic calibration adjustment unit adjusts calibration parameters according to environmental changes and system performance changes. The calibration effect evaluation unit quantifies the improvement effect by comparing the performance before and after calibration.

[0098] This implementation uses commercial Intel 5300 modules to replace dedicated sensing devices, while utilizing existing WiFi infrastructure, eliminating the need for additional wiring and significantly reducing system deployment and maintenance costs. The Intel 5300 modules are inexpensive, easy to deploy, and have low maintenance costs, offering a clear cost advantage over traditional dedicated sensing devices, making them suitable for large-scale commercial applications and widespread adoption. Furthermore, the system supports flexible configuration of the Intel 5300 modules, allowing for dynamic adjustment of the number of nodes according to application scenario requirements, resulting in a highly scalable and adaptable system architecture. The system supports dynamic addition and removal of devices, accommodating deployment needs of varying scales, from home environments to large office spaces, demonstrating promising commercial application prospects.

Claims

1. A human motion perception method based on multi-node collaboration, applied to a human motion perception system including transmitting nodes and sensing nodes, wherein the sensing nodes receive detection signals transmitted by the transmitting nodes, extract CSI data from them, and then transmit the data, characterized in that, include: Collect CSI data emitted by each sensing node, then estimate the angle information of different sensing nodes, and use the triangulation algorithm to calculate the human body position; Based on the distance from the sensing node to the human body location and the signal quality, several optimal sensing nodes are selected. Extract time-domain features, frequency-domain features, and time-frequency-domain features from the CSI data returned by the optimal sensing node, and construct a multi-dimensional feature vector for each optimal sensing node; An action recognition model is used to generate action classification results for corresponding nodes based on the input multidimensional feature vector. The action classification results of all the best perception nodes are fused into action recognition results, and the fusion weights are dynamically adjusted according to the signal quality.

2. The human motion perception method according to claim 1, characterized in that, The process of selecting several optimal sensing nodes based on the distance from the sensing node to the human body location and signal quality includes: The sensing node closest to the human body location is selected as the primary sensing node. If the closest sensing node is not unique, the sensing node with the best signal quality is selected as the primary sensing node. Select the next closest sensing node as an auxiliary sensing node or put it into standby mode; The selected primary and secondary sensing nodes are designated as the optimal sensing nodes.

3. The human motion perception method according to claim 1, characterized in that, The signal quality is evaluated based on signal strength, stability, and noise level.

4. The human motion perception method according to claim 3, characterized in that, The signal quality is calculated using the following formula: in, Indicates signal quality. Indicates signal strength. Indicates signal stability. Indicates the noise level. , , These are the weighting coefficients.

5. The human motion perception method according to claim 1, characterized in that, The calculation of human body position using the triangulation algorithm is achieved by constructing virtual sensing nodes and using a multi-node collaborative method based on the CSI data of the sensing nodes and virtual sensing nodes.

6. The human motion perception method according to claim 1, characterized in that, It also includes a step of periodically checking the data consistency between various sensing nodes using cross-validation methods, and calibrating when the data deviation exceeds a set threshold.

7. The human motion perception method according to claim 1, characterized in that, For scenarios with prominent temporal or spatial features, action recognition models based on LSTM or CNN are selected respectively.

8. The human motion perception method according to claim 1, characterized in that, The fusion of action classification results from all optimal perception nodes into action recognition results is achieved using a weighted voting or Bayesian fusion method.

9. The human motion perception method according to claim 1, characterized in that, The terminal nodes are evenly distributed in the target area.

10. A human motion perception system based on multi-node collaboration, characterized in that, include: Transmitting nodes are used to transmit detection signals; The sensing node is used to receive the detection signal transmitted by the transmitting module, extract the CSI data from it, and then transmit it. The positioning module is used to collect CSI data emitted by each sensing node, thereby estimating the angle information of different sensing nodes, and using the triangulation algorithm to calculate the human body position. The node selection module is used to filter several optimal sensing nodes based on the distance of the sensing node from the human body location and the signal quality. The action recognition module extracts time-domain, frequency-domain, and time-frequency-domain features from the CSI data returned by the optimal sensing nodes, constructing a multi-dimensional feature vector for each optimal sensing node; utilizing... The action recognition model generates action classification results for corresponding nodes based on the input multidimensional feature vectors; it merges the action classification results of all the best perception nodes into an action recognition result, and dynamically adjusts the fusion weights according to the signal quality.

Images