Motion gesture recognition method and system for wearable smart device

Abstract

The application provides a motion gesture recognition method and system for a wearable smart device, which collects device acceleration and angular velocity data, calculates an acceleration modulus value and a gravity modulus value deviation first, and demarcates a motion feasible subspace according to a dynamic threshold value of a previous moment gesture; an angular velocity is projected and integrated to update a rotation matrix, when the deviation is lower than the threshold value, an acceleration correction is used to obtain a constrained rotation matrix, when the deviation is not lower than the threshold value, a projected integrated gesture rotation matrix is used as a current constrained gesture rotation matrix, a six-dimensional main gesture vector is constructed from the first two columns of the matrix, a minimum rotation axis angle vector of an alignment reference matrix is solved as a perturbation vector, the main gesture vector is input into a classification model, a weight adjustment is performed on a classification confidence in combination with a perturbation vector modulus value, a main body gesture and a credible state thereof are generated, a corresponding micro-motion recognizer is activated, a distribution feature of the perturbation vector in a time sequence window is extracted and input into the recognizer, and a complete motion recognition result is output in combination with the main body gesture.

Description

Technical Field

[0001] This application belongs to the field of recognition, and in particular relates to a method and system for motion posture recognition of wearable smart devices. Background Technology

[0002] Wearable devices have been widely used in medical rehabilitation, sports, virtual reality, and human-computer interaction. In these applications, accurately capturing and recognizing the wearer's motion posture is crucial for achieving intelligent sensing, state assessment, and interactive control. Wearable devices typically rely on built-in microelectromechanical systems (MEMS) inertial sensors, such as accelerometers and gyroscopes, to acquire inertial data during human movement in real time. Traditional motion posture recognition methods generally collect angular velocity and acceleration data from the device, use posture fusion algorithms to calculate posture representations such as quaternions and rotation matrices, and then combine this with machine learning models to classify and predict human motion categories. This type of recognition method based on inertial sensor data has advantages such as being unaffected by ambient light, strong privacy protection, and flexible deployment, and has become an important technical path in human motion recognition. However, with the continuous expansion of application scenarios, especially with the increasing demand for high-precision, fine-grained motion analysis, the shortcomings of existing conventional recognition methods are gradually becoming apparent.

[0003] When the human body is in a state of rapid dynamic motion, the data collected by accelerometers often includes significant non-gravitational accelerations, making the attitude correction mechanism, which relies on the standard gravity direction, prone to deviation. Existing methods typically use a single fixed threshold for filtering or correction, lacking the ability to adaptively adjust constraints based on the current motion state. This makes it difficult to reliably suppress interference in complex dynamic scenarios, thus affecting the stability and accuracy of attitude calculation. Furthermore, natural human motion often has multi-level composite characteristics; that is, while performing macroscopic main movements, there may also be accompanying micro-movements of specific patterns. Existing attitude classification models often input the overall motion data uniformly and perform hybrid modeling, making it difficult to effectively distinguish between large-scale main motion features and small-scale perturbation features. This results in micro-movement information being easily masked by the main motion signal. The aforementioned coarse-grained recognition mechanism limits the wearable device's ability to finely perceive multimodal superimposed movements, making it difficult to meet the need for hierarchical recognition of main motion and accompanying micro-movements in complex dynamic scenarios. Summary of the Invention

[0004] To improve the ability of wearable devices to perceive and recognize complex movements in a precise and hierarchical manner, one or more embodiments of the present invention provide a motion posture recognition method for wearable smart devices, comprising the following steps: Acquire the angular velocity and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold and determine the feasible motion subspace according to the subject's motion posture category at the previous moment; The angular velocity data is projected onto the motion feasible subspace to obtain a projected angular velocity vector; the attitude rotation matrix is ​​updated by integration based on the projected angular velocity vector; when the absolute value of the deviation is lower than the dynamic motion threshold, the attitude rotation matrix updated by integration is corrected using the acceleration data to generate a constrained attitude rotation matrix. Extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional master attitude vector; calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector; The six-dimensional main attitude vector is input into the attitude classification model. The output confidence of the attitude classification model is adjusted in real time using the magnitude of the perturbation attitude vector, and the current subject motion attitude and its confidence state are generated. The corresponding auxiliary micro-motion pattern recognizer is activated according to the current subject motion attitude. The statistical distribution characteristics of the perturbation attitude vector in the continuous time window are input into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and the recognition result is output in combination with the current subject motion attitude.

[0005] Further, the step of calculating the absolute value of the deviation between the current acceleration modulus and the standard gravitational modulus based on the acceleration data includes: Calculate the vector magnitude corresponding to the acceleration data at the current sampling time; The absolute value of the deviation is obtained by subtracting the set standard gravity magnitude from the vector magnitude and taking the absolute value of the difference.

[0006] Further, the step of projecting the angular velocity data onto the motion feasible subspace to obtain the projected angular velocity vector includes: Extract the set of unit orthogonal basis vectors from the pre-set library that are mapped to the subject's motion posture category at the previous moment to construct the subspace reference matrix; The product of the transpose of the subspace reference matrix and the column vector of the currently input angular velocity data is calculated as the projection coefficient vector; The projection angular velocity vector is obtained by multiplying the subspace reference matrix and the projection coefficient vector.

[0007] Further, the step of using the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix includes: Multiply the transpose of the current attitude rotation matrix after integration update with the preset standard gravity direction unit vector to obtain the reference gravity direction vector in the device coordinate system. The alignment error vector is obtained by calculating the cross product of the normalized current acceleration data column vector and the reference gravity direction vector. The feedback micro-rotation vector is obtained by multiplying the set feedback proportional constant with the alignment error vector. Calculate the magnitude of the feedback micro-rotation vector; When the modulus is not lower than the preset small angle threshold, a local transformation matrix is ​​constructed according to the Rodriguez formula. The local transformation matrix is ​​the sum of a third-order identity matrix, a first correction term, and a second correction term. Wherein, the first correction term is the ratio of the sine of the modulus to the modulus multiplied by the antisymmetric matrix corresponding to the feedback micro-rotation vector; the second correction term is a matrix term obtained by multiplying the second correction coefficient by the square of the antisymmetric matrix corresponding to the feedback micro-rotation vector; wherein, the second correction coefficient is the ratio between the difference obtained by subtracting the cosine of the modulus and the square of the modulus. When the modulus is lower than a preset small angle threshold, the local transformation matrix is ​​calculated using the Taylor approximation values ​​corresponding to the ratio of the sine value to the modulus and the ratio of the cosine difference to the square of the modulus. The updated constrained attitude rotation matrix is ​​obtained by multiplying the current attitude rotation matrix updated by integration with the local transformation matrix.

[0008] Further, the calculation aligns the constrained attitude rotation matrix to the minimum rotational three-dimensional axis angle vector required to update the reference attitude rotation matrix based on the integral of the angular velocity data, as the perturbation attitude vector, including: The reference attitude rotation matrix is ​​generated by integrating the angular velocity data from the previous moment. The alignment error matrix is ​​obtained by multiplying the transpose of the constrained attitude rotation matrix with the reference attitude rotation matrix. Calculate the trace of the alignment error matrix, and calculate the true rotation angle based on the inverse cosine relationship between the trace and the rotation angle; Extract the rotation axis unit vector from the antisymmetric portion of the alignment error matrix; A three-dimensional axis-angle vector with a magnitude of the true rotation angle and a direction that is the unit vector of the rotation axis is constructed as the perturbation attitude vector.

[0009] Furthermore, the step of using the magnitude of the perturbation attitude vector to perform real-time weighted adjustment of the output confidence of the attitude classification model, and generating the current subject motion attitude and its reliable state, includes: The square root of the sum of the squares of the components of the perturbation attitude vector on the three orthogonal axes is used as the magnitude of the perturbation attitude vector. The weighting adjustment coefficient is calculated based on the magnitude of the perturbation attitude vector and the set adjustment intensity parameter; wherein the weighting adjustment coefficient is the reciprocal of a first value, and the first value is the sum of 1 and the product of the adjustment intensity parameter and the magnitude. The product of the weighting adjustment coefficient and the initial confidence level output by the attitude classification model is calculated as the adjusted confidence level; The class with the highest initial confidence output by the posture classification model is selected as the current subject's motion posture, and a reliable state is generated based on the adjusted confidence corresponding to that class. When the adjusted confidence level is not lower than the preset confidence threshold, a normal confidence state is generated. A low-confidence status label is generated when the adjusted confidence level corresponding to the category is lower than the preset confidence threshold.

[0010] Further, the step of inputting the statistical distribution characteristics of the perturbation attitude vector within a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and combining the current subject motion posture to output the recognition result, includes: Within a set continuous time window, continuously acquired perturbation attitude vectors are cached in the data queue; The mean, variance, and skewness of each independent component of the perturbation attitude vector in the data queue are extracted as the statistical distribution characteristics within the continuous time window; The statistical distribution features are concatenated in a preset order to form a feature vector, which is then input into the currently activated auxiliary micro-motion pattern recognizer network. After forward calculation by the auxiliary micro-motion pattern recognizer network, the auxiliary micro-motion pattern determination result under the current subject motion posture category is output.

[0011] One or more embodiments of the present invention also provide a motion posture recognition system for wearable smart devices, comprising the following modules: The calculation module is used to acquire the angular velocity and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold according to the subject's motion posture category at the previous moment and determine the feasible motion subspace; The generation module is used to project the angular velocity data onto the motion feasible subspace to obtain a projected angular velocity vector; to perform an integral update on the attitude rotation matrix based on the projected angular velocity vector; and when the absolute value of the deviation is lower than the dynamic motion threshold, to use the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix. The construction module is used to extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional master attitude vector; and to calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector. The recognition module is used to input the six-dimensional main attitude vector into the attitude classification model, adjust the output confidence of the attitude classification model in real time using the magnitude of the perturbation attitude vector, and generate the current subject motion attitude and its confidence state; activate the corresponding auxiliary micro-motion pattern recognizer according to the current subject motion attitude; input the statistical distribution characteristics of the perturbation attitude vector in a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and output the recognition result in combination with the current subject motion attitude.

[0012] Preferably, the step of calculating the absolute value of the deviation between the current acceleration modulus and the standard gravitational modulus based on the acceleration data includes: Calculate the vector magnitude corresponding to the acceleration data at the current sampling time; The absolute value of the deviation is obtained by subtracting the set standard gravity magnitude from the vector magnitude and taking the absolute value of the difference.

[0013] Preferably, the step of projecting the angular velocity data onto the motion feasible subspace to obtain the projected angular velocity vector includes: Extract the set of unit orthogonal basis vectors from the pre-set library that are mapped to the subject's motion posture category at the previous moment to construct the subspace reference matrix; The product of the transpose of the subspace reference matrix and the column vector of the currently input angular velocity data is calculated as the projection coefficient vector; The projection angular velocity vector is obtained by multiplying the subspace reference matrix and the projection coefficient vector.

[0014] Preferably, the step of using the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix includes: Multiply the transpose of the current attitude rotation matrix after integration update with the preset standard gravity direction unit vector to obtain the reference gravity direction vector in the device coordinate system. The alignment error vector is obtained by calculating the cross product of the normalized current acceleration data column vector and the reference gravity direction vector. The feedback micro-rotation vector is obtained by multiplying the set feedback proportional constant with the alignment error vector. Calculate the magnitude of the feedback micro-rotation vector; When the modulus is not lower than the preset small angle threshold, a local transformation matrix is ​​constructed according to the Rodriguez formula. The local transformation matrix is ​​the sum of a third-order identity matrix, a first correction term, and a second correction term. Wherein, the first correction term is the ratio of the sine of the modulus to the modulus multiplied by the antisymmetric matrix corresponding to the feedback micro-rotation vector; the second correction term is a matrix term obtained by multiplying the second correction coefficient by the square of the antisymmetric matrix corresponding to the feedback micro-rotation vector; wherein, the second correction coefficient is the ratio between the difference obtained by subtracting the cosine of the modulus and the square of the modulus. When the modulus is lower than a preset small angle threshold, the local transformation matrix is ​​calculated using the Taylor approximation values ​​corresponding to the ratio of the sine value to the modulus and the ratio of the cosine difference to the square of the modulus. The updated constrained attitude rotation matrix is ​​obtained by multiplying the current attitude rotation matrix updated by integration with the local transformation matrix.

[0015] Preferably, the calculation of aligning the constrained attitude rotation matrix to the minimum rotational three-dimensional axis angle vector required to update the reference attitude rotation matrix based on the integral of the angular velocity data as the perturbation attitude vector includes: The reference attitude rotation matrix is ​​generated by integrating the angular velocity data from the previous moment. The alignment error matrix is ​​obtained by multiplying the transpose of the constrained attitude rotation matrix with the reference attitude rotation matrix. Calculate the trace of the alignment error matrix, and calculate the true rotation angle based on the inverse cosine relationship between the trace and the rotation angle; Extract the rotation axis unit vector from the antisymmetric portion of the alignment error matrix; A three-dimensional axis-angle vector with a magnitude of the true rotation angle and a direction that is the unit vector of the rotation axis is constructed as the perturbation attitude vector.

[0016] Preferably, the step of using the magnitude of the perturbation attitude vector to perform real-time weighted adjustment of the output confidence of the attitude classification model and generating the current subject motion attitude and its reliable state includes: The square root of the sum of the squares of the components of the perturbation attitude vector on the three orthogonal axes is used as the magnitude of the perturbation attitude vector. The weighting adjustment coefficient is calculated based on the magnitude of the perturbation attitude vector and the set adjustment intensity parameter; wherein the weighting adjustment coefficient is the reciprocal of a first value, and the first value is the sum of 1 and the product of the adjustment intensity parameter and the magnitude. The product of the weighting adjustment coefficient and the initial confidence level output by the attitude classification model is calculated as the adjusted confidence level; The class with the highest initial confidence output by the posture classification model is selected as the current subject's motion posture, and a reliable state is generated based on the adjusted confidence corresponding to that class. When the adjusted confidence level is not lower than the preset confidence threshold, a normal confidence state is generated. A low-confidence status label is generated when the adjusted confidence level corresponding to the category is lower than the preset confidence threshold.

[0017] Preferably, the step of inputting the statistical distribution characteristics of the perturbation attitude vector within a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and combining the current subject motion posture to output the recognition result, includes: Within a set continuous time window, continuously acquired perturbation attitude vectors are cached in the data queue; The mean, variance, and skewness of each independent component of the perturbation attitude vector in the data queue are extracted as the statistical distribution characteristics within the continuous time window; The statistical distribution features are concatenated in a preset order to form a feature vector, which is then input into the currently activated auxiliary micro-motion pattern recognizer network. After forward calculation by the auxiliary micro-motion pattern recognizer network, the auxiliary micro-motion pattern determination result under the current subject motion posture category is output.

[0018] This invention calculates the acceleration magnitude deviation and combines it with the previous moment's attitude category to determine a specific threshold and a feasible motion subspace. Angular velocity is projected into this space for integration and updating. When the deviation condition is met, acceleration data is used for correction, reducing irrelevant noise interference and improving the accuracy and stability of attitude matrix calculation. The constrained attitude rotation matrix is ​​decomposed into a six-dimensional main attitude vector and a perturbation attitude vector, reducing the data processing dimensionality while separating the main motion trend from subtle perturbation features. The perturbation attitude vector magnitude is used to weight and adjust the output confidence of the main classification model, reducing the absolute value of the confidence under perturbation states. A low-confidence state label or confidence level auxiliary information is output in conjunction with a preset confidence threshold. By activating the corresponding micro-motion pattern recognizer, the distribution pattern of perturbation features within a continuous time window is deeply explored, achieving joint and accurate recognition of macroscopic main motion and microscopic detail actions, enhancing the refined perception capability of wearable devices in complex motion scenarios. Attached Figure Description

[0019] Figure 1 A flowchart of a motion posture recognition method for wearable smart devices; Figure 2 This is a schematic diagram showing the absolute value of the acceleration deviation. Figure 3 This is a schematic diagram illustrating the changing trends of statistical characteristics. Detailed Implementation

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] In the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] References to "one embodiment" or "some embodiments" as described in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in another embodiment," "in yet another embodiment," etc., appearing in different parts of this application specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0024] This specification provides one or more embodiments of a motion posture recognition method for wearable smart devices, such as... Figure 1 As shown, it includes the following steps: S1, acquire the angular velocity data and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold according to the subject's motion posture category at the previous moment and determine the feasible motion subspace.

[0025] The gyroscope and accelerometer in the inertial measurement unit (IMU) of the microelectromechanical system (MEMS) read triaxial angular velocity and triaxial acceleration data. The raw analog signals are preprocessed using a Butterworth low-pass filter to remove high-frequency noise, resulting in digital angular velocity and acceleration vectors. The square root of the sum of the squares of the three components of the triaxial acceleration vector is calculated as the current acceleration magnitude. The difference between the current acceleration magnitude and the standard gravitational acceleration magnitude of 9.80665 is calculated, and the absolute value of this difference is taken as the absolute value of the deviation. An attitude parameter mapping table is maintained in memory, indexed by the subject's motion attitude category. This table records the dynamic motion threshold and feasible subspace basis vector matrix corresponding to each subject's motion attitude category. The subject's motion attitude category label output by the attitude classification model at the previous moment is used as the key to query this hash table, obtaining the corresponding numerical dynamic motion threshold and a 3x2 unit orthogonal basis matrix composed of two mutually orthogonal three-dimensional column vectors with a magnitude of 1. This matrix serves as the basis vector matrix for determining the feasible subspace. When the current sampling period is the first sampling period after the system starts, or when the subject's motion posture category has not been effectively output at the previous moment, the preset initial subject motion posture category, default dynamic motion threshold, and default motion feasible subspace basis vector matrix are used for initialization query and posture calculation.

[0026] In some embodiments, calculating the absolute value of the deviation between the current acceleration modulus and the standard gravitational modulus based on the acceleration data includes: Calculate the vector magnitude corresponding to the acceleration data at the current sampling time; The absolute value of the deviation is obtained by subtracting the set standard gravity magnitude from the vector magnitude and taking the absolute value of the difference.

[0027] The control system extracts the three-dimensional acceleration data column vector at the current time t within each data sampling period. It then calls the floating-point unit to perform the sum of squares and square root operation of the three-dimensional Euclidean distance, calculating the vector magnitude. The pre-calibrated standard gravitational modulus constant is extracted from the underlying non-volatile storage parameter configuration area. In practical wearable device engineering applications, this constant is used to match the average gravity at the Earth's latitude. The preferred configuration range is limited to 9.80 to 9.82 m / s. 2 Between these values, a typical example is taken with the standard gravitational acceleration of 9.80665 m / s². 2 .

[0028] Perform basic floating-point subtraction to obtain the difference in magnitude. The difference is then converted into a non-negative absolute deviation value using a hardware-level absolute value arithmetic unit or a software-level absolute value function. For example, when the sensor samples triaxial acceleration data at a certain moment and the value is a specific value, the instantaneous vector magnitude of the acceleration data is calculated to be approximately 9.687; with a set standard gravity magnitude... When the value is 9.807, subtracting the two results in a negative difference of 0.120. Taking the absolute value of this difference gives the current absolute value of the deviation. =0.120. This scalar transformation calculation step maps acceleration fluctuations in three-dimensional space into a single one-dimensional evaluation index, characterizing the degree of deviation of the current acceleration data from the standard gravity reference. It provides a quantitative basis for subsequent dynamic threshold judgment and conditional attitude correction. This calculation process only involves square root, subtraction, and absolute value operations, making it suitable for low-latency execution in embedded processors. An illustration of the absolute value of acceleration deviation is shown below. Figure 2 As shown.

[0029] S2, the angular velocity data is projected onto the motion feasible subspace to obtain the projected angular velocity vector; the attitude rotation matrix is ​​updated by integration based on the projected angular velocity vector; when the absolute value of the deviation is lower than the dynamic motion threshold, the attitude rotation matrix updated by integration is corrected using the acceleration data to generate a constrained attitude rotation matrix.

[0030] Using linear algebra, based on the fact that the orthogonal basis matrix U satisfies The property of the current three-dimensional angular velocity data column vector The transpose of the orthogonal basis matrix Multiplication yields the projection coefficient vector The projection angular velocity vector is obtained by multiplying the orthogonal basis matrix U with the projection coefficient vector c. Subsequently, using the Runge-Kutta integral algorithm or the Picard integral algorithm, the antisymmetric matrix corresponding to the projected angular velocity vector is substituted into the attitude update equation. The third-order attitude rotation matrix of the previous moment is updated using the matrix exponential function or the Rodriguez rotation formula. The attitude rotation matrix obtained based on the integral of the projected angular velocity, before acceleration correction, is denoted as the projected integral attitude rotation matrix. The Runge-Kutta integral algorithm is a type of numerical integration method used to more accurately predict the state at the next moment by estimating the intermediate slope multiple times, given the current state and rate of change. A conditional judgment statement is written to compare the absolute value of the deviation with the dynamic motion threshold. When the condition is true (i.e., the absolute value of the deviation is lower than the dynamic motion threshold), it indicates that the current acceleration data mainly reflects the direction of gravity. The control program calculates the alignment error vector based on the gravity direction predicted by the projected integral attitude rotation matrix and the normalized acceleration direction. This alignment error vector generates a feedback micro-rotation vector, which is then right-multiplied and corrected using a local transformation matrix to generate a constrained attitude rotation matrix. When the absolute value of the deviation is not lower than the dynamic motion threshold, it is determined that the current acceleration data contains motion acceleration interference unsuitable for gravity direction correction. The control program does not perform attitude correction based on the acceleration direction and resets the projected integral attitude rotation matrix. Directly used as the constrained attitude rotation matrix for the current cycle This ensures that the subsequent steps of constructing the six-dimensional main attitude vector and calculating the perturbation attitude vector are executed continuously.

[0031] In some embodiments, projecting the angular velocity data onto the motion feasible subspace to obtain a projected angular velocity vector includes: Extract the set of unit orthogonal basis vectors from the pre-set library that are mapped to the subject's motion posture category at the previous moment to construct the subspace reference matrix; The product of the transpose of the subspace reference matrix and the column vector of the currently input angular velocity data is calculated as the projection coefficient vector; The projection angular velocity vector is obtained by multiplying the subspace reference matrix and the projection coefficient vector.

[0032] The system's internal read-only memory pre-configures a posture mapping library, storing unit orthogonal basis vector sets corresponding to dozens of common biological movement postures. Each unit orthogonal basis vector set consists of two mutually orthogonal three-dimensional unit column vectors with a magnitude of 1, used to define the two main degrees of freedom that can be retained under that subject's movement posture. During real-time operation, the control unit, based on the subject's movement posture category output by the posture classification model at the previous moment (e.g., if the previous movement posture category was determined to be walking on flat ground), indexes and retrieves the corresponding unit orthogonal basis vector set from the library, concatenates the two three-dimensional unit column vectors along their column directions, and constructs a subspace reference matrix U with a mathematical dimension of 3×2. The three-dimensional angular velocity data column vectors output in real-time by the current three-axis gyroscope are used. The hardware matrix multiplier is used to multiply the column vectors by the transpose of the subspace reference matrix. Perform linear algebraic multiplication to calculate the dimension-reduced two-dimensional projection coefficient vector. .

[0033] After obtaining the projection coefficients, the matrix operation instruction library is used to remap the dimensionality-reduced projection coefficient vector back to the original three-dimensional Cartesian physical space. This involves performing a matrix dot product operation between the subspace reference matrix U and the projection coefficient vector c to obtain a projection angular velocity vector that reduces interference from infeasible dimensions. As a specific application example, suppose that for a particular walking posture, the retrieved reference matrix consists exactly of two basis vectors extending along the horizontal plane. The projection coefficient vector obtained by multiplying the current angular velocity vector detected by the sensor with its transpose matrix is ​​then used to remap and calculate the projected angular velocity vector, which filters out interference components from infeasible directions. This ensures that the subsequent posture integration process mainly updates along the kinematic degrees of freedom corresponding to the current subject's motion posture. This projection constraint process based on matrix algebraic mapping ensures that the posture calculation proceeds along the physical kinematic degrees of freedom that conform to ergonomics and biomechanics.

[0034] In some embodiments, the step of using the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix includes: Multiply the transpose of the current attitude rotation matrix after integration update with the preset standard gravity direction unit vector to obtain the reference gravity direction vector in the device coordinate system. The alignment error vector is obtained by calculating the cross product of the normalized current acceleration data column vector and the reference gravity direction vector. The feedback micro-rotation vector is obtained by multiplying the set feedback proportional constant with the alignment error vector. Calculate the magnitude of the feedback micro-rotation vector; When the modulus is not lower than the preset small angle threshold, a local transformation matrix is ​​constructed according to the Rodriguez formula. The local transformation matrix is ​​the sum of a third-order identity matrix, a first correction term, and a second correction term. Wherein, the first correction term is the ratio of the sine of the modulus to the modulus multiplied by the antisymmetric matrix corresponding to the feedback micro-rotation vector; the second correction term is a matrix term obtained by multiplying the second correction coefficient by the square of the antisymmetric matrix corresponding to the feedback micro-rotation vector; wherein, the second correction coefficient is the ratio between the difference obtained by subtracting the cosine of the modulus and the square of the modulus. When the modulus is lower than a preset small angle threshold, the local transformation matrix is ​​calculated using the Taylor approximation values ​​corresponding to the ratio of the sine value to the modulus and the ratio of the cosine difference to the square of the modulus. The updated constrained attitude rotation matrix is ​​obtained by multiplying the current attitude rotation matrix updated by integration with the local transformation matrix.

[0035] To perform attitude fusion correction, the microcontroller will use a pre-calibrated global standard gravity direction unit vector. As an initial reference value, the projected integral attitude rotation matrix, which is initially updated by the projection integral of the gyroscope angular velocity, is obtained through matrix multiplication logic. transpose matrix Multiply by the reference value, that is, perform the operation. Extract the reference gravity direction vector in the device coordinate system. The control program normalizes the magnitude of the currently acquired triaxial acceleration data column vector to obtain a standard unit vector. and invoke the vector outer product operation instruction. Calculate the three-dimensional alignment error vector *e*, representing the correction direction required to align the reference gravity direction vector to the current acceleration direction. Retrieve the configured feedback scaling constant from memory. The preferred setting range for this parameter is limited to 0.01 to 0.05, typically 0.02. This scalar constant is then multiplied by the alignment error vector e to obtain the feedback micro-rotation vector. The Euclidean space magnitude of the feedback micro-rotation vector is obtained using the square root function. This represents the specific minute rotation angle value that the system actually needs to compensate for within the current period. After entering the local transformation matrix construction stage, the Lie algebra Rodrigues series expansion formula under nonlinear exponential mapping is used to ensure the orthogonality of the transformation matrix.

[0036] Extract the standard third-order identity matrix I from the matrix pool; calculate the first nonlinear correction coefficient and multiply it by the antisymmetric matrix uniquely determined by the feedback micro-rotation vector; calculate the second nonlinear correction coefficient and multiply it by the square of the antisymmetric matrix; perform matrix addition on these three parts to generate the local transformation matrix. Specifically, when θ is not lower than a preset small angle threshold When θ is below a preset small angle threshold, the first nonlinear correction coefficient is sinθ / θ, and the second nonlinear correction coefficient is (1-cosθ) / θ². When the first nonlinear correction coefficient is taken as 1-θ² / 6, and the second nonlinear correction coefficient is taken as 1 / 2-θ² / 24, the... Desirable radians to The values ​​are within the radian range. The calculation process of the aforementioned local transformation matrix selects the corresponding branch to execute based on the comparison result between the modulus and the preset small angle threshold. When the modulus is lower than the preset small angle threshold, the trigonometric ratio operation containing the modulus as the denominator is no longer directly executed, thus avoiding division by zero anomalies when the compensation angle approaches zero. In the calculation stage, the underlying DSP matrix multiplication module is activated to perform the projection integral attitude rotation matrix... Perform matrix multiplication sequentially with the local transformation matrix ΔR, i.e., perform the operation This fusion computing strategy reduces the risk of division-to-zero overflow anomalies that may be induced when the compensation angle approaches zero, and also reduces integral accumulation drift caused by gravity mapping through smoothing constraints.

[0037] S3, extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional main attitude vector; calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector.

[0038] The first and second 3D column vectors of the constrained attitude rotation matrix (3x3) are read using memory index slicing. These two column vectors are then vertically stacked and flattened in memory using array concatenation to form a one-dimensional array containing six floating-point elements, thus constructing the six-dimensional master attitude vector. Using the raw angular velocity data obtained above without projection into the feasible motion subspace, a Picard integral algorithm is performed to update and obtain a reference attitude rotation matrix without acceleration correction. The attitude rotation matrix The transpose matrix and the reference attitude rotation matrix Multiply them to calculate the relative error rotation matrix. Use the trace function to calculate the sum of the main diagonal elements of the relative error rotation matrix, and then use the relationship between this trace value and the rotation angle. Find the rotation angle, extract the rotation axis vector based on the antisymmetric part of the relative error rotation matrix, and construct a three-dimensional axis-angle vector representing the minimum rotation required for alignment as the perturbation attitude vector by multiplying the rotation angle with the rotation axis vector.

[0039] In some embodiments, the calculation of aligning the constrained attitude rotation matrix to the minimum rotational three-dimensional axis angle vector required to update the reference attitude rotation matrix based on the integral of the angular velocity data as the perturbation attitude vector includes: The reference attitude rotation matrix is ​​generated by integrating the angular velocity data from the previous moment. The alignment error matrix is ​​obtained by multiplying the transpose of the constrained attitude rotation matrix with the reference attitude rotation matrix. Calculate the trace of the alignment error matrix, and calculate the true rotation angle based on the inverse cosine relationship between the trace and the rotation angle; Extract the rotation axis unit vector from the antisymmetric portion of the alignment error matrix; A three-dimensional axis-angle vector with a magnitude of the true rotation angle and a direction that is the unit vector of the rotation axis is constructed as the perturbation attitude vector.

[0040] Within each fixed sampling period, the unprojected raw angular velocity data transmitted by the sensor is used. The sampling frequency of the system is preferably set to 100Hz to 500Hz, with 200Hz being a specific example. The system then uses the standard first-order Runge-Kutta numerical integration algorithm to sample the previous moment's body attitude rotation matrix stored in the register. Perform kinematic update calculations under conditions without external gravity constraints. Combine this with the Schmidt orthogonalization method to generate the pure kinematic reference attitude rotation matrix for the current time step. Trigger the matrix multiplier to convert the transpose of the constrained attitude rotation matrix determined in the current cycle. With the newly generated reference attitude rotation matrix Perform matrix multiplication, the formula is: This yields a third-order alignment error matrix representing the three-dimensional spatial difference between the output constrained state and the free and unconstrained state. To analyze the underlying rotational geometry of the alignment error matrix, the program iterates through the three core elements on the main diagonal of the error matrix and sums them to calculate the mathematical trace of the matrix.

[0041] Based on rotation matrix feature inversion, according to Solve for the true rotation angle required for system compensation The output unit is radians and the range is limited to the interval from zero to pi. The `clip` function is used to restrict the input of the inverse cosine function to the range [-1, 1] to avoid floating-point calculation errors causing abnormal inverse cosine operations. The purely antisymmetric part of the error matrix is ​​extracted using the difference subtraction method of the execution matrix, from which three independent off-diagonal difference elements determining the rotational attitude are decomposed. When the angle is below a preset small angle threshold, the perturbation attitude vector is set to zero or the perturbation attitude vector is constructed using the effective rotation axis direction of the previous sampling period; when... Not lower than the preset micro-angle threshold and When the denominator value is higher than a preset angle threshold, the off-diagonal independent difference element is divided by... Standardize and scale to extract the unit vector of the rotation axis; when Approaching π and When the value is not higher than the preset angle denominator threshold, candidate rotation axis directions are extracted based on the diagonal elements of the alignment error matrix, and normalized to obtain the rotation axis unit vector. The true rotation angle is then expressed in scalar form. By performing a scalar scalar multiplication with the rotation axis unit vector u in vector form, a vector with the error direction directional property and a magnitude equal to the rotation intensity is constructed. 3D axis-angle vector Assuming the trace of the alignment error matrix is ​​calculated to be 2.998 in actual operation, the true rotation angle obtained by arccosine is approximately 0.0447 radians. In the example where the unit vector of the rotation axis is given, the output perturbation attitude vector value will map to the corresponding deviation in radians.

[0042] S4, input the six-dimensional main attitude vector into the attitude classification model, use the magnitude of the perturbation attitude vector to adjust the output confidence of the attitude classification model in real time, and generate the current subject motion attitude and its confidence state; activate the corresponding auxiliary micro-motion pattern recognizer according to the current subject motion attitude; input the statistical distribution characteristics of the perturbation attitude vector in the continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and output the recognition result in combination with the current subject motion attitude.

[0043] Multiple lightweight auxiliary micro-motion pattern recognizer network models trained on different macroscopic movements are pre-loaded in memory. A multi-branch selection control structure, such as a switch statement, is written to receive the current subject motion posture label, load and initialize the weights of the specific auxiliary micro-motion pattern recognizer network corresponding to that label. A fixed-length sliding circular queue is initialized as a continuous time window, pushing newly generated perturbation posture vectors into the queue and removing the oldest data. All perturbation posture vectors in the current queue are traversed, and embedded statistical calculation functions, fixed-point or floating-point matrix operation functions are called to calculate the mean, variance, and skewness of each dimension of the three-dimensional vector. During the host computer training or simulation phase, equivalent statistical calculations are performed, and these parameters consistent with the statistical distribution characteristics are concatenated into a feature vector, which is input into the newly activated specific auxiliary micro-motion pattern recognizer network. The predicted label of the auxiliary micro-motion pattern is output through the network's forward computation. The current subject motion posture category label and the auxiliary micro-motion pattern label are merged and packaged into a single data object using string concatenation operations or data structure combinations, and output as the recognition result via serial communication or wireless transmission protocol.

[0044] In some embodiments, the step of real-time weighted adjustment of the output confidence of the posture classification model using the magnitude of the perturbation posture vector, and generating the current subject motion posture and its reliable state, includes: The square root of the sum of the squares of the components of the perturbation attitude vector on the three orthogonal axes is used as the magnitude of the perturbation attitude vector. The weighting adjustment coefficient is calculated based on the magnitude of the perturbation attitude vector and the set adjustment intensity parameter; wherein the weighting adjustment coefficient is the reciprocal of a first value, and the first value is the sum of 1 and the product of the adjustment intensity parameter and the magnitude. The product of the weighting adjustment coefficient and the initial confidence level output by the attitude classification model is calculated as the adjusted confidence level; The class with the highest initial confidence output by the posture classification model is selected as the current subject's motion posture, and a reliable state is generated based on the adjusted confidence corresponding to that class. When the adjusted confidence level is not lower than the preset confidence threshold, a normal confidence state is generated. A low-confidence status label is generated when the adjusted confidence level corresponding to the category is lower than the preset confidence threshold.

[0045] Upon receiving the input command, the independent numerical components of the three-dimensional perturbation attitude vector constructed in the current cycle along the three orthogonal physical axes X, Y, and Z in the device body coordinate system are extracted and set as parameters. , , The three floating-point parameters are squared concurrently, then summed, and finally the arithmetic square root is calculated. The expression is: This allows us to derive the magnitude M of the perturbation attitude vector, representing the current spatial attitude uncertainty and jitter intensity level of the device. The set adjustment intensity parameters are read from the read-only environment configuration file. This hyperparameter is used to control the confidence penalty caused by irregular external vibrations, with an optimal range of 10.0 to 50.0, and a typical preferred value of 20.0. The control unit obtains the weighting adjustment coefficient by performing a mixed operation of multiplication and constant addition followed by taking the reciprocal. This function constructs a decay weight mapping relationship that is between 0 and 1 and decreases monotonically as the modulus increases.

[0046] The pose classification model at the front end is a lightweight neural network model. The network model's structure includes a data input layer, multiple fully connected hidden layers for extracting pose mapping relationships, and a classification output layer using a normalization function. The input to this network model is a six-dimensional master pose vector constructed in the current sampling period. In embodiments requiring enhanced temporal stability, the six-dimensional master pose vectors from multiple consecutive sampling periods can be stacked sequentially to form a six-dimensional master pose vector sequence, but the multi-axis spatial kinematic feature vectors and frequency feature matrices, which are independent of the six-dimensional master pose vectors, are no longer used as alternative inputs.

[0047] The output of this network model is the initial confidence distribution features for all candidate biological motion posture categories. After the lightweight posture classification prediction model outputs an initial confidence distribution matrix covering each category through probability mapping, it is assumed that the initial confidence prediction value for a certain potential action category is... The decision control logic will adjust the scalar weighting coefficient w obtained from the calculation along with the initial confidence levels of each predicted category. Perform scalar multiplication step by step to obtain the adjusted confidence vector after decay and convergence. In practical inference scenarios, for example, when the calculated current frame perturbation mode value M reaches 0.05 radians due to severe user jitter, the calculated weight reduction adjustment coefficient is 0.50 under the default adjustment intensity parameter of 20. At this time, if the model determines that the initial confidence peak value of a certain main pose category was originally 0.90, after multiplication intervention adjustment, the confidence response intensity will drop to 0.45.

[0048] Since the weighting adjustment coefficient w is a positive scalar that applies uniformly to all candidate categories, this multiplication process does not change the relative size order between candidate categories. Therefore, the weighting adjustment in this embodiment is mainly used to reduce the absolute confidence level under perturbation conditions and to compare it with a preset confidence threshold. The decision evaluation system determines the subject motion posture category corresponding to the highest initial confidence level based on the initial confidence distribution output by the posture classification model, and compares the adjusted confidence level corresponding to this category with the preset confidence threshold to determine the confidence status of the output result. When the adjusted confidence level is not lower than the preset confidence threshold, the system determines that the current classification result is reliable, generates a normal confidence status, and directly uses the current subject motion posture to activate the corresponding auxiliary micro-motion pattern recognizer. When the adjusted confidence level corresponding to the subject motion posture category is lower than the preset confidence threshold, the system determines that it is currently in a high-perturbation or uncertain situation, adds a low-confidence status label to the subject motion posture category, and can temporarily use the previous stable subject motion posture as the activation basis for the auxiliary micro-motion pattern recognizer when the confidence level is lower than the preset confidence threshold for multiple consecutive sampling periods.

[0049] In some embodiments, the step of inputting the statistical distribution characteristics of the perturbation attitude vector within a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and combining the current subject motion posture to output the recognition result includes: Within a set continuous time window, continuously acquired perturbation attitude vectors are cached in the data queue; The mean, variance, and skewness of each independent component of the perturbation attitude vector in the data queue are extracted as the statistical distribution characteristics within the continuous time window; The statistical distribution features are concatenated in a preset order to form a feature vector, which is then input into the currently activated auxiliary micro-motion pattern recognizer network. After forward calculation by the auxiliary micro-motion pattern recognizer network, the auxiliary micro-motion pattern determination result under the current subject motion posture category is output.

[0050] A first-in, first-out (FIFO) circular data buffer with a fixed pointer addressing length is allocated in the memory region, and a continuous time window with a flexibly configurable time span is constructed and maintained. To cover the frequency band characteristics of normal human hand micro-movement cycles, the duration of this time window parameter is preferably set within the range of 0.5 seconds to 2.0 seconds. For example, under a fixed sampling rate of 100Hz, a sliding capacity data queue containing 50 to 200 sampling cycles is configured. The control kernel relies on timer interrupts to push the latest calculated 3D perturbation attitude vector into the front of the data queue in time sequence during each acquisition cycle, and automatically discards the oldest redundant historical data. As the buffer queue fills up, the data preprocessing engine immediately traverses the data queue through the vector operation instruction set, and calculates the arithmetic mean for evaluating the basic deviation from the independent 3D data sets of the perturbation attitude vector X-axis, Y-axis, and Z-axis using basic statistical formulas. The statistical distribution characteristics are extracted using three major physical and mathematical parameters: the overall variance used to measure high-frequency discrete volatility, and the distribution skewness S used to detect the nonlinear directional tendency of vibration. The changing trend of the statistical characteristics of the perturbation attitude vector within a continuous time window can be seen as follows: Figure 3 As shown, Figure 3 Using the time series as the horizontal axis and the normalized feature values ​​as the vertical axis, this diagram illustrates how statistical features such as the mean and variance change as the time window slides.

[0051] After feature parameter extraction, the vector combination unit initiates a data shaping process, concatenating the calculated X-axis, Y-axis, and Z-axis statistical subarrays one by one along the column vector direction to reconstruct a one-dimensional feature vector F with a fixed size of 9×1 dimensions. This concatenated feature vector is then loaded into the currently activated specific auxiliary micro-motion pattern recognizer network. The auxiliary micro-motion pattern recognizer network is a lightweight deep learning neural network model. The network structure includes an input layer responsible for receiving feature vector data, a hidden layer (preferably a fully connected layer or a two-layer one-dimensional convolutional neural network containing local receptive field features), and a terminal probability mapping output layer for multi-class classification. The network input is the one-dimensional feature vector F reconstructed through concatenation. The network output is a multi-path prediction confidence score array corresponding to the auxiliary micro-motion pattern under the current subject motion posture category. Following the network's multi-layer hidden layer multiplier-adder matrix operations and nonlinear forward inference propagation, the terminal neurons calculate and output the multi-path classification prediction confidence score array for the current working condition. Given that the main motion state has been determined, such as when it has been identified as a walking posture, micro-motion information is acquired, and the determination result of the auxiliary micro-motion mode under the current main posture is obtained.

[0052] One or more embodiments of this specification also provide a motion posture recognition system for wearable smart devices, including the following modules: The calculation module is used to acquire the angular velocity and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold according to the subject's motion posture category at the previous moment and determine the feasible motion subspace; The generation module is used to project the angular velocity data onto the motion feasible subspace to obtain a projected angular velocity vector; to perform an integral update on the attitude rotation matrix based on the projected angular velocity vector; and when the absolute value of the deviation is lower than the dynamic motion threshold, to use the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix. The construction module is used to extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional master attitude vector; and to calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector. The recognition module is used to input the six-dimensional main attitude vector into the attitude classification model, adjust the output confidence of the attitude classification model in real time using the magnitude of the perturbation attitude vector, and generate the current subject motion attitude and its confidence state; activate the corresponding auxiliary micro-motion pattern recognizer according to the current subject motion attitude; input the statistical distribution characteristics of the perturbation attitude vector in a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and output the recognition result in combination with the current subject motion attitude.

[0053] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for motion posture recognition in wearable smart devices, characterized in that, Includes the following steps: Acquire the angular velocity and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold and determine the feasible motion subspace according to the subject's motion posture category at the previous moment; The angular velocity data is projected onto the motion feasible subspace to obtain a projected angular velocity vector; the attitude rotation matrix is ​​updated by integration based on the projected angular velocity vector; when the absolute value of the deviation is lower than the dynamic motion threshold, the attitude rotation matrix updated by integration is corrected using the acceleration data to generate a constrained attitude rotation matrix. Extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional master attitude vector; calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector; The six-dimensional main attitude vector is input into the attitude classification model. The output confidence of the attitude classification model is adjusted in real time using the magnitude of the perturbation attitude vector, and the current subject motion attitude and its confidence state are generated. The corresponding auxiliary micro-motion pattern recognizer is activated according to the current subject motion attitude. The statistical distribution characteristics of the perturbation attitude vector in the continuous time window are input into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and the recognition result is output in combination with the current subject motion attitude.

2. The method according to claim 1, characterized in that, The calculation of the absolute value of the deviation between the current acceleration modulus and the standard gravitational modulus based on the acceleration data includes: Calculate the vector magnitude corresponding to the acceleration data at the current sampling time; The absolute value of the deviation is obtained by subtracting the set standard gravity magnitude from the vector magnitude and taking the absolute value of the difference.

3. The method according to claim 1, characterized in that, The step of projecting the angular velocity data onto the motion feasible subspace to obtain the projected angular velocity vector includes: Extract the set of unit orthogonal basis vectors from the pre-set library that are mapped to the subject's motion posture category at the previous moment to construct the subspace reference matrix; The product of the transpose of the subspace reference matrix and the column vector of the currently input angular velocity data is calculated as the projection coefficient vector; The projection angular velocity vector is obtained by multiplying the subspace reference matrix and the projection coefficient vector.

4. The method according to claim 3, characterized in that, The step of using the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix includes: Multiply the transpose of the current attitude rotation matrix after integration update with the preset standard gravity direction unit vector to obtain the reference gravity direction vector in the device coordinate system. The alignment error vector is obtained by calculating the cross product of the normalized current acceleration data column vector and the reference gravity direction vector. The feedback micro-rotation vector is obtained by multiplying the set feedback proportional constant with the alignment error vector. Calculate the magnitude of the feedback micro-rotation vector; When the modulus is not lower than the preset small angle threshold, a local transformation matrix is ​​constructed according to the Rodriguez formula. The local transformation matrix is ​​the sum of a third-order identity matrix, a first correction term, and a second correction term. Wherein, the first correction term is the ratio of the sine of the modulus to the modulus multiplied by the antisymmetric matrix corresponding to the feedback micro-rotation vector; the second correction term is a matrix term obtained by multiplying the second correction coefficient by the square of the antisymmetric matrix corresponding to the feedback micro-rotation vector; wherein, the second correction coefficient is the ratio between the difference obtained by subtracting the cosine of the modulus and the square of the modulus. When the modulus is lower than a preset small angle threshold, the local transformation matrix is ​​calculated using the Taylor approximation values ​​corresponding to the ratio of the sine value to the modulus and the ratio of the cosine difference to the square of the modulus. The updated constrained attitude rotation matrix is ​​obtained by multiplying the current attitude rotation matrix updated by integration with the local transformation matrix.

5. The method according to claim 1, characterized in that, The calculation aligns the constrained attitude rotation matrix to the minimum rotational three-dimensional axis angle vector required to update the reference attitude rotation matrix based on the integral of the angular velocity data, using this vector as the perturbation attitude vector. This includes: The reference attitude rotation matrix is ​​generated by integrating the angular velocity data from the previous moment. The alignment error matrix is ​​obtained by multiplying the transpose of the constrained attitude rotation matrix with the reference attitude rotation matrix. Calculate the trace of the alignment error matrix, and calculate the true rotation angle based on the inverse cosine relationship between the trace and the rotation angle; Extract the rotation axis unit vector from the antisymmetric portion of the alignment error matrix; A three-dimensional axis-angle vector with a magnitude of the true rotation angle and a direction that is the unit vector of the rotation axis is constructed as the perturbation attitude vector.

6. The method according to claim 1, characterized in that, The step of using the magnitude of the perturbation attitude vector to perform real-time weighted adjustment of the output confidence of the attitude classification model, and generating the current subject motion attitude and its reliable state, includes: The square root of the sum of the squares of the components of the perturbation attitude vector on the three orthogonal axes is used as the magnitude of the perturbation attitude vector. The weighting adjustment coefficient is calculated based on the magnitude of the perturbation attitude vector and the set adjustment intensity parameter; wherein the weighting adjustment coefficient is the reciprocal of a first value, and the first value is the sum of 1 and the product of the adjustment intensity parameter and the magnitude. The product of the weighting adjustment coefficient and the initial confidence level output by the attitude classification model is calculated as the adjusted confidence level; The class with the highest initial confidence output by the posture classification model is selected as the current subject's motion posture, and a reliable state is generated based on the adjusted confidence corresponding to that class. When the adjusted confidence level is not lower than the preset confidence threshold, a normal confidence state is generated. A low-confidence status label is generated when the adjusted confidence level corresponding to the category is lower than the preset confidence threshold.

7. The method according to claim 1, characterized in that, The step of inputting the statistical distribution characteristics of the perturbation attitude vector within a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and combining the current subject motion posture to output the recognition result, includes: Within a set continuous time window, continuously acquired perturbation attitude vectors are cached in the data queue; The mean, variance, and skewness of each independent component of the perturbation attitude vector in the data queue are extracted as the statistical distribution characteristics within the continuous time window; The statistical distribution features are concatenated in a preset order to form a feature vector, which is then input into the currently activated auxiliary micro-motion pattern recognizer network. After forward calculation by the auxiliary micro-motion pattern recognizer network, the auxiliary micro-motion pattern determination result under the current subject motion posture category is output.

8. A motion posture recognition system for wearable smart devices, characterized in that, Includes the following modules: The calculation module is used to acquire the angular velocity and acceleration data of the device; calculate the absolute value of the deviation between the current acceleration modulus and the standard gravity modulus based on the acceleration data; query the dynamic motion threshold according to the subject's motion posture category at the previous moment and determine the feasible motion subspace; The generation module is used to project the angular velocity data onto the motion feasible subspace to obtain a projected angular velocity vector; to perform an integral update on the attitude rotation matrix based on the projected angular velocity vector; and when the absolute value of the deviation is lower than the dynamic motion threshold, to use the acceleration data to correct the integrally updated attitude rotation matrix to generate a constrained attitude rotation matrix. The construction module is used to extract the first two columns of the constrained attitude rotation matrix to construct a six-dimensional master attitude vector; and to calculate the minimum three-dimensional rotation axis angle vector required to align the constrained attitude rotation matrix to the reference attitude rotation matrix updated based on the integral of the angular velocity data as the perturbation attitude vector. The recognition module is used to input the six-dimensional main attitude vector into the attitude classification model, adjust the output confidence of the attitude classification model in real time using the magnitude of the perturbation attitude vector, and generate the current subject motion attitude and its confidence state; activate the corresponding auxiliary micro-motion pattern recognizer according to the current subject motion attitude; input the statistical distribution characteristics of the perturbation attitude vector in a continuous time window into the activated auxiliary micro-motion pattern recognizer to identify the auxiliary micro-motion pattern, and output the recognition result in combination with the current subject motion attitude.

9. The system according to claim 8, characterized in that, The calculation of the absolute value of the deviation between the current acceleration modulus and the standard gravitational modulus based on the acceleration data includes: Calculate the vector magnitude corresponding to the acceleration data at the current sampling time; The absolute value of the deviation is obtained by subtracting the set standard gravity magnitude from the vector magnitude and taking the absolute value of the difference.

10. The system according to claim 8, characterized in that, The step of projecting the angular velocity data onto the motion feasible subspace to obtain the projected angular velocity vector includes: Extract the set of unit orthogonal basis vectors from the pre-set library that are mapped to the subject's motion posture category at the previous moment to construct the subspace reference matrix; The product of the transpose of the subspace reference matrix and the column vector of the currently input angular velocity data is calculated as the projection coefficient vector; The projection angular velocity vector is obtained by multiplying the subspace reference matrix and the projection coefficient vector.

Images