Human body feature sensing method and system based on large model and sensing terminal, and medium
By collecting and preprocessing channel state information through multi-source sensing terminals, and combining large models and transfer learning techniques, a cross-individual feature perception model is trained, which solves the problem of insufficient cross-individual generalization ability of traditional small models and achieves efficient and accurate human feature perception on new individuals.
Patent Information
- Application Number
- CN202511496889.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-20
- Publication Date
- 2026-01-27
AI Technical Summary
Traditional small-model methods based on WiFi CSI have shortcomings in cross-individual generalization ability and recognition of general human features. They cannot effectively utilize the generalization advantage of large models to achieve fine-grained human feature perception. Furthermore, existing models are not robust enough to dynamic interference and are difficult to apply in complex environments.
The message channel state information is collected by a multi-source sensing terminal. After preprocessing, a fine-tuned large model is used for perception to train a cross-individual feature perception large model. The pre-training capability of the large model is used to extract common features from a large amount of individual data, reducing the dependence on individual-specific data. Cross-individual common features are extracted through CFR encoder and Transformer coding layer.
It achieves good generalization performance on new individuals and efficient and accurate human feature perception, making it suitable for fields such as smart healthcare and sports science.
Smart Images
Figure CN121412640A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large model perception technology, specifically to human feature perception methods, systems, and media based on large models and sensing terminals. Background Technology
[0002] Wireless sensing technology is gradually penetrating core aspects of human life, driven by urgent needs in areas such as smart homes, medical monitoring, and security emergency response, including applications like contactless health monitoring, gesture interaction, and fall detection. In complex indoor environments, sensing technologies based on wireless signals (such as Wi-Fi and millimeter-wave radar) have become a research hotspot due to their advantages such as no need for wearable devices and privacy protection. Currently, Wi-Fi-based wireless sensing technology is attracting significant attention due to its ubiquitous device availability and environmental reusability. Existing smartphones, IoT terminals, and other devices generally support Wi-Fi communication modules, allowing for direct detection of human behavior using their Channel State Information (CSI) or Received Signal Strength Indicator (RSSI), without the need for additional dedicated hardware.
[0003] Currently, Received Signal Strength Indication (RSSI)-based wireless sensing technology is a common method for environmental perception and behavior recognition. This technology relies on the target device periodically broadcasting Wi-Fi probe request frames or actively sending data packets to obtain signal strength information. However, RSSI is susceptible to multipath interference, obstacle obstruction, and environmental changes, with signal strength fluctuations exceeding 10 dB, leading to a high false positive rate in behavior recognition. Therefore, RSSI-based sensing schemes face significant challenges in terms of generalization and practicality. Furthermore, RSSI data has a limited data dimension and coarse granularity. In contrast, Channel State Information (CSI)-based wireless sensing technology offers a superior solution to these problems. CSI provides more refined and multi-dimensional signal characteristics than RSSI, containing rich phase and amplitude information, effectively capturing subtle environmental changes and target behavioral characteristics. Because CSI is sensitive to multipath effects, it can utilize detailed information from multipath propagation to enhance sensing accuracy, thereby significantly reducing the impact of environmental interference. In addition, CSI data has finer granularity and higher dimensionality, making it suitable for high-precision human perception in complex scenarios.
[0004] Current small-scale wireless sensing models based on CSI face significant bottlenecks in generalization capabilities: limited model architecture capacity makes it difficult to effectively capture the nonlinear correlation between complex human features and multipath signals, leading to decreased recognition accuracy during cross-scene migration; their heavy reliance on manual feature engineering (such as manually designing time-frequency domain features) makes them prone to losing key signal features when the environment changes; furthermore, existing models lack robustness to dynamic interference (such as moving obstacles and sudden noise) and lack a unified multi-task learning framework, requiring independent modeling for different sensing tasks. These issues restrict the large-scale application of wireless sensing technology in real-world complex environments.
[0005] Large-scale cue word engineering and full-parameter training have significant limitations in the field of wireless sensing. Cue word engineering relies heavily on human experience for design, making it difficult to systematically capture the complex physical relationships between wireless signals and human behavior. Full-parameter training faces the dual challenges of computational efficiency and deployment feasibility: models with billions of parameters require training on kilocalorie-level clusters, consuming over 400 times more energy than traditional small models, and the model size expands to over 300GB after convergence, making it difficult to embed in resource-constrained IoT terminals. Furthermore, full-parameter updates lack modular scalability, severely hindering the efficiency of engineering implementation. Summary of the Invention
[0006] The technical problem this invention aims to solve is that traditional small-model methods based on WiFi CSI have shortcomings in cross-individual generalization ability and recognition of general human features (such as gender and body weight), and cannot effectively utilize the generalization advantage of large models to achieve fine-grained human feature perception. The purpose of this invention is to provide a human feature perception method, system, and medium based on large models and sensing terminals. By collecting message channel state information through multi-source sensing terminals, and using a fine-tuned large model for perception after preprocessing, it can effectively utilize wireless network channel state information for human feature perception and achieve cross-individual generalization. Traditional human feature perception methods based on wireless network channel state information usually rely on data from specific individuals for training, resulting in poor model performance on new individuals. However, this invention, through pre-training a large model, can extract common features from a large amount of wireless network channel state information data from different individuals, thereby achieving better generalization performance on new individuals and reducing dependence on individual-specific data.
[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:
[0008] This solution provides a human feature perception method based on large models and sensing terminals, including:
[0009] Continuously collect raw message channel state information from different individuals through authorized channels;
[0010] The original message channel state information is preprocessed to obtain the original training dataset.
[0011] A large-scale model for cross-individual feature perception was trained based on the original dataset.
[0012] The real-time message channel state information of the individual under test is collected in real time. The real-time message channel state information is preprocessed and encoded into text format data, which is then input into the cross-individual feature perception big model to obtain the human body features of the individual under test.
[0013] A further optimized scheme is as follows: the original message channel state information includes timing signals, image data, and text data; the timing signals include: heart rate, gait, action sequence, and speech waveform; the image data includes: human posture video, facial expression video, and behavior video; the text data includes: user self-reported status and environmental context description.
[0014] A further optimization scheme involves preprocessing the original message channel state information to obtain the original training dataset, including the following methods:
[0015] The Hampel filtering method is used to detect and replace outliers in each spatial stream using a sliding window.
[0016] Apply two-dimensional median filtering to the Hampel-filtered data in the subcarrier-time two-dimensional plane;
[0017] The data from all spatial streams within each sampling period are concatenated according to the spatial stream dimension to form a tensor group containing four spatial streams.
[0018] A further optimization involves training a large-scale cross-individual feature perception model based on the original dataset, including the following methods:
[0019] The individual features in the preprocessed original training dataset are encoded into a unified representation dataset based on the CFR encoder.
[0020] Using a pre-trained large model as the base model, the base model is fine-tuned based on a unified representation dataset and efficient parameter fine-tuning techniques to obtain a cross-individual feature perception large model.
[0021] A further optimization scheme involves encoding individual features from the original training dataset into a unified representation dataset based on the CFR encoder; including the following methods:
[0022] Data from individual k Mapped to a common feature space Z, the features of which are unrelated to individual identity;
[0023] Based on the encoder from the data The preliminary features of individual k are extracted;
[0024] Based on the Transformer encoding layer, cross-individual common features of individual k are extracted from the initial features. During the process, a domain discriminator is introduced. The system distinguishes which individual features originate from and learns invariant features of that individual by minimizing a maximum game, thus enabling the domain discriminator to... The inability to distinguish the individuals from which the characteristics originate makes it possible to identify common characteristics across individuals. It becomes a common characteristic that is irrelevant to individuals.
[0025] A further optimized solution is that the minimization of the maximum game is represented as:
[0026] ;
[0027] in, The expectation operator represents the cross-individual common characteristics across all individuals. Expectation calculation; log represents exponentiation; domain discriminator Common characteristics across individuals The output of .
[0028] A further optimization scheme is as follows: the pre-trained large model includes: an embedding layer, a self-attention module, and a multivariate perception layer; the embedding layer is used to obtain a high-dimensional representation of the unified representation dataset; the self-attention module and the multivariate perception layer are used to extract features from the high-dimensional representation of the unified representation dataset;
[0029] The embedding layers are distributed at the beginning and end of the pre-trained large model, and the self-attention module and the multivariate perception layer are alternately set between the embedding layers on both the beginning and end sides.
[0030] A further optimization scheme is proposed, in which the fine-tuning method for the basic large model includes:
[0031] Prepare a subject weight matrix and a trainable matrix for each individual;
[0032] Insert adapters into certain layers of the basic large model;
[0033] Freeze the principal weight matrix in the base model and introduce a low-rank increment ΔW to adjust the trainable matrix in the base model.
[0034] A further optimization is that the operation of the adapter includes:
[0035] h′=h+W up ·σ(W down ·h);
[0036] Where h represents the input feature vector of the layer where the adapter is located, W down Let represent the dimension reduction matrix, with dimensions r×d, where r < d, used to project the input features of the adapter layer from d dimensions to r dimensions; σ() represents the non-linear activation function; Wup represents the dimension increase matrix, with dimensions d×r, used to project the dimension-reduced input features from r dimensions to d dimensions; h′ represents the output feature vector of the adapter.
[0037] A further optimization scheme involves freezing the main weight matrix and introducing a low-rank weight increment ΔW to adjust the trainable matrix; including the following methods:
[0038] The trainable matrix is adjusted according to the following formula:
[0039] ;
[0040] Where x represents the input vector of the layer containing the adapter; W represents the original trainable matrix of the layer containing the adapter; A represents the matrix projected onto the low-dimensional space; B represents the matrix projected back onto the output space; and h′ represents the output feature vector of the adapter.
[0041] This solution provides a human feature perception system based on a large model and a sensing terminal, used to implement the aforementioned human feature perception method based on a large model and a sensing terminal. The system includes:
[0042] The acquisition module is used to continuously acquire raw message channel state information from different individuals from authorized channels;
[0043] The preprocessing module is used to preprocess the original message channel state information to obtain the original training dataset;
[0044] The training module is used to train a large cross-individual feature perception model based on the original dataset;
[0045] The perception module is used to collect real-time message channel status information of the individual under test in real time, preprocess the real-time message channel status information, encode it into text format data, and then input it into the cross-individual feature perception big model to obtain the human body features of the individual under test.
[0046] This solution also provides a computer-readable medium storing a computer program that, when executed by a processor, can implement the human feature perception method based on a large model and a sensing terminal as described above.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] 1. This invention provides a method, system, and medium for human feature perception based on a large model and sensing terminals. It collects message channel state information through multi-source sensing terminals, preprocesses it, and then uses a finely tuned large model for perception. This effectively utilizes wireless network channel state information for human feature perception and achieves cross-individual generalization. Traditional human feature perception methods based on wireless network channel state information typically rely on data from specific individuals for training, resulting in poor model performance on new individuals. In contrast, this invention, through pre-training a large model, can extract common features from a large amount of wireless network channel state information data from different individuals, thereby achieving better generalization performance on new individuals and reducing dependence on individual-specific data.
[0049] 2. This invention provides a method, system, and medium for human feature perception based on a large model and sensing terminals. It trains a large model with cross-individual recognition capabilities offline, and then rapidly adapts to new individuals online using lightweight adaptation technology, achieving efficient and accurate human feature perception. This method combines multimodal sensing data, a large model, and transfer learning techniques, making it applicable to multiple fields such as smart healthcare and sports science. Attached Figure Description
[0050] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:
[0051] Figure 1 A method for human feature perception based on large models and sensing terminals;
[0052] Figure 2 A schematic diagram illustrating the principle of human feature perception based on a large model and a sensing terminal;
[0053] Figure 3 This is a human feature perception system based on a large model and a sensing terminal. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.
[0055] Traditional small-model methods based on WiFi CSI have shortcomings in cross-individual generalization ability and recognition of general human features (such as gender and body weight), and cannot effectively utilize the generalization advantage of large models to achieve fine-grained human feature perception. In view of this, this solution provides the following embodiments to solve the above-mentioned technical problems.
[0056] Example 1: This example provides a human feature perception method based on a large model and a sensing terminal, such as... Figure 1 As shown, the method includes:
[0057] Step 1: Continuously collect raw message channel status information from different individuals through authorized channels. The raw message channel status information in this step includes timing signals, image data, and text data. The timing signals include heart rate, gait, action sequences, and speech waveforms. The image data includes human posture videos, facial expression videos, and behavioral videos. The text data includes user self-reported status and environmental context descriptions.
[0058] Step two: Preprocess the original message channel state information to obtain the original training dataset; such as... Figure 2 As shown, this step specifically includes the following methods:
[0059] The Hampel filtering method is used to detect and replace outliers in each spatial stream using a sliding window.
[0060] Apply two-dimensional median filtering to the Hampel-filtered data in the subcarrier-time two-dimensional plane;
[0061] The data from all spatial streams within each sampling period are concatenated according to the spatial stream dimension to form a tensor group containing four spatial streams.
[0062] In this embodiment, the preprocessing process is also directly implemented by the CFR architecture.
[0063] Step 3: Train a large-scale cross-individual feature perception model based on the original dataset; this step specifically includes the following methods:
[0064] S31, based on the CFR encoder, encodes the individual features in the preprocessed original training dataset into a unified representation dataset; this step specifically includes the following methods:
[0065] Data from individual k Mapped to a common feature space Z, the features of which are unrelated to individual identity;
[0066] Based on the encoder from the data The preliminary features of individual k are extracted;
[0067] Based on the Transformer encoding layer, cross-individual common features of individual k are extracted from the initial features. During the process, a domain discriminator is introduced. The system distinguishes which individual features originate from and learns invariant features of that individual by minimizing a maximum game, thus enabling the domain discriminator to... The inability to distinguish the individuals from which the characteristics originate makes it possible to identify common characteristics across individuals. It becomes a common characteristic that is irrelevant to individuals.
[0068] The minimization-maximum game is represented as follows:
[0069] ;
[0070] in, The expectation operator represents the cross-individual common characteristics across all individuals. Expectation calculation; log represents exponentiation; domain discriminator Common characteristics across individuals The output of .
[0071] S32, using a pre-trained large model as the base large model, the base large model is fine-tuned based on a unified representation dataset and efficient parameter fine-tuning techniques to obtain a cross-individual feature perception large model.
[0072] The pre-trained large model includes: an embedding layer, a self-attention module, and a multivariate perception layer; the embedding layer is used to obtain a high-dimensional representation of the unified representation dataset; the self-attention module and the multivariate perception layer are used to extract features from the high-dimensional representation of the unified representation dataset;
[0073] Embedding layers are distributed at the beginning and end of the pre-trained large model, while self-attention modules and multivariate perception layers are alternately set between the embedding layers on both the beginning and end sides.
[0074] Fine-tuning methods for the basic large model include:
[0075] S321, prepare a main weight matrix and a trainable matrix for each individual;
[0076] S322, Insert an adapter into a portion of the layers in the base model; the operation of the adapter includes:
[0077] h′=h+W up ·σ(W down ·h);
[0078] Where h represents the input feature vector of the layer where the adapter is located, W down Let represent the dimension reduction matrix, with dimensions r×d, where r < d, used to project the input features of the adapter layer from d dimensions to r dimensions; σ() represents the non-linear activation function; Wup represents the dimension increase matrix, with dimensions d×r, used to project the dimension-reduced input features from r dimensions to d dimensions; h′ represents the output feature vector of the adapter.
[0079] A layer adapter inserts a small neural network (usually a bottleneck structure) into a layer of the base model. This network is first reduced in dimensionality and then increased in dimensionality, and then added to the original input via residual connections. In this way, the layer adapter introduces only a small number of parameters (controlled by r) yet can adjust the feature representation. During fine-tuning, we fix the parameters of the pre-trained model and only train the parameters of the adapter (i.e., W). up and W down ) and possible LayerNorm parameters.
[0080] S323 freezes the main weight matrix in the basic large model and introduces a low-rank increment ΔW to adjust the trainable matrix in the basic large model.
[0081] Step four involves real-time acquisition of the real-time message channel state information of the individual under test, preprocessing the real-time message channel state information, encoding it into text format data, and then inputting it into the cross-individual feature perception big data model to obtain the human body features of the individual under test. This step specifically includes the following methods:
[0082] The trainable matrix is adjusted according to the following formula:
[0083] ;
[0084] Where x represents the input vector of the layer containing the adapter; W represents the original trainable matrix of the layer containing the adapter; A represents the matrix projected onto the low-dimensional space; B represents the matrix projected back onto the output space; and h′ represents the output feature vector of the adapter.
[0085] Example 2:
[0086] This embodiment provides a human feature perception system based on a large model and a sensing terminal, used to implement the human feature perception method based on a large model and a sensing terminal described in Embodiment 1, such as... Figure 3 As shown, the system includes:
[0087] The acquisition module is used to continuously acquire raw message channel status information of different individuals from authorized channels; the specific hardware devices of the acquisition module include: CSI transceiver, sensed target, and server.
[0088] The CSI transceiver is used to collect Wi-Fi packets and analyze the received CSI data.
[0089] The preprocessing module is used to preprocess the original message channel state information to obtain the original training dataset; based on the preprocessing module, the perceived CSI information of each individual is preprocessed to remove outliers and reduce noise interference.
[0090] The training module is used to train a large cross-individual feature perception model based on the original dataset;
[0091] The perception module is used to collect real-time message channel status information of the individual under test in real time, preprocess the real-time message channel status information, encode it into text format data, and then input it into the cross-individual feature perception big model to obtain the human body features of the individual under test.
[0092] Example 3: This example provides a computer-readable medium storing a computer program, which, when executed by a processor, can implement the human feature perception method based on a large model and a sensing terminal as described in Example 1; specifically, the following steps are performed:
[0093] Step 1: Continuously collect raw message channel state information from different individuals using the authorized channels;
[0094] Step 2: Preprocess the original message channel state information to obtain the original training dataset;
[0095] Step 3: Train a large-scale cross-individual feature perception model based on the original dataset;
[0096] Step four: Real-time message channel state information of the individual to be tested is collected, the real-time message channel state information is preprocessed and encoded into text format data, and then input into the cross-individual feature perception big model to obtain the human body features of the individual to be tested.
[0097] like Figure 2 As shown, the entire process of the human feature perception method based on large-scale model and sensing terminal can be divided into two parts: an offline stage and an online stage. In the offline stage, data is first collected from the target area using acquisition devices to capture message channel state information passing through people. Next, the collected raw message channel state information undergoes preprocessing, including cleaning, filtering, and smoothing steps, to remove noise and irrelevant data and improve data quality. The preprocessed data is reshaped and encoded into a format suitable for processing by the cross-individual feature perception large-scale model (LLM). Subsequently, a pre-trained LLM model is used as the base model for fine-tuning. Reparameterization fine-tuning techniques are used to make the model more adaptable to the specific task, ultimately resulting in a fine-tuned cross-individual feature perception large-scale model. In the online stage, message channel state information of the target individual is collected in real time, and the message channel state information is preprocessed to ensure data quality and stability. The preprocessed real-time data is encoded into a format suitable for cross-individual feature perception large-scale model processing, and then the fine-tuned cross-individual feature perception large-scale model is used to identify human features from the real-time data. The entire process, through offline data collection and model fine-tuning, and online real-time data processing and feature recognition, ensures efficient system operation and accurate feature recognition.
[0098] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A human feature perception method based on a large model and a sensing terminal, characterized in that, The methods include: Continuously collect raw message channel state information from different individuals through authorized channels; The original message channel state information is preprocessed to obtain the original training dataset. A large-scale model for cross-individual feature perception was trained based on the original dataset. The real-time message channel state information of the individual under test is collected in real time. The real-time message channel state information is preprocessed and encoded into text format data, which is then input into the cross-individual feature perception big model to obtain the human body features of the individual under test.
2. The human feature perception method based on a large model and a sensing terminal according to claim 1, characterized in that, The original message channel state information includes timing signals, image data, and text data; the timing signals include: heart rate, gait, action sequence, and speech waveform; the image data includes: human posture video, facial expression video, and behavior video; the text data includes: user self-reported status and environmental context description; The original training dataset is obtained by preprocessing the original message channel state information, including the following methods: The Hampel filtering method is used to detect and replace outliers in each spatial stream using a sliding window. Apply two-dimensional median filtering to the Hampel-filtered data in the subcarrier-time two-dimensional plane; The data from all spatial streams within each sampling period are concatenated according to the spatial stream dimension to form a tensor group containing four spatial streams.
3. The human feature perception method based on a large model and a sensing terminal according to claim 1, characterized in that, A large-scale cross-individual feature perception model was trained based on the original dataset, including the following methods: The CFR encoder encodes the individual features in the preprocessed original training dataset into a unified representation dataset; data from individual k... Mapped to a common feature space Z, the features of which are independent of individual identity; based on the encoder, from the data... The preliminary features of individual k are extracted; and the cross-individual common features of individual k are extracted from the preliminary features based on the Transformer encoding layer. During the process, a domain discriminator is introduced. The system distinguishes which individual features originate from and learns invariant features of that individual by minimizing a maximum game, thus enabling the domain discriminator to... The inability to distinguish the individuals from which the characteristics originate makes it possible to identify common characteristics across individuals. It becomes a common characteristic that is irrelevant to individuals; Using a pre-trained large model as the base model, the base model is fine-tuned based on a unified representation dataset and efficient parameter fine-tuning techniques to obtain a cross-individual feature perception large model.
4. The human feature perception method based on a large model and a sensing terminal according to claim 3, characterized in that, The minimization-maximum game is represented as follows: ； in, The expectation operator represents the cross-individual common characteristics across all individuals. Expectation calculation; log represents exponentiation; domain discriminator Common characteristics across individuals The output of .
5. The human feature perception method based on a large model and a sensing terminal according to claim 3, characterized in that, The pre-trained large model includes: an embedding layer, a self-attention module, and a multivariate perception layer; the embedding layer is used to obtain a high-dimensional representation of the unified representation dataset; the self-attention module and the multivariate perception layer are used to extract features from the high-dimensional representation of the unified representation dataset; The embedding layers are distributed at the beginning and end of the pre-trained large model, and the self-attention module and the multivariate perception layer are alternately set between the embedding layers on both the beginning and end sides.
6. The human feature perception method based on a large model and a sensing terminal according to claim 3, characterized in that, The fine-tuning methods for the basic large model include: Prepare a subject weight matrix and a trainable matrix for each individual; Insert adapters into certain layers of the basic large model; Freeze the principal weight matrix in the base model and introduce a low-rank increment ΔW to adjust the trainable matrix in the base model.
7. The human feature perception method based on a large model and a sensing terminal according to claim 6, characterized in that, The operation of the adapter includes: h′=h+W up ·σ(W down ·h); Where h represents the input feature vector of the layer where the adapter is located, and W down Let represent the dimension reduction matrix, with dimensions r×d, where r < d, used to project the input features of the adapter layer from d dimensions to r dimensions; σ() represents the non-linear activation function; Wup represents the dimension increase matrix, with dimensions d×r, used to project the dimension-reduced input features from r dimensions to d dimensions; h′ represents the output feature vector of the adapter.
8. The human feature perception method based on a large model and a sensing terminal according to claim 2, characterized in that, The process involves freezing the main weight matrix and introducing a low-rank weight increment ΔW to adjust the trainable matrix; including... method: The trainable matrix is adjusted according to the following formula: ； Where x represents the input vector of the layer containing the adapter; W represents the original trainable matrix of the layer containing the adapter; A represents the matrix that projects the original trainable matrix onto the low-dimensional space; and B represents the matrix that projects matrix A back to the output space. h′ represents the output feature vector of the adapter.
9. A human feature perception system based on a large model and a sensing terminal, characterized in that, The system is used to implement the human feature perception method based on a large model and a sensing terminal as described in any one of claims 1-8, the system comprising: The acquisition module is used to continuously acquire raw message channel status information from different individuals from authorized channels; The preprocessing module is used to preprocess the original message channel state information to obtain the original training dataset; The training module is used to train a large cross-individual feature perception model based on the original dataset; The perception module is used to collect real-time message channel status information of the individual under test in real time, preprocess the real-time message channel status information, encode it into text format data, and then input it into the cross-individual feature perception big model to obtain the human body features of the individual under test.
10. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, executed by a processor, can implement the human feature perception method based on a large model and a sensing terminal as described in any one of claims 1-8.