A target recognition method and system fusing polarization features, a terminal and a medium

Abstract

The application discloses a target recognition method and system fusing polarization features, a terminal and a medium, and the method comprises the following steps: synchronously collecting channel state information on each subcarrier by using a dual-polarized antenna, and constructing a polarization scattering matrix based on the channel state information; decomposing the polarization scattering matrix by using a polarization target decomposition technology to obtain a polarization coherent vector matrix; converting the polarization coherent vector matrix into a polarization word sequence, modeling the polarization word sequence through a Transformer layer of a wireless basic model to obtain a global feature vector; and inputting the global feature vector into a downstream task head to obtain a target recognition result. By introducing polarization position embedding, the wireless basic model can learn the polarization semantics of electromagnetic waves, so that deeper signal feature understanding is realized, and the performance of downstream perception tasks is improved.

Description

Technical Field

[0001] This invention relates to the field of wireless radio frequency sensing and channel signal processing technology, and in particular to a target identification method, system, terminal and medium that integrates polarization features. Background Technology

[0002] Against the backdrop of the rapid development of IoT, human-computer interaction, and intelligent environmental sensing technologies, non-contact passive sensing technology, with its advantages of no device dependence, strong concealment, and wide adaptability, has become a core research direction in the field of intelligent sensing, and is widely used in many scenarios such as human behavior recognition, gesture perception, indoor positioning, and environmental monitoring. Radio frequency sensing technology based on Channel State Information (CSI) can capture the dynamic changes in amplitude and phase during the propagation of spatial electromagnetic waves, mapping the temporal and spatial variation characteristics of wireless signals to target location, human actions, and environmental feature information. This overcomes the limitations of traditional visual sensing, which is restricted by lighting, occlusion, and privacy, and possesses the advantages of all-weather, non-contact, and interference-resistant sensing, making it the mainstream technology solution in the current passive radio frequency sensing field. Traditional CSI radio frequency sensing technology mainly relies on radio frequency equipment to complete data acquisition. It then cleans the raw CSI data using preprocessing methods such as filtering and denoising, and phase calibration. Next, it extracts shallow physical features such as amplitude, phase, angle of arrival, and Doppler shift, and uses deep learning models such as CNN, RNN, and LSTM to establish a mapping relationship between features and sensing targets, achieving various sensing tasks. With the development of cross-modal fusion technology, existing mainstream optimization schemes have introduced a visual base model and a cross-modal comparative knowledge distillation mechanism. Using the visual model as the teacher model and the radio frequency (RF) model as the student model, semantic knowledge transfer from video images compensates for the lack of semantic features in RF signals, effectively improving the recognition accuracy and generalization ability of the RF sensing model. However, existing RF sensing technologies based on wireless base models still have significant technical shortcomings, and their limitations are becoming increasingly apparent. Specifically, current mainstream enhancement solutions rely entirely on cross-modal data such as video to drive the knowledge distillation process, imbuing radio frequency signals with high-level semantic information solely through visual semantics. This over-reliance on external modal data not only increases the cost of data acquisition and model training but also makes them susceptible to issues such as cross-modal data matching bias and poor scene adaptability. Furthermore, existing technologies generally ignore the polarization characteristics naturally carried during the propagation and scattering of radio electromagnetic waves. Polarization features, as inherent physical characteristics that can accurately distinguish targets of different materials and possess strong robustness, have not been effectively explored and utilized. This results in existing radio frequency sensing models having insufficient ability to distinguish target behavior in different materials and complex environments. The physical feature dimensions of the sensing models are singular, and their feature representation capabilities are limited, making it difficult to adapt to complex and ever-changing indoor and outdoor sensing scenarios. This significantly restricts the sensing accuracy, anti-interference capability, and scene adaptability of CSI radio frequency sensing technology, hindering its practical application and technological iteration in high-precision intelligent sensing scenarios.

[0003] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a target recognition method, system, terminal and medium that integrates polarization features, in response to the above-mentioned defects of the prior art. It aims to solve the problems that existing radio frequency sensing methods for wireless basic models rely too much on cross-modal data such as video to assign semantic information and fail to explore and utilize the polarization characteristics carried during the scattering of wireless signals, resulting in single feature dimensions, weak target discrimination ability and poor anti-interference ability, making it difficult to guarantee the performance of downstream sensing tasks in complex environments.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a target recognition method that integrates polarization features, wherein the method includes: Channel state information on each subcarrier is collected synchronously using a dual-polarized antenna, and a polarization scattering matrix is ​​constructed based on the channel state information. The polarization scattering matrix is ​​decomposed using polarization target decomposition technology to obtain the polarization coherence vector matrix; The polarization coherence vector matrix is ​​transformed into a polarization word sequence, and the polarization word sequence is modeled through the Transformer layer of the wireless basic model to obtain a global feature vector. The global feature vector is input into the downstream task head to obtain the target recognition result.

[0006] In one implementation, the step of synchronously acquiring channel state information on each subcarrier using a dual-polarized antenna includes: A dual-polarized antenna with both horizontal and vertical polarization is used as the radio frequency transceiver node. Within a preset time period, data from each subcarrier channel is collected synchronously through the radio frequency transceiver node; Extract the channel state information corresponding to each subcarrier from the channel data.

[0007] In one implementation, constructing the polarization scattering matrix based on the channel state information includes: Based on the channel state information, the complex values ​​of the channel state information corresponding to horizontal transmission and horizontal reception, horizontal transmission and vertical reception, vertical transmission and horizontal reception, and vertical transmission and vertical reception are extracted in units of a single subcarrier and a single timestamp. The extracted complex values ​​of each channel state information are filled into the matrix according to their corresponding positions to obtain the polarization scattering matrix.

[0008] In one implementation, the step of decomposing the polarization scattering matrix using polarization target decomposition technology to obtain a polarization coherence vector matrix includes: The Pauli coherent target decomposition method is used to perform polarimetric target decomposition on the polarimetric scattering matrix to obtain the polarimetric coherent vector matrix.

[0009] In one implementation, converting the polarization coherence vector matrix into a polarization lexical sequence includes: Using the polarization coherence vector matrix as input, a block operation is performed in the time-frequency plane to divide the data into several micro-blocks of different sizes. By using a polarization sensing neural network, all physical information within each micro-block is mapped to the embedding space to obtain the corresponding embedding vectors. Based on the proportion of scattering components within each microblock, polarization coding is performed on each embedding vector to obtain each polarization-coded vector; The polarization coding vectors are integrated to generate a polarization word sequence.

[0010] In one implementation, the polarized lexical sequence is modeled through the Transformer layer of the wireless underlying model to obtain a global feature vector, including: The polarized lexical sequence is input into the Transformer layer of the wireless basic model, and the physical relationship between lexical units is modeled through a multi-head self-attention mechanism to obtain a global feature vector.

[0011] In one implementation, the target recognition result includes material recognition result, action recognition result, and state recognition result.

[0012] Secondly, embodiments of the present invention also provide a target recognition system that integrates polarization features, wherein the system includes: The polarization scattering matrix construction module is used to synchronously collect channel state information on each subcarrier using a dual-polarized antenna, and construct a polarization scattering matrix based on the channel state information. The polarization coherence vector matrix acquisition module is used to decompose the polarization scattering matrix using polarization target decomposition technology to obtain the polarization coherence vector matrix; The global feature vector acquisition module is used to convert the polarization coherence vector matrix into a polarization word sequence, and to model the polarization word sequence through the Transformer layer of the wireless basic model to obtain the global feature vector. The target recognition result acquisition module is used to input the global feature vector into the downstream task head to obtain the target recognition result.

[0013] Thirdly, embodiments of the present invention also provide a terminal, wherein the terminal includes a memory, a processor, and a target recognition program with fused polarization features stored in the memory and executable on the processor. When the processor executes the target recognition program with fused polarization features, it implements the steps of the target recognition method with fused polarization features as described in any of the above schemes.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a target recognition program with fused polarization features, and when the target recognition program with fused polarization features is executed by a processor, it implements the steps of the target recognition method with fused polarization features as described in any of the above schemes.

[0015] Beneficial Effects: This invention provides a target recognition method that integrates polarization features. Compared with existing technologies, this invention first utilizes dual-polarized antennas to synchronously collect channel state information on each subcarrier, thereby accurately capturing the multi-dimensional raw data such as polarization and time delay contained in the wireless signal during propagation. Then, based on the channel state information, a polarization scattering matrix is ​​constructed, fully preserving the polarization scattering characteristics and channel detail features of the wireless signal in the target scene. This provides comprehensive and accurate raw data support for subsequent feature mining and target recognition, avoiding the drawbacks of information loss and single-feature limitations of traditional single-polarization acquisition methods. Next, the polarization scattering matrix is ​​decomposed and analyzed using polarization target decomposition technology to remove redundant noise and interference components in the channel, extracting a polarization coherence vector matrix that can characterize the essential attributes of the target. This process purifies and reduces the dimensionality of the original matrix data, effectively filtering out environmental clutter, multipath interference, and other adverse factors, significantly improving the purity of effective features. Then, the polarization coherence vector matrix is ​​converted into a polarization word sequence adapted to the input of the wireless fundamental model. Leveraging the powerful global modeling and contextual learning capabilities of the wireless fundamental model's Transformer layer, deep feature mining is performed on the polarization word sequence to fully capture the long-distance correlations between different words and different subcarrier polarization features, extracting a global feature vector that combines local details and global correlation characteristics. This overcomes the limitation of traditional algorithms that can only extract shallow local features. Finally, the learned global feature vector is input into the downstream task head, and through task-specific feature mapping and classification logic, accurate target recognition results are efficiently output. Overall, this invention forms a complete technical chain from multi-dimensional channel information acquisition, polarization feature purification, deep global feature modeling to intelligent recognition output. It fully explores the hidden feature value of the polarization dimension of wireless signals, and realizes deep adaptive learning and global fusion of polarization features based on artificial intelligence models. It effectively solves the industry problems of traditional wireless target recognition technology, such as one-sided feature extraction, weak anti-interference ability, low recognition accuracy in complex scenarios, and poor generalization. It significantly improves the stability and accuracy of wireless target recognition in complex electromagnetic environments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a specific implementation of the target recognition method that integrates polarization features, as provided in an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating a preferred embodiment of the target recognition method that integrates polarization features provided in this invention.

[0018] Figure 3 This is a principle block diagram of the performance evaluation system for the hybrid power generation system provided in the embodiments of the present invention.

[0019] Figure 4 This is a block diagram illustrating the internal structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] Against the backdrop of the rapid development of IoT, human-computer interaction, and intelligent environmental sensing technologies, non-contact passive sensing technology, with its advantages of no device dependence, strong concealment, and wide adaptability, has become a core research direction in the field of intelligent sensing, and is widely used in many scenarios such as human behavior recognition, gesture perception, indoor positioning, and environmental monitoring. Radio frequency sensing technology based on Channel State Information (CSI) can capture the dynamic changes in amplitude and phase during the propagation of spatial electromagnetic waves, mapping the temporal and spatial variation characteristics of wireless signals to target location, human actions, and environmental feature information. This overcomes the limitations of traditional visual sensing, which is restricted by lighting, occlusion, and privacy, and possesses the advantages of all-weather, non-contact, and interference-resistant sensing, making it the mainstream technology solution in the current passive radio frequency sensing field. Traditional CSI radio frequency sensing technology mainly relies on radio frequency equipment to complete data acquisition. It then cleans the raw CSI data using preprocessing methods such as filtering and denoising, and phase calibration. Next, it extracts shallow physical features such as amplitude, phase, angle of arrival, and Doppler shift, and uses deep learning models such as CNN, RNN, and LSTM to establish a mapping relationship between features and sensing targets, achieving various sensing tasks. With the development of cross-modal fusion technology, existing mainstream optimization schemes have introduced a visual base model and a cross-modal comparative knowledge distillation mechanism. Using the visual model as the teacher model and the radio frequency (RF) model as the student model, semantic knowledge transfer from video images compensates for the lack of semantic features in RF signals, effectively improving the recognition accuracy and generalization ability of the RF sensing model. However, existing RF sensing technologies based on wireless base models still have significant technical shortcomings, and their limitations are becoming increasingly apparent. Specifically, current mainstream enhancement solutions rely entirely on cross-modal data such as video to drive the knowledge distillation process, imbuing radio frequency signals with high-level semantic information solely through visual semantics. This over-reliance on external modal data not only increases the cost of data acquisition and model training but also makes them susceptible to issues such as cross-modal data matching bias and poor scene adaptability. Furthermore, existing technologies generally ignore the polarization characteristics naturally carried during the propagation and scattering of radio electromagnetic waves. Polarization features, as inherent physical characteristics that can accurately distinguish targets of different materials and possess strong robustness, have not been effectively explored and utilized. This results in existing radio frequency sensing models having insufficient ability to distinguish target behavior in different materials and complex environments. The physical feature dimensions of the sensing models are singular, and their feature representation capabilities are limited, making it difficult to adapt to complex and ever-changing indoor and outdoor sensing scenarios. This significantly restricts the sensing accuracy, anti-interference capability, and scene adaptability of CSI radio frequency sensing technology, hindering its practical application and technological iteration in high-precision intelligent sensing scenarios.

[0022] To address the aforementioned issues, this embodiment provides a target recognition method that integrates polarization features. Specifically, this embodiment first utilizes a dual-polarized antenna to synchronously acquire channel state information on each subcarrier, thereby accurately capturing the multi-dimensional raw data, such as polarization and delay dimensions, inherent in the wireless signal during propagation. Then, based on the channel state information, a polarization scattering matrix is ​​constructed, fully preserving the polarization scattering characteristics and channel detail features of the wireless signal in the target scene. This provides comprehensive and accurate raw data support for subsequent feature mining and target recognition, avoiding the drawbacks of information loss and single-feature limitations inherent in traditional single-polarization acquisition methods. Next, the polarization scattering matrix is ​​decomposed and analyzed using polarization target decomposition technology to remove redundant noise and interference components from the channel, extracting the polarization coherence vector that characterizes the essential attributes of the target. The matrix is ​​used to purify and reduce the dimensionality of the original matrix data, effectively filtering out the influence of environmental clutter, multipath interference and other adverse factors, and significantly improving the purity of effective features. Then, the polarization coherence vector matrix is ​​converted into a polarization word sequence adapted to the input of the wireless basic model. With the powerful global modeling and context association learning capabilities of the Transformer layer of the wireless basic model, deep feature mining is performed on the polarization word sequence to fully capture the long-distance correlation between polarization features of different words and different subcarriers, and extract a global feature vector that has both local details and global correlation characteristics, breaking through the limitation of traditional algorithms that can only extract shallow local features. Finally, the learned global feature vector is input into the downstream task head, and through task-specific feature mapping and classification logic, the accurate target recognition results are efficiently output. Overall, this invention forms a complete technical chain from multi-dimensional channel information acquisition, polarization feature purification, deep global feature modeling to intelligent recognition output. It fully explores the hidden feature value of the polarization dimension of wireless signals, and realizes deep adaptive learning and global fusion of polarization features based on artificial intelligence models. It effectively solves the industry problems of traditional wireless target recognition technology, such as one-sided feature extraction, weak anti-interference ability, low recognition accuracy in complex scenarios, and poor generalization. It significantly improves the stability and accuracy of wireless target recognition in complex electromagnetic environments.

[0023] The target recognition method fused with polarization features provided in this embodiment can be applied to smart terminals, such as... Figure 1 As shown, the specific steps include the following: Step S100: Synchronously collect channel state information on each subcarrier using a dual-polarized antenna, and construct a polarization scattering matrix based on the channel state information.

[0024] In this embodiment, the channel state information of each subcarrier is synchronously collected by a dual-polarized antenna. The polarization scattering matrix is ​​constructed based on the collected channel state information. This method can completely preserve the propagation characteristics and spatial correlation characteristics of the channel in the polarization and subcarrier dimensions, providing complete and accurate basic matrix data support for subsequent channel feature analysis, polarization diversity utilization, and communication performance optimization. At the same time, the synchronous acquisition method can avoid channel information distortion caused by timing deviations, ensuring the authenticity and reliability of the polarization scattering matrix in representing the actual channel propagation characteristics.

[0025] Specifically, step S100 includes the following steps: Step S101: Use a dual-polarized antenna with horizontal and vertical polarization as the radio frequency transceiver node; Step S102: Within a preset time period, synchronously collect data from each subcarrier channel through the radio frequency transceiver node; Step S103: Extract the channel state information corresponding to each subcarrier from the channel data; Step S104: Based on the channel state information, extract the complex values ​​of the channel state information corresponding to horizontal transmission and horizontal reception, horizontal transmission and vertical reception, vertical transmission and horizontal reception, and vertical transmission and vertical reception, using a single subcarrier and a single timestamp as units. Step S105: Fill the extracted complex values ​​of each channel state information into the matrix according to their corresponding positions to obtain the polarization scattering matrix.

[0026] In one implementation, a dual-polarized antenna with both horizontal and vertical polarization is first used as the radio frequency transceiver node, such as... Figure 2 As shown, through the horizontal polarization on the radio frequency receiver ( Vertical polarization ( The antenna synchronously acquires channel data for each subcarrier within a preset time period. Then, it extracts the channel state information (CSI) corresponding to each subcarrier from the channel data. Using a single subcarrier f and a single timestamp t as units, it extracts the complex values ​​of the CSI for horizontal transmission / horizontal reception, horizontal transmission / vertical reception, vertical transmission / horizontal reception, and vertical transmission / vertical reception. The extracted complex values ​​of the CSI are then filled into a matrix according to their corresponding positions to construct a 2×2 polarization scattering matrix. : ,in, Represents the CSI complex value, where the first letter of the subscript indicates the transmit polarization direction and the second letter indicates the receive polarization direction (e.g., ...). (For horizontal transmission and vertical reception) For time indexing, Frequency subcarrier indexing fully preserves the cross-polarization characteristics of electromagnetic waves after reflection from the target, laying a data foundation for subsequent differentiation of object materials.

[0027] Step S200: Decompose the polarization scattering matrix using polarization target decomposition technology to obtain the polarization coherence vector matrix.

[0028] In this embodiment, as Figure 2 As shown, firstly, the Pauli coherent target decomposition method is used to construct the polarization scattering matrix. Polarization target decomposition is performed to transform the abstract complex matrix form into a vector space with clearly defined physical meaning. Specifically, the matrix is ​​decomposed... Transformed into polarization coherence vector matrix Its expression is: ,in, It is a three-dimensional complex vector; its components Represents single-path scattering (such as from a wall or torso plane); This represents quadratic angular scattering (such as the dihedral angle formed by limbs). This represents volume scattering or rotational reflection (such as from limb edges or rough surfaces). Through this physics-based polarization vector decomposition process, the radio frequency signal is transformed into "scattering type" features, significantly reducing the learning burden of subsequent base models on environmental multipath noise, and outputting a polarization coherence vector matrix. (in The total number of time frames and the total number of subcarriers (3 represents the three channels of the Pauli decomposition) are used as inputs for the subsequent lexicalization step.

[0029] Step S300: Convert the polarization coherence vector matrix into a polarization lexical sequence, and model the polarization lexical sequence through the Transformer layer of the wireless basic model to obtain a global feature vector.

[0030] In this embodiment, the acquired polarimetric coherence vector matrix is ​​converted into a polarimetric lexical sequence adapted to the input format of the wireless fundamental model. Global correlation modeling of the polarimetric lexical sequence is carried out with the help of the Transformer layer of the wireless fundamental model, which fully explores the implicit temporal and spatial correlation features within the polarimetric domain data. This allows for the capture of global dependencies of polarimetric information from a holistic perspective, effectively condensing global feature vectors with strong representational capabilities. This provides high-dimensional and robust feature support for subsequent downstream tasks such as polarimetric signal recognition, channel estimation, and state awareness.

[0031] Specifically, step S300 includes the following steps: Step S301: Using the polarization coherence vector matrix as input, perform a block operation on the time-frequency plane to divide the space into several micro-blocks of different sizes. Step S302: Using a polarization sensing neural network, all physical information within each micro-block is mapped to the embedding space to obtain the corresponding embedding vectors. Step S303: Based on the proportion of scattering components in each microblock, perform polarization coding on each embedding vector to obtain each polarization coded vector; Step S304: Integrate the polarization coding vectors to generate a polarization word sequence; Step S305: Input the polarized lexical sequence into the Transformer layer of the wireless basic model, and model the physical relationship between lexical units through a multi-head self-attention mechanism to obtain a global feature vector.

[0032] In one implementation, such as Figure 2 As shown, firstly, the polarimetric coherence vector matrix K is taken as input, which is a T×F×3 three-dimensional tensor composed of time-frequency polarimetric data. Then, patching is performed in the time-frequency plane, dividing the matrix into several micro-patterns of size t×f. Each patch contains a set of local polarimetric scattering evolution information. Next, a polar-aware DNN maps all the physical information within each patch to a D-dimensional embedding space, obtaining the corresponding embedding vectors. Then, based on the proportion of scattering components within each patch, polarimetric encoding is performed on each embedding vector to explicitly identify the physical affiliation of the signal block (e.g., plane reflection, secondary scattering, volume scattering, etc.), obtaining each polarimetric encoded vector. Finally, the polarimetric encoded vectors are integrated to generate a polarimetric term sequence. This process bridges the gap between "physical quantities" and "semantic tokens," giving tokens a clear physical meaning and laying the foundation for subsequent processing. Then, the polarized token sequence is input into the Transformer layer of the wireless foundation model. A multi-head self-attention mechanism models the physical relationships between tokens: on one hand, it identifies token regions with "single specular reflection" polarization and highly stable phase, classifying them as static environments (such as walls or furniture), and suppressing them in the attention weight distribution; on the other hand, for tokens with complex depolarization effects, it captures the dynamic relationship between volume scattering and secondary scattering components, accurately filtering signals caused solely by human reflection from a cluttered background. After layers of attention filtering and aggregation, the wireless foundation model finally outputs a vector representing global features, achieving precise filtering and feature extraction of reflection signals from specific targets.

[0033] Step S400: Input the global feature vector into the downstream task head to obtain the target recognition result.

[0034] In this embodiment, as Figure 2 As shown, firstly, the global feature vector, which incorporates polarization features, is input into different downstream task heads (including classification heads) to support different perception tasks and obtain target recognition results. These results include material recognition, action recognition, and state recognition. Material recognition utilizes the differences in electromagnetic polarization response contained in the token to distinguish objects with similar physical shapes but different materials (such as wooden furniture and metal equipment). By extracting the polarization attribute ratio from the global vector, it can effectively distinguish objects with similar shapes but different materials. Action recognition combines the dynamic changes in polarization information to classify human actions. Due to the elimination of environmental interference, the accuracy of action recognition is significantly improved. State detection targets long-term stillness or specific postures of the target, utilizing the stable "physical fingerprint" exhibited by polarization features. By monitoring the stability of the global feature vector over a long period, it achieves the detection of the target's activity state or positional shift. By empowering global feature vectors that integrate polarization information to various downstream sensing tasks, the polarization features are fully utilized to represent materials, dynamic changes, and long-term states. From material differentiation and action recognition to state monitoring, the robustness and task adaptability of the sensing system in complex scenarios are comprehensively improved, providing more comprehensive and accurate information support for various sensing applications.

[0035] In summary, this embodiment first utilizes a dual-polarized antenna to synchronously acquire channel state information on each subcarrier, thereby accurately capturing the multi-dimensional raw data, such as polarization and delay dimensions, contained in the wireless signal during propagation. Then, based on the channel state information, a polarization scattering matrix is ​​constructed, fully preserving the polarization scattering characteristics and channel detail features of the wireless signal in the target scene. This provides comprehensive and accurate raw data support for subsequent feature mining and target identification, avoiding the drawbacks of information loss and single-feature limitations inherent in traditional single-polarization acquisition methods. Next, the polarization scattering matrix is ​​decomposed and analyzed using polarization target decomposition technology, stripping away redundant noise and interference components in the channel and extracting the polarization coherence vector matrix that characterizes the essential attributes of the target. This achieves purification and feature extraction of the original matrix data. Feature reduction effectively filters out the influence of environmental clutter, multipath interference, and other adverse factors, significantly improving the purity of effective features. Then, the polarization coherence vector matrix is ​​converted into a polarization word sequence adapted to the input of the wireless fundamental model. Leveraging the powerful global modeling and contextual learning capabilities of the wireless fundamental model's Transformer layer, deep feature mining is performed on the polarization word sequence to fully capture the long-distance correlation between polarization features of different words and different subcarriers. This extracts a global feature vector that combines local details with global correlation characteristics, breaking through the limitation of traditional algorithms that can only extract shallow local features. Finally, the learned global feature vector is input into the downstream task head, and through task-specific feature mapping and classification logic, accurate target recognition results are efficiently output. Overall, this invention forms a complete technical chain from multi-dimensional channel information acquisition, polarization feature purification, deep global feature modeling to intelligent recognition output. It fully explores the hidden feature value of the polarization dimension of wireless signals, and realizes deep adaptive learning and global fusion of polarization features based on artificial intelligence models. It effectively solves the industry problems of traditional wireless target recognition technology, such as one-sided feature extraction, weak anti-interference ability, low recognition accuracy in complex scenarios, and poor generalization. It significantly improves the stability and accuracy of wireless target recognition in complex electromagnetic environments.

[0036] like Figure 3As shown in the illustration, this embodiment also provides a target recognition system that integrates polarization features. The system includes: a polarization scattering matrix construction module 10, a polarization coherence vector matrix acquisition module 20, a global feature vector acquisition module 30, and a target recognition result acquisition module 40. Specifically, the polarization scattering matrix construction module 10 is used to synchronously collect channel state information on each subcarrier using a dual-polarized antenna, and construct a polarization scattering matrix based on the channel state information. The polarization coherence vector matrix acquisition module 20 is used to decompose the polarization scattering matrix using polarization target decomposition technology to obtain a polarization coherence vector matrix. The global feature vector acquisition module 30 is used to convert the polarization coherence vector matrix into a polarization word sequence, and model the polarization word sequence through the Transformer layer of the wireless basic model to obtain a global feature vector. The target recognition result acquisition module 40 is used to input the global feature vector into a downstream task head to obtain a target recognition result, wherein the target recognition result includes material recognition result, action recognition result, and state recognition result.

[0037] In one implementation, the polarization scattering matrix construction module 10 includes: The standardized data acquisition unit is used to perform time synchronization, outlier removal, and filtering on the running data to obtain standardized data.

[0038] In one implementation, the reference value acquisition module 20 includes: The radio frequency transceiver node acquisition unit is used to employ a dual-polarized antenna containing horizontal and vertical polarization as a radio frequency transceiver node. The channel data acquisition unit is used to synchronously acquire channel data of each subcarrier through the radio frequency transceiver node within a preset time period; The channel state information acquisition unit is used to extract the channel state information corresponding to each subcarrier from the channel data; The channel state information complex value acquisition unit is used to extract the channel state information complex values ​​corresponding to horizontal transmission and horizontal reception, horizontal transmission and vertical reception, vertical transmission and horizontal reception, and vertical transmission and vertical reception based on the channel state information, using a single subcarrier and a single timestamp as units. The polarization scattering matrix acquisition unit is used to fill the extracted complex values ​​of each channel state information into the matrix according to their corresponding positions to obtain the polarization scattering matrix.

[0039] In one implementation, the polarization coherence vector matrix acquisition module 20 includes: The polarization coherence vector matrix acquisition unit is used to perform polarization target decomposition on the polarization scattering matrix using the Pauli coherence target decomposition method to obtain the polarization coherence vector matrix.

[0040] In one implementation, the global feature vector acquisition module 30 includes: The micro-block acquisition unit is used to perform block operation in the time-frequency plane with the polarization coherence vector matrix as input, and divide it into micro-blocks of several sizes; Each embedding vector acquisition unit is used to map all the physical information in each micro-block to the embedding space through the polarization sensing neural network to obtain the corresponding embedding vectors. Each polarization coding vector acquisition unit is used to perform polarization coding on each embedding vector according to the proportion of scattering components in each micro-block, so as to obtain each polarization coding vector; The polarization lexical sequence generation unit is used to integrate and process the various polarization coding vectors to generate a polarization lexical sequence; The global feature vector acquisition unit is used to input the polarized lexical sequence into the Transformer layer of the wireless basic model, and to model the physical relationship between lexical units through a multi-head self-attention mechanism to obtain the global feature vector.

[0041] The working principle of each module in the target recognition system that integrates polarization features in this embodiment is the same as that of each step in the above method embodiment, and will not be repeated here.

[0042] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 4 As shown. The terminal may include one or more processors 100 ( Figure 4 (Only one is shown in the image), memory 101, and a computer program 102 stored in memory 101 and executable on one or more processors 100, such as a target recognition program that fuses polarization features. When one or more processors 100 execute computer program 102, they can implement the various steps in the target recognition method embodiment that fuses polarization features. Alternatively, when one or more processors 100 execute computer program 102, they can implement the functions of various modules / units in the target recognition method embodiment that fuses polarization features, which is not limited here.

[0043] In one embodiment, the processor 100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0044] In one embodiment, memory 101 may be an internal storage unit of an electronic device, such as a hard drive or RAM. Memory 101 may also be an external storage device of the electronic device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, memory 101 may include both internal and external storage units. Memory 101 is used to store computer programs and other programs and data required by the terminal. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0045] Those skilled in the art will understand that Figure 4 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0046] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, operational databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual operating data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target recognition method that integrates polarization features, characterized in that, The method includes: Channel state information on each subcarrier is collected synchronously using a dual-polarized antenna, and a polarization scattering matrix is ​​constructed based on the channel state information. The polarization scattering matrix is ​​decomposed using polarization target decomposition technology to obtain the polarization coherence vector matrix; The polarization coherence vector matrix is ​​transformed into a polarization word sequence, and the polarization word sequence is modeled through the Transformer layer of the wireless basic model to obtain a global feature vector. The global feature vector is input into the downstream task head to obtain the target recognition result.

2. The target recognition method based on fused polarization features according to claim 1, characterized in that, The method of synchronously acquiring channel state information on each subcarrier using a dual-polarized antenna includes: A dual-polarized antenna with both horizontal and vertical polarization is used as the radio frequency transceiver node. Within a preset time period, data from each subcarrier channel is collected synchronously through the radio frequency transceiver node; Extract the channel state information corresponding to each subcarrier from the channel data.

3. The target recognition method based on fused polarization features according to claim 1, characterized in that, The construction of the polarization scattering matrix based on the channel state information includes: Based on the channel state information, the complex values ​​of the channel state information corresponding to horizontal transmission and horizontal reception, horizontal transmission and vertical reception, vertical transmission and horizontal reception, and vertical transmission and vertical reception are extracted in units of a single subcarrier and a single timestamp. The extracted complex values ​​of each channel state information are filled into the matrix according to their corresponding positions to obtain the polarization scattering matrix.

4. The target recognition method based on fused polarization features according to claim 1, characterized in that, The process of decomposing the polarization scattering matrix using polarization target decomposition techniques to obtain the polarization coherence vector matrix includes: The Pauli coherent target decomposition method is used to perform polarimetric target decomposition on the polarimetric scattering matrix to obtain the polarimetric coherent vector matrix.

5. The target recognition method based on fused polarization features according to claim 1, characterized in that, The step of converting the polarization coherence vector matrix into a polarization lexical sequence includes: Using the polarization coherence vector matrix as input, a block operation is performed in the time-frequency plane to divide the data into several micro-blocks of different sizes. By using a polarization sensing neural network, all physical information within each micro-block is mapped to the embedding space to obtain the corresponding embedding vectors. Based on the proportion of scattering components within each microblock, polarization coding is performed on each embedding vector to obtain each polarization-coded vector; The polarization coding vectors are integrated to generate a polarization word sequence.

6. The target recognition method based on fused polarization features according to claim 5, characterized in that, The polarized lexical sequence is modeled using the Transformer layer of the wireless basic model to obtain a global feature vector, including: The polarized lexical sequence is input into the Transformer layer of the wireless basic model, and the physical relationship between lexical units is modeled through a multi-head self-attention mechanism to obtain a global feature vector.

7. The target recognition method based on fused polarization features according to claim 1, characterized in that, The target recognition results include material recognition results, motion recognition results, and state recognition results.

8. A target recognition system that integrates polarization features, characterized in that, The system includes: The polarization scattering matrix construction module is used to synchronously collect channel state information on each subcarrier using a dual-polarized antenna, and construct a polarization scattering matrix based on the channel state information. The polarization coherence vector matrix acquisition module is used to decompose the polarization scattering matrix using polarization target decomposition technology to obtain the polarization coherence vector matrix; The global feature vector acquisition module is used to convert the polarization coherence vector matrix into a polarization word sequence, and to model the polarization word sequence through the Transformer layer of the wireless basic model to obtain the global feature vector. The target recognition result acquisition module is used to input the global feature vector into the downstream task head to obtain the target recognition result.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a target recognition program with fused polarization features stored in the memory and executable on the processor. When the processor executes the target recognition program with fused polarization features, it implements the steps of the target recognition method with fused polarization features as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a target recognition program that incorporates fused polarization features. When the target recognition program that incorporates fused polarization features is executed by a processor, it implements the steps of the target recognition method that incorporates fused polarization features as described in any one of claims 1-7.

Images