Non-contact personnel identity and posture recognition method based on motion instability compensation

Abstract

The application belongs to the field of millimeter wave radar perception and human posture recognition, and discloses a non-contact personnel identity and posture recognition method based on motion instability compensation, comprising the following steps: step 1, static background points are adaptively screened, the radar self-motion speed is estimated, and the point cloud coordinates and the micro-Doppler spectrum are compensated; step 2, a GesAuthNet double-branch deep neural network is built to obtain shape parameters and posture parameters; step 3, an identity authenticator is constructed in combination with the shape parameters to authenticate the legitimacy of the user, and a gesture classifier is constructed in combination with the posture parameters to identify the gesture of the user. The application has the advantages of motion instability compensation capability, robust gesture feature extraction capability, high-precision identity authentication and gesture recognition, and strong stability and robustness.

Description

Technical Field

[0001] This application belongs to the field of millimeter-wave radar perception and human posture recognition, specifically involving a non-contact method for human posture recognition and identity authentication based on motion instability compensation. Background Technology

[0002] Human behavior perception is a key enabling technology for many human-centered applications, such as human-computer interaction, virtual reality, and augmented reality. With increasing user concerns about privacy and personalized services, joint verification of human behavior and identity has gradually become a core requirement for secure human-computer interaction applications. In such applications, the system needs to understand who is engaged in what activities, which enhances system security and provides more accurate services in specific scenarios. The academic community has devoted considerable effort to achieving joint verification of user behavior and identity.

[0003] Wearable device-based solutions use sensors worn by users to perceive human activity in real time, obtaining information about their identity and activities. However, these solutions typically require users to wear additional devices, causing inconvenience. Camera-based solutions infer identity and activity by capturing images of people. However, these solutions often expose the privacy of people and their surroundings and are frequently affected by poor lighting conditions, leading to poor performance. In recent years, wireless signal-based solutions have perceived human identity and activity by analyzing wireless signals rather than images or videos, fundamentally avoiding privacy leaks. Furthermore, they offer advantages such as user-friendliness and low environmental sensitivity. Among these, millimeter-wave radar-based methods, benefiting from their large bandwidth and high resolution, have gained significant attention in academia and are being gradually developed for applications such as user behavior and authentication. For example, GesturePrint and PolyLite-RadarNet utilize millimeter-wave radar to jointly infer human behavior and identity information.

[0004] However, these solutions have not yet been extended to scenarios where radar equipment is mobile. In real-world scenarios, millimeter-wave radar equipment may be mounted on mobile platforms, requiring the simultaneous inference of user behavior and identity while the equipment is moving with the platform. Therefore, developing a scalable solution that can adapt to mobile platforms is crucial. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a non-contact human posture recognition and identity authentication method based on motion instability compensation. This method solves the problem of reliable perception on a motion platform through a non-contact human posture recognition system. By compensating for motion instability, it filters background point clouds, estimates radar self-motion, compensates for and eliminates signal interference caused by radar motion, and uses a GesAuthNet dual-branch deep neural network to extract identity-related morphological features and gesture-related posture features respectively. Through an identity prior guidance mechanism, it separates robust gesture features that are not affected by identity, ultimately achieving high-precision identity authentication and gesture recognition.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] This application presents a non-contact method for personnel identification and posture recognition based on motion instability compensation, which specifically includes the following steps:

[0008] Step 1: Extract personnel signals for motion instability compensation: By adaptively filtering static background points, the radar self-motion velocity is estimated using the consistency in time and space, and the point cloud coordinates and micro-Doppler spectrum are compensated to eliminate signal interference introduced by the radar self-motion velocity.

[0009] Step 2: Construct a GesAuthNet dual-branch deep neural network. The GesAuthNet dual-branch deep neural network includes an identity representation construction branch and a gesture motion representation construction branch. The identity representation construction branch obtains morphological parameters related to human behavior. The gesture motion representation constructs branches to obtain posture parameters related to the gesture motion. ;

[0010] Step 3, Identity Verification and Gesture Recognition: Combining Morphological Parameters Build an identity authenticator to authenticate the user's legitimacy, combining gesture parameters. Build a gesture classifier to recognize user gestures.

[0011] A further improvement of this application is that step 1 includes the following steps:

[0012] Step 1.1: Utilize the consistency of point cloud velocity residuals in time and space to adaptively filter the confidence of each background point belonging to the static background, and optimize the radar self-motion velocity based on the confidence weight.

[0013] Step 1.2: Distinguish between background points and dynamic points based on confidence level, and perform point cloud correction and micro-Doppler spectrum compensation based on radar self-motion velocity.

[0014] A further improvement to this application is that step 1.1 specifically includes the following steps:

[0015] Step 1.1.1: Define the point cloud velocity residual. :

[0016] ;

[0017] ;

[0018] ;

[0019] in, The speed of radar movement. For the first Frame contains The first point The radial velocity magnitude of a point cloud, The distance between the background point and the radar. Unit vector in spatial direction, point cloud coordinates , It is the azimuth angle. The pitch angle, For transpose; For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in The velocity of the point cloud in the direction of motion, if the background points originate from a static background, is the point cloud velocity residual. If the background point originates from a dynamic object, then the residual... ;

[0020] Step 1.1.2: Based on the point cloud velocity residual Calculate the confidence score of each background point belonging to the static background. To achieve adaptive filtering of background points:

[0021]

[0022] Where σ is the variance of all point cloud velocity residuals;

[0023] Step 1.1.3: Combining confidence levels Estimating the optimal radar self-motion velocity based on consistency in time and space. :

[0024]

[0025] in, Indicates the first Frame-estimated radar self-motion velocity, Indicates the first Frame-estimated radar self-motion velocity, For time consistency constraints, For spatial consistency constraints, It is the weighted directional covariance matrix of the background points. The weighted directional covariance matrix of the background points is calculated as the sum of the second largest and smallest eigenvalues.

[0026] .

[0027] A further improvement in this application is that step 1.2 specifically includes the following steps:

[0028] Step 1.2.1: Calculate the radar from the current... Rotation matrix from frame coordinates to reference coordinates, i.e., coordinates of frame 1. and translation vector And migrate the point cloud to the reference coordinates, after migration the first Frame number The coordinates of the points are:

[0029] ;

[0030] Step 1.2.1: Map the radar's self-motion velocity to the spatial direction of the point, and eliminate the mapped radar self-motion velocity. After correction... Frame number The magnitude of the radial velocity at each point is:

[0031] ;

[0032] Step 1.2.3, Micro-Doppler Spectrum Compensation: For each time moment The compensation factor for Doppler shift is expressed as:

[0033]

[0034] Where λ is the carrier wavelength. For the target direction unit vector, the microDoppler spectrum Every moment Execution frequency compensation:

[0035] .

[0036] A further improvement in this application is that: in step 2, the identity representation construction branch constructs morphological parameter representations related to personnel identity, and the gesture motion representation construction branch constructs posture parameter representations related to gesture motion, wherein:

[0037] The identity representation construction branch includes two steps: structural biometrics extraction and morphological parameter representation extraction. Structural biometrics extraction is used to extract global biometrics of the human body, and morphological parameter representation extraction uses the extracted global biometrics to construct morphological parameter representations related to the person's identity. Structural biometrics extraction includes anchor point localization, local biometrics extraction, and global biometrics construction.

[0038] The gesture motion representation construction branch consists of two steps: gesture motion feature extraction and pose parameter representation extraction. Gesture motion feature extraction extracts global time-frequency features of gesture motion from the micro-Doppler spectrum, and pose parameter representation extraction uses the extracted global time-frequency features to construct a pose parameter representation related to gesture motion.

[0039] A further improvement in this application is that the identity representation construction branch specifically includes the following steps:

[0040] Step 2.1.1: Identity Representation - Branch Anchor Point Location: Based on the human body reflection point cloud, points representing different limbs in the human body are automatically inferred, and the first... Human body reflection point cloud in frames High-dimensional point cloud feature representation Aggregated high-dimensional point cloud feature representation Obtain global point cloud representation Finally, the global point cloud representation is used. Calculate Anchor points = , For the first High-dimensional feature representation of each anchor point;

[0041] Step 2.1.2: Local biometric feature extraction. The calculated... Set each anchor point as There are cluster centers, and the points are divided according to the correlation between each point and the anchor point. A subset of point clouds for a key location, the first The first frame The point and the first The correlation calculation method for each anchor point is as follows:

[0042] ;

[0043] in, It is the first The first frame High-dimensional point cloud feature representation of points Indicates the first Frame number High-dimensional point cloud feature representation of each point is extracted using a self-attention mechanism. Local biomarkers of key areas ;

[0044] Step 2.1.3: Global biometric feature construction and fusion of the identity representation construction branch. Local biomarkers of key areas Constructing global biosignatures ;

[0045] Step 2.1.4: The morphological parameter representation extraction of the identity representation construction branch includes a morphological parameter estimator, utilizing the obtained global biometric features. Estimating morphological parameters related to human behavior Among them, the loss function for calculating morphological parameters Used for supervision The estimation process: .

[0046] A further improvement in this application is that, in step 2.1.1, the anchor point diversity constraint loss function is defined. To supervise the calculation process of the anchor points:

[0047]

[0048] in, It is the high-dimensional feature representation of the j-th anchor point. This represents the total number of frames in the point cloud.

[0049] A further improvement of this application is that the gesture motion representation construction branch includes the following steps:

[0050] Step 2.2.1: Gesture motion representation construction branch local frequency feature extraction uses a two-dimensional convolutional network to extract local frequency features at each moment of the spectrum along the time dimension;

[0051] Step 2.2.2: Extracting global time-frequency features for the gesture motion representation branch. The importance of local frequency features at each time step is measured, and the local frequency features extracted at each time step are weighted and fused according to their importance to obtain global time-frequency features.

[0052] Step 2.2.3: The gesture motion representation construction branch includes a pose parameter representation extraction, which uses the extracted global time-frequency features to estimate the pose parameters related to the gesture motion. As a gesture feature, it is used for gesture recognition.

[0053] A further improvement of this application is that, in step 2.2.3, the loss function of the posture parameters is calculated. Used to monitor posture parameters The estimation process ensures that the estimated pose parameters are accurate. Consistent with the actual pose parameters, represented as Using identity representation to construct branches to obtain morphological parameters related to personnel identity This indicates that as prior identity information, from The component related to personnel identity is removed from the middle. Decoupling yields attitude parameters independent of identity. As a representation of gesture movement, this process is represented as

[0054]

[0055] in, , express In the feature space, define an orthogonal constraint loss function for the orientation. Supervision The decoupling process:

[0056] .

[0057] A further improvement in this application is that step 3 specifically includes the following steps:

[0058] Step 3.1: An identity authenticator is constructed to authenticate the user's legitimacy: given the morphological parameters extracted at the current moment. Identity authenticator The user's identity is determined using a multi-layered perception structure. The determination process is represented as follows:

[0059] ,

[0060] The calculation represents the difference between the identity inferred by the identity authenticator and the actual identity of the person. :

[0061] ;

[0062] in, and For real user tags and predicted user tags, For the number of users, Total number of frames;

[0063] Step 3.2: A gesture classifier was constructed to recognize the user's gestures: given the gesture parameters extracted at the current moment. Gesture Recognition A multi-layered perception structure is used to recognize user gestures. The recognition process is represented as follows:

[0064]

[0065] Calculate the gap between the gesture category recognized by the gesture classifier and the true gesture category. :

[0066]

[0067] in, and For real gesture labels and predicted gesture labels, Number of gestures;

[0068] Step 3.3: Calculate the gesture recognition loss function and identity recognition loss function The sum :

[0069]

[0070]

[0071]

[0072] in, and The loss functions for gesture recognition and identity recognition are respectively for the gesture recognition branch and the identity recognition branch.

[0073] The beneficial effects of this application are: the present invention has the advantages of motion instability compensation capability, robust gesture feature extraction capability, high-precision identity authentication and gesture recognition, and strong stability and robustness, specifically:

[0074] This application can accurately estimate the self-velocity of a millimeter-wave radar sensor mounted on a mobile platform, and eliminate the reflection signal offset caused by the radar's self-motion based on the estimated radar self-velocity, thereby recovering the reflection signal related to the person's identity and gesture movement, and has the ability to compensate for motion instability.

[0075] The gesture motion representation construction branch designed in this application can use the morphological parameter representation related to the person's identity as a priori, and decouple the gesture feature to obtain the posture parameter representation that is independent of the person's identity, so as to realize the gesture recognition of any person.

[0076] This application constructs a GesAuthNet dual-branch deep neural network, which can extract morphological parameter representations related to personnel identity from the identity representation construction branch and extract pose parameter representations related to gesture movement from the gesture movement branch, thereby achieving joint identity authentication and gesture recognition.

[0077] The personnel signal extraction method based on motion instability compensation established in this application can eliminate the interference of radar self-motion on the reflected signals from personnel, while the designed GesAuthNet eliminates the information coupling between personnel identity and gesture motion, and has strong stability and robustness in personnel identity and gesture recognition. Attached Figure Description

[0078] Figure 1 This is a flowchart of the identification method of this application.

[0079] Figure 2 This is an example diagram of the application of this application.

[0080] Figure 3 This is the system architecture diagram of this application.

[0081] Figure 4 This is the architecture diagram of the GesAuthNet dual-branch deep neural network in this application.

[0082] Figure 5 These are the 10 everyday gestures involved in this application.

[0083] Figure 6 This is a schematic diagram illustrating how this application achieves gesture recognition with an average gesture recognition accuracy of 97.67% across 10 users.

[0084] Figure 7 This is a diagram illustrating how this application achieves user authentication with an average accuracy rate of 97.70% across 10 users. Detailed Implementation

[0085] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the present invention. That is, in some embodiments of the present invention, these practical details are not essential. In addition, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.

[0086] like Figures 1-3 As shown, this application discloses a non-contact personnel identification and posture recognition method based on motion instability compensation. The non-contact personnel identification and posture recognition method specifically includes the following steps:

[0087] Step 1: Extract personnel signals for motion instability compensation: By adaptively filtering static background points, the radar's self-motion velocity is estimated using temporal and spatial consistency, and precise compensation is performed on the point cloud coordinates and micro-Doppler spectrum to eliminate signal interference introduced by the radar's self-motion velocity. Specifically, this includes the following steps:

[0088] Step 1.1, Adaptive Dynamic Interference Robust Self-Motion Estimation: Utilizing the temporal and spatial consistency of the point cloud velocity residuals, the confidence level of each background point belonging to the static background is adaptively filtered, and the radar self-motion velocity is optimized based on a confidence-weighted average. Specifically, this includes the following steps:

[0089] Step 1.1.1 Adaptive estimation of background points based on temporal and spatial consistency: Define point cloud velocity residuals :

[0090] ;

[0091] ;

[0092] ;

[0093] in, The speed of radar movement. For the first Frame contains The first point The radial velocity magnitude of a point cloud, The distance between the background point and the radar. Unit vector in spatial direction, point cloud coordinates , It is the azimuth angle. The pitch angle, For transpose; For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in The velocity of the point cloud in the direction of motion, if the background points originate from a static background, is the point cloud velocity residual. If the background point originates from a dynamic object, then the residual... ;

[0094] Step 1.1.2: Based on the point cloud velocity residual Calculate the confidence score of each background point belonging to the static background. To achieve adaptive filtering of background points:

[0095]

[0096] Where σ is the variance of all point cloud velocity residuals, representing the fluctuation range of the point cloud velocity residuals. The smaller the residual, the higher the confidence level; the larger the residual, the lower the confidence level.

[0097] Step 1.1.3: Combining confidence levels Estimating the optimal radar self-motion velocity based on consistency in time and space. :

[0098]

[0099] in, Indicates the first Frame-estimated radar self-motion velocity, Indicates the first Frame-estimated radar self-motion velocity, Temporal consistency constraints are used to suppress unreasonable jumps between adjacent frames and improve estimation stability. This spatial consistency constraint encourages points participating in the estimation to be sufficiently distributed spatially, while avoiding the dominance of local dynamic groups in the estimation results. It quantitatively characterizes the directional distribution of points in space and is the weighted directional covariance matrix of background points. The weighted directional covariance matrix of the background points is calculated as the sum of the second largest and smallest eigenvalues.

[0100] .

[0101] If the point cloud is sufficiently distributed, the eigenvalues ​​of the matrix will be relatively balanced; if they are concentrated in a single direction, they will exhibit a low-rank structure.

[0102] Step 1.2: Differentiate between background points and dynamic points based on confidence level, and perform point cloud correction and micro-Doppler spectrum compensation based on radar self-motion velocity. This specifically includes the following steps:

[0103] Step 1.2.1: Calculate the radar from the current... Rotation matrix from frame coordinates to reference coordinates, i.e., coordinates of frame 1. and translation vector And migrate the point cloud to the reference coordinates, after migration the first Frame number The coordinates of the points are:

[0104] ;

[0105] Step 1.2.1: Map the radar's self-motion velocity to the spatial direction of the point, and eliminate the mapped radar self-motion velocity. After correction... Frame number The magnitude of the radial velocity at each point is:

[0106] ;

[0107] Step 1.2.3, Micro-Doppler Spectrum Compensation: For each time moment The compensation factor for Doppler shift is expressed as:

[0108]

[0109] Where λ is the carrier wavelength. For the target direction unit vector, the microDoppler spectrum Every moment Execution frequency compensation:

[0110] .

[0111] Step 2, Feature Extraction from a Two-Branch Deep Neural Network: Construct a GesAuthNet two-branch deep neural network, such as... Figure 4 As shown, the GesAuthNet dual-branch deep neural network includes an identity representation construction branch and a gesture motion representation construction branch. The identity representation construction branch constructs morphological parameter representations related to a person's identity, while the gesture motion representation construction branch constructs pose parameter representations related to gesture motion. The former utilizes point clouds to construct morphological parameter representations related to a person's identity, while the latter utilizes micro-Doppler spectra to construct pose parameter representations related to gesture motion. The identity representation construction branch obtains morphological parameters related to a person's behavior. The gesture motion representation constructs branches to obtain posture parameters related to the gesture motion. .

[0112] In step 2, the identity representation construction branch constructs morphological parameter representations related to personnel identity, and the gesture motion representation construction branch constructs pose parameter representations related to gesture motion, wherein:

[0113] The identity representation construction branch includes two steps: structural biometrics extraction and morphological parameter representation extraction. Structural biometrics extraction is used to extract global biometrics of the human body. Morphological parameter representation extraction uses the extracted global biometrics to construct morphological parameter representations related to the person's identity. Structural biometrics extraction includes anchor point localization, local biometrics extraction, and global biometrics construction. Anchor point localization separates points related to the human limbs from the point cloud. Local biometrics extraction extracts local biometrics related to the limbs from the points related to the limbs. Global biometrics construction uses the fusion of local biometrics of different limbs to obtain global biometrics.

[0114] The gesture motion representation construction branch consists of two steps: gesture motion feature extraction and posture parameter representation extraction. Gesture motion feature extraction extracts global time-frequency features of gesture motion from the micro-Doppler spectrum. Posture parameter representation extraction uses the extracted global time-frequency features to construct posture parameter representations related to gesture motion. Gesture motion feature extraction includes local frequency feature extraction and global time-frequency feature extraction. Local frequency feature extraction extracts the local frequency features corresponding to each moment in the micro-Doppler spectrum. Global time-frequency feature extraction aggregates the local frequency features at each moment to construct global time-frequency features.

[0115] The identity representation construction branch specifically includes the following steps:

[0116] Step 2.1.1: Identity Representation Construction Branch Anchor Point Location. Based on the human body reflection point cloud, points representing different limbs in the human body are automatically inferred, and the first limb is extracted through a multilayer perceptron. Human body reflection point cloud in frames High-dimensional point cloud feature representation Then, the high-dimensional point cloud feature representation is aggregated using max pooling. Obtain global point cloud representation Finally, the global point cloud representation is used. By using fully connected layers, key human body parts that are body-specific can be calculated based on the spatial location of each point. Anchor points = , For the first High-dimensional feature representation of each anchor point;

[0117] Define the anchor point diversity constraint loss function The calculation process of anchor points is supervised to minimize the correlation between different anchor points, ensuring that each learned anchor point focuses on different spatial geometric patterns and preventing multiple anchor points from degenerating into the same region. Anchor point diversity constrains the loss function. The correlation between different anchor points was measured and expressed as:

[0118]

[0119] in, It is the first High-dimensional feature representation of each anchor point This represents the total number of frames in the point cloud.

[0120] Step 2.1.2: Local Biometric Feature Extraction for Identity Representation Construction Branch. Based on the correlation between anchor points and points, the point cloud subsets for each limb part are divided, and local biometric features are extracted from the point cloud subsets of each key part of the human body. Specifically, the local biometric feature extraction uses the K-means clustering algorithm to calculate... Set each anchor point as There are cluster centers, and the points are divided according to the correlation between each point and the anchor point. A subset of point clouds for a key location, the first The first frame The point and the first The correlation calculation method for each anchor point is as follows:

[0121] ;

[0122] in, It is the first The first frame High-dimensional point cloud feature representation of points Indicates the first Frame number High-dimensional point cloud feature representation of individual points, identity representation construction of local biometric feature extraction branches, and further extraction using a self-attention mechanism. Local biomarkers of key areas ;

[0123] Step 2.1.3: The global biometric feature construction of the identity representation branch adopts a cross-attention (CA) fusion mechanism. Local biomarkers of key areas Constructing global biosignatures ;

[0124] Step 2.1.4, the morphological parameter representation extraction of the identity representation construction branch includes a morphological parameter estimator that uses a multilayer perceptron structure to utilize the obtained global biometric features. Estimating morphological parameters related to human behavior , used to represent physical structure information related to a person's identity.

[0125] Loss function for calculating morphological parameters Used for supervision The estimation process ensures that the estimated morphological parameters are consistent with the actual morphological parameters. The estimated morphological parameters were measured. With actual morphological parameters The difference between them is expressed as .

[0126] The gesture motion representation branch construction includes the following steps:

[0127] Step 2.2.1: Gesture motion representation construction branch local frequency feature extraction uses a two-dimensional convolutional network to extract local frequency features at each moment of the spectrum along the time dimension;

[0128] Step 2.2.2: Gesture motion representation construction branch global time-frequency feature extraction. The importance of local frequency features at each moment is measured by time aggregation algorithm, and the local frequency features extracted at each moment are weighted and fused according to importance to obtain global time-frequency features, which contain rich gesture motion information;

[0129] Step 2.2.3, the gesture motion representation construction branch of the pose parameter representation extraction includes a pose parameter estimator, which uses a multilayer perceptron structure to estimate the pose parameters related to the gesture motion using the extracted global time-frequency features. As a gesture feature, it is used for gesture recognition.

[0130] Loss function for calculating pose parameters Used to monitor posture parameters The estimation process ensures that the estimated pose parameters are accurate. Consistent with actual pose parameters, The estimated pose parameters were measured. Compared with actual pose parameters The difference between them is expressed as Considering This step, which couples personnel identity information with gesture motion information, further utilizes the identity representation to construct morphological parameters related to personnel identity obtained from the branch. This indicates that as prior identity information, from The component related to personnel identity is removed from the middle. Decoupling yields attitude parameters independent of identity. As a representation of gesture movement, this process is represented as

[0131]

[0132] in, , express In the feature space, this step defines the orthogonal constraint loss function. Supervision The decoupling process ensures and Lowest correlation Measured and Correlation between them: .

[0133] Step 3, Identity Verification and Gesture Recognition: Combining Morphological Parameters Build an identity authenticator to authenticate the user's legitimacy, combining gesture parameters. Build a gesture classifier to recognize user gestures. This includes the following steps:

[0134] Step 3.1: An identity authenticator is constructed to authenticate the user's legitimacy: given the morphological parameters extracted at the current moment. Identity authenticator The user's identity is determined using a multi-layered perception structure. The determination process is represented as follows:

[0135] ,

[0136] The calculation represents the difference between the identity inferred by the identity authenticator and the actual identity of the person. :

[0137] ;

[0138] in, and For real user tags and predicted user tags, For the number of users, Total number of frames;

[0139] Step 3.2: A gesture classifier was constructed to recognize the user's gestures: given the gesture parameters extracted at the current moment. Gesture Recognition A multi-layered perception structure is used to recognize user gestures. The recognition process is represented as follows:

[0140]

[0141] Calculate the gap between the gesture category recognized by the gesture classifier and the true gesture category. :

[0142] ,

[0143] in, and For real gesture labels and predicted gesture labels, Number of gestures;

[0144] Step 3.3: Calculate the gesture recognition loss function and identity recognition loss function The sum :

[0145]

[0146]

[0147]

[0148] in, and These are the loss functions for the gesture recognition branch and the identity recognition branch, respectively. (Gesture recognition loss function) The gesture motion representation represents the sum of the loss terms for constructing branches and gesture classification; the identity recognition loss function. This represents the sum of the loss terms for identity representation construction and identity recognition.

[0149] Figure 5 This application presents 10 common hand gestures used by users to interact with a mobile platform (hereinafter referred to as the interactive target) equipped with millimeter-wave radar sensing equipment. Different gestures represent different commands. Users face the interactive target while performing all gestures. Specifically: (a) A left gesture requires the user to wave from right to left, indicating a command to turn left to the interactive target. (b) A right gesture requires the user to wave from left to right, indicating a command to turn right to the interactive target. (c) An up gesture requires the user to raise their hand from bottom to top, indicating a command to move upwards to the interactive target. (d) A down gesture requires the user to lower their hand from top to bottom, indicating a command to move downwards to the interactive target. (e) A forward push gesture requires the user to extend their hand towards the interactive target, indicating a command to move backwards to the interactive target. (f) A backward pull gesture requires the user to retract their extended arm from the direction of the interactive target, indicating a command to move forward to the interactive target. (g) A clockwise gesture requires the user to rotate their arm clockwise, indicating a command to reduce movement speed to the interactive target. (h) A counter-clockwise gesture requires the user to rotate their arm counter-clockwise, indicating that the user is giving a command to the interaction target to increase movement speed. (i) A waving gesture requires the user to extend their arms to the sides and wave them up and down, indicating that the user is giving a command to the interaction target to stop moving. (j) In a stationary posture, the user does not perform any gesture, indicating that the user stops giving commands to the interaction target.

[0150] Based on Figure 6 and Figure 7 As can be seen, this application can achieve a gesture recognition accuracy of 97.67% and an identity authentication accuracy of 97.70% under radar motion conditions, which is significantly better than the existing benchmark scheme, verifying its excellent stability and robustness for human posture recognition and identity authentication under motion instability conditions.

[0151] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A non-contact method for personnel identity and posture recognition based on motion instability compensation, characterized in that: The non-contact personnel identification and posture recognition method specifically includes the following steps: Step 1: Extract personnel signals for motion instability compensation: By adaptively filtering static background points, the radar self-motion velocity is estimated using the consistency in time and space, and the point cloud coordinates and micro-Doppler spectrum are compensated to eliminate signal interference introduced by the radar self-motion velocity. Step 2: Construct a GesAuthNet dual-branch deep neural network. The GesAuthNet dual-branch deep neural network includes an identity representation construction branch and a gesture motion representation construction branch. The identity representation construction branch obtains morphological parameters related to human behavior. The gesture motion representation constructs branches to obtain posture parameters related to the gesture motion. ; Step 3, Identity Verification and Gesture Recognition: Combining Morphological Parameters Build an identity authenticator to authenticate the user's legitimacy, combining gesture parameters. Build a gesture classifier to recognize user gestures.

2. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Utilize the consistency of point cloud velocity residuals in time and space to adaptively filter the confidence of each background point belonging to the static background, and optimize the radar self-motion velocity based on the confidence weight. Step 1.2: Distinguish between background points and dynamic points based on confidence level, and perform point cloud correction and micro-Doppler spectrum compensation based on radar self-motion velocity.

3. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 2, characterized in that: Step 1.1 specifically includes the following steps: Step 1.1.1: Define the point cloud velocity residual. : ； ； ； in, The speed of radar movement. For the first Frame contains The first point The magnitude of the radial velocity of a point cloud. The distance between the background point and the radar. Unit vector in spatial direction, point cloud coordinates , It is the azimuth angle. The pitch angle, For transpose; For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in Self-motion velocity in the direction, For the first Frame radar in The velocity of the point cloud in the direction of motion, if the background points originate from a static background, is the point cloud velocity residual. If the background point originates from a dynamic object, then the residual... ; Step 1.1.2: Based on the point cloud velocity residual Calculate the confidence score of each background point belonging to the static background. To achieve adaptive filtering of background points: Where σ is the variance of all point cloud velocity residuals; Step 1.1.3: Combine confidence level Estimating the optimal radar self-motion velocity based on consistency in time and space. : in, Indicates the first Frame-estimated radar self-motion velocity, Indicates the first Frame-estimated radar self-motion velocity, For time consistency constraints, For spatial consistency constraints, It is the weighted directional covariance matrix of the background points. The weighted directional covariance matrix of the background points is calculated as the sum of the second largest and smallest eigenvalues. 。 4. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 2, characterized in that: Step 1.2 specifically includes the following steps: Step 1.2.1: Calculate the radar from the current... Rotation matrix from frame coordinates to reference coordinates, i.e., coordinates of frame 1. and translation vector And migrate the point cloud to the reference coordinates, after migration the first Frame number The coordinates of the points are: ； Step 1.2.1: Map the radar's self-motion velocity to the spatial direction of the point, and eliminate the mapped radar self-motion velocity. After correction... Frame number The magnitude of the radial velocity at each point is: ； Step 1.2.3, Micro-Doppler Spectrum Compensation: For each time moment The compensation factor for Doppler shift is expressed as: Where λ is the carrier wavelength. For the target direction unit vector, the microDoppler spectrum Every moment Execution frequency compensation: 。 5. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 1, characterized in that: In step 2, the identity representation construction branch constructs morphological parameter representations related to personnel identity, and the gesture motion representation construction branch constructs pose parameter representations related to gesture motion, wherein: The identity representation construction branch includes two steps: structural biometrics extraction and morphological parameter representation extraction. Structural biometrics extraction is used to extract global biometrics of the human body, and morphological parameter representation extraction uses the extracted global biometrics to construct morphological parameter representations related to the person's identity. Structural biometrics extraction includes anchor point localization, local biometrics extraction, and global biometrics construction. The gesture motion representation construction branch consists of two steps: gesture motion feature extraction and pose parameter representation extraction. Gesture motion feature extraction extracts global time-frequency features of gesture motion from the micro-Doppler spectrum, and pose parameter representation extraction uses the extracted global time-frequency features to construct a pose parameter representation related to gesture motion.

6. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 5, characterized in that: The identity representation branch construction process includes the following steps: Step 2.1.1: Identity Representation - Branch Anchor Point Location: Based on the human body reflection point cloud, points representing different limbs in the human body are automatically inferred, and the first... Human body reflection point cloud in frames High-dimensional point cloud feature representation Aggregated high-dimensional point cloud feature representation Obtain global point cloud representation Finally, the global point cloud representation is used. Calculate Anchor points = , For the first High-dimensional feature representation of each anchor point; Step 2.1.2: Local biometric feature extraction. The calculated... Set each anchor point as There are cluster centers, and the points are divided according to the correlation between each point and the anchor point. A subset of point clouds for a key location, the first The first frame The point and the first The correlation calculation method for each anchor point is as follows: ； in, It is the first The first frame High-dimensional point cloud feature representation of points Indicates the first Frame number High-dimensional point cloud feature representation of each point is extracted using a self-attention mechanism. Local biomarkers of key areas ; Step 2.1.3: Global biometric feature construction and fusion of the identity representation construction branch. Local biomarkers of key areas Constructing global biosignatures ; Step 2.1.4: The morphological parameter representation extraction of the identity representation construction branch includes a morphological parameter estimator, utilizing the obtained global biometric features. Estimating morphological parameters related to human behavior Among them, the loss function for calculating morphological parameters Used for supervision The estimation process: .

7. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 6, characterized in that: In step 2.1.1, the anchor point diversity constraint loss function is defined. To supervise the calculation process of the anchor points: in, It is the first High-dimensional feature representation of each anchor point This represents the total number of frames in the point cloud.

8. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 5, characterized in that: The gesture motion representation branch construction includes the following steps: Step 2.2.1: Gesture motion representation construction branch local frequency feature extraction uses a two-dimensional convolutional network to extract local frequency features at each moment of the spectrum along the time dimension; Step 2.2.2: Extracting global time-frequency features for the gesture motion representation branch. The importance of local frequency features at each time step is measured, and the local frequency features extracted at each time step are weighted and fused according to their importance to obtain global time-frequency features. Step 2.2.3: The gesture motion representation construction branch includes a pose parameter representation extraction, which uses the extracted global time-frequency features to estimate the pose parameters related to the gesture motion. As a gesture feature, it is used for gesture recognition.

9. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 8, characterized in that: In step 2.2.3, the loss function for the pose parameters is calculated. Used to monitor posture parameters The estimation process ensures that the estimated pose parameters are accurate. Consistent with the actual pose parameters, represented as Using identity representation to construct branches to obtain morphological parameters related to personnel identity This indicates that as prior identity information, from The component related to personnel identity is removed from the middle. Decoupling yields attitude parameters independent of identity. As a representation of gesture movement, this process is represented as in, , express In the feature space, define an orthogonal constraint loss function for the orientation. Supervision The decoupling process: 。 10. The non-contact personnel identification and posture recognition method based on motion instability compensation according to claim 9, characterized in that: Step 3 specifically includes the following steps: Step 3.1: An identity authenticator is constructed to authenticate the user's legitimacy: given the morphological parameters extracted at the current moment. Identity authenticator The user's identity is determined using a multi-layered perception structure. The determination process is represented as follows: ， The calculation represents the difference between the identity inferred by the identity authenticator and the actual identity of the person. : ； in, and For real user tags and predicted user tags, For the number of users, Total number of frames; Step 3.2: A gesture classifier was constructed to recognize the user's gestures: given the gesture parameters extracted at the current moment. Gesture Recognition A multi-layered perception structure is used to recognize user gestures. The recognition process is represented as follows: Calculate the gap between the gesture category recognized by the gesture classifier and the true gesture category. : in, and For real gesture labels and predicted gesture labels, Number of gestures; Step 3.3: Calculate the gesture recognition loss function and identity recognition loss function The sum : in, and These are the gesture recognition loss function and the identity recognition loss function, respectively.

Images