An identity detection method and system based on inertial sensor

Abstract

The present invention discloses an identity detection method and system based on an inertial sensor. The method comprises: extracting effective motion data; performing attitude resolution on effective motion data based on an attitude resolution model of an extended Kalman filter with dynamic noise feedback; restoring motion trajectory based on attitude information; extracting features of effective motion data and motion trajectory in time domain and frequency domain respectively to obtain multiple motion feature vectors; randomly generating multiple groups of interfering motion data and extracting feature vectors of the interfering motion data; training a random forest model based on interfering motion data feature vectors, motion feature vectors, action categories corresponding to interfering motion data feature vectors, and action categories corresponding to motion feature vectors to obtain a classification model; and performing identity detection through the classification model. The present invention can realize accurate action recognition and identity detection on a small sample data set.

Description

Technical Field

[0001] The present invention relates to the technical field of identity detection, and in particular to an identity detection method and system based on an inertial sensor. Background Art

[0002] Inertial sensors are widely used in a variety of small mobile terminals such as mobile phones, watches, and remote controls due to their portability, low power consumption, and low cost. In the future era of the Internet of Things, human motion sensing technology based on smart terminals and intelligent algorithms will become an important part of life. Among them, using the sensing technology of inertial sensors to identify different individual users is a major research hotspot in society.

[0003] Using inertial sensor data for identity detection is actually analyzing the motion data of specific actions of different users. The key to this step is that the algorithm model extracts effective features from the motion data of specific actions and gives different output results according to different features. In this process, multiple frames and frame losses of data corresponding to specific actions, and errors in the calculation of signal features will cause the loss of motion characteristics, affecting the feature extraction of identity detection tasks and the classification results of the model. Therefore, in the task of using inertial sensors for identity detection, there are two important issues: one is how to accurately extract the inertial sensor data corresponding to specific actions, and the other is how to extract effective motion features from the inertial sensor data corresponding to specific actions.

[0004] In addition, the low production cost and construction mechanism of inertial sensors result in limited measurement accuracy and the inability to provide accurate motion information. However, in identity detection tasks, motion trajectory is the most intuitive feature quantity, and an accurate motion trajectory can almost reflect most of the characteristics of an action. In terms of inertial navigation, laser, radar, vision and other sensors are usually used to fuse motion information and correct the trajectory solution results of the inertial sensor, which will make the task have greater limitations and bring higher costs.

[0005] In essence, identity detection is a classification task. In order to obtain a high accuracy rate, a large amount of training data needs to be provided in advance and useful information needs to be obtained from this data. In the identity detection task, the simple data collection process guarantees the user's human-computer interaction experience. Therefore, under the premise of ensuring recognition accuracy, how to train the model to achieve ideal results on a small sample set is also a major problem. Summary of the invention

[0006] Based on the above problems, the present invention provides an identity detection method and system based on inertial sensors. Through a large number of optimization projects (data segmentation, design of filtering algorithms), feature engineering (solving posture trajectories, extracting frequency domain and time domain statistical features), and designing machine learning models, accurate action recognition and identity detection can be achieved on small sample data sets.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] An identity detection method based on an inertial sensor, comprising:

[0009] Collecting motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points;

[0010] Extracting effective motion data from the motion data;

[0011] An attitude calculation model of an extended Kalman filter based on dynamic noise feedback is used to calculate the attitude of the effective motion data to obtain attitude information;

[0012] Restoring the motion trajectory based on the posture information;

[0013] Extracting features of the effective motion data and the motion trajectory in the time domain and the frequency domain respectively to obtain a plurality of motion feature vectors;

[0014] Randomly generate multiple groups of interfering motion data, and extract feature vectors of the interfering motion data;

[0015] Training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model;

[0016] Identity detection is performed through the classification model.

[0017] Optionally, extracting effective motion data from the motion data specifically includes:

[0018] Based on the motion data, a multi-sliding window method is used to determine motion state data;

[0019] Based on the motion state data, a multi-sliding window method is used to determine effective data segmentation points;

[0020] Valid motion data is determined based on the valid data segmentation points.

[0021] Optionally, the process of determining the attitude solution model of the extended Kalman filter based on dynamic noise feedback is as follows:

[0022] Calculate initial state information based on the valid motion data;

[0023] The weight of the extended Kalman filter is adjusted based on the initial state information to obtain an attitude solution model of the extended Kalman filter based on dynamic noise feedback.

[0024] Optionally, training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector specifically includes:

[0025] Inputting the interference motion data feature vector and the motion feature vector into a first random forest model to obtain a first output result;

[0026] Determining whether the first output result is an action category corresponding to the feature vector of the interference motion data;

[0027] If so, the output is rejected;

[0028] If not, the motion feature vector is input into a second random forest model to obtain the probability of each action category to which the motion feature vector belongs;

[0029] If the maximum probability is less than the probability threshold, the output is rejected;

[0030] If the maximum probability is greater than the probability threshold, the action category corresponding to the maximum probability is taken as the classification result.

[0031] The present invention provides an identity detection system based on an inertial sensor, comprising:

[0032] A data acquisition module, used to collect motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points;

[0033] An effective motion data extraction module, used to extract effective motion data from the motion data;

[0034] An attitude information solving module is used for solving the attitude solving model of the extended Kalman filter based on dynamic noise feedback, solving the attitude of the effective motion data to obtain attitude information;

[0035] A motion trajectory restoration module, used to restore the motion trajectory based on the posture information;

[0036] A feature extraction module, used to extract features from the effective motion data and the motion trajectory in the time domain and the frequency domain respectively, to obtain a plurality of motion feature vectors;

[0037] An interference motion data feature vector generation module, used to randomly generate multiple groups of interference motion data and extract feature vectors of the interference motion data;

[0038] A training module, used for training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model;

[0039] The identity detection module is used to perform identity detection through the classification model.

[0040] Optionally, the effective motion data extraction module specifically includes:

[0041] A motion state data determining unit, configured to determine the motion state data based on the motion data by using a multi-sliding window method;

[0042] An effective data segmentation point determination unit, configured to determine effective data segmentation points based on the motion state data by using a multi-sliding window method;

[0043] The valid motion data determining unit is used to determine the valid motion data based on the valid data segmentation points.

[0044] Optionally, the posture information solving module includes:

[0045] an initial state information calculation unit, configured to calculate initial state information based on the valid motion data;

[0046] The attitude solution model determination unit is used to adjust the weight of the extended Kalman filter based on the initial state information to obtain an attitude solution model of the extended Kalman filter based on dynamic noise feedback.

[0047] Optionally, it specifically includes:

[0048] A first input unit, used for inputting the interference motion data feature vector and the motion feature vector into a first random forest model to obtain a first output result;

[0049] A judging unit, configured to judge whether the first output result is an action category corresponding to a feature vector of the interference motion data;

[0050] A first output unit, configured to output a rejection of recognition when the first output result is an action category corresponding to the feature vector of the interference motion data;

[0051] A second input unit is used for inputting the motion feature vector into a second random forest model to obtain the probability of each action category to which the motion feature vector belongs when the first output result is not the action category corresponding to the interference motion data feature vector;

[0052] The second output unit is used to output a rejection of recognition when the maximum probability is less than the probability threshold; and to use the action category corresponding to the maximum probability as the classification result when the maximum probability is greater than the probability threshold.

[0053] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0054] 1) The present invention performs segmentation and extraction of inertial sensor data based on multi-resolution features, can accurately separate the effective motion data collected by the inertial sensor, eliminate redundant interference data, and extract sufficient effective features from concise specific actions;

[0055] 2) The present invention performs attitude solution and trajectory restoration based on the attitude solution model of the extended Kalman filter with dynamic noise feedback. During the attitude solution process, the noise parameters in the filter are fed back according to the measured values ​​of acceleration and angular velocity, and the weight of the filter is adjusted, which greatly improves the dynamic accuracy of attitude solution.

[0056] 3) The present invention performs a lot of feature processing on the original signal, screens out some feature indicators, and builds an identity detection algorithm based on the traditional machine learning model. Due to the completion of a large amount of feature engineering in the early stage, the model not only has a very small amount of calculation, but can also give detection results in real time in specific tasks, and has a high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0058] Figure 1 is a flow chart of an identity detection method based on an inertial sensor according to an embodiment of the present invention;

[0059] Figure 2 For action 1, "press the button";

[0060] Figure 3 For action 2, "vibrate";

[0061] Figure 4 This is the segmentation result of “Action 1”;

[0062] Figure 5 The segmentation result for “action 2”;

[0063] Figure 6 It is the calculation process of the classification model;

[0064] Figure 7 The overall process of identity detection. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The purpose of the present invention is to provide an identity detection method and system based on inertial sensors, which can realize accurate action recognition and identity detection on a small sample data set through a large number of optimization projects (data segmentation, design of filtering algorithms), feature engineering (solving posture trajectories, extracting frequency domain and time domain statistical features), and designing a machine learning model.

[0067] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] The present invention is based on the identity detection task of inertial sensors. In essence, it performs feature engineering on the inertial sensor data corresponding to a specific action, screens out favorable feature information, and designs a corresponding classification model. The present invention has made optimizations and innovations in feature engineering and model design. Below, it will be specifically introduced from the perspectives of experimental design and data acquisition, extraction of effective motion data based on multi-resolution features, improved attitude solution model of extended Kalman filtering based on dynamic noise feedback, motion trajectory solution based on attitude quaternions, feature extraction of original data and trajectory data in time domain and frequency domain, identity detection, and other processes.

[0069] like Figure 1 As shown, the present invention provides an identity detection method based on an inertial sensor, comprising the following steps:

[0070] Step 101: collecting motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points.

[0071] To use inertial sensors alone to complete identity detection, you first need to design specific actions and collect data on the corresponding actions to create a data set. (Inertial sensors can measure the three-dimensional acceleration and three-dimensional angular velocity data of moving objects, and are small in size and easy to wear.)

[0072] Taking the detection of TV users as an example: Fix the inertial sensor to the remote control, then place the remote control on the table and keep it relatively stationary. At this time, open the script program for storing data on the computer side. (Before performing specific actions, keep the sensor stationary, which facilitates the subsequent segmentation and extraction of valid data.)

[0073] The experimenter picks up the remote control, performs specific actions in the air, and then places the sensor back on the table. (The present invention designs two specific actions, requiring the actions to be simple enough and to contain certain motion information of the experimenter. One specific action is for the experimenter to pick up the remote control and perform the action of "pressing a button" in the air, denoted as Action 1, as shown in Figure 2 . Another is for the experimenter to pick up the remote control and perform the action of "vibrating" in the air, denoted as Action 2, as shown in Figure 3 .)

[0074] After completing the acquisition of motion data, close the data storage script on the computer side. (Assume that the inertial sensor starts recording data at time t1, starts performing specific actions at time t2, ends the specific actions at time t3, and stops recording data at time t4. Obviously, t1 < t2 < t3 < t4, and the data recorded by the sensor during the period from t1 to t2 and from t3 to t4 is the data in the stationary state. The data recorded by the sensor during the period from t2 to t3 is called the valid motion data.)

[0075] Step 102: Extract valid motion data from the action data. Specifically, it includes: determining motion state data by using the method of multiple sliding windows based on the motion data; determining valid data segmentation points by using the method of multiple sliding windows based on the motion state data; and determining valid motion data based on the valid data segmentation points.

[0076] In subsequent identity detection, the three-dimensional acceleration and three-dimensional angular velocity data collected by the sensor during the period from t2 to t3 are required, and the key to data preprocessing is to find the two time points t2 and t3 within the interval [t1, t4]. Next, the present invention uses the method of multiple sliding windows to determine t2 and t3.

[0077] Denote S’=(s’1, s’2, …, s’ m ) as a segment of time series signal data containing m data points collected in Step 101. Select a sliding window with a window length of L and a sliding step size of 1, and slide through the time series signal data S’ respectively. Record the variance of the data within the sliding window during the process as

[0078] V=(v1, v2, …, v m-L+1 )

[0079] The present invention defines that within the window, if the variance is greater than ε=0.05, the sensor is in motion, which is recorded as 1, otherwise it is in a stationary state, which is recorded as 0.

[0080] State=(state1,state2,…,state m-L+1 )

[0081] The specific action data collected has a relatively obvious pattern. In the time period before and after the movement, the sensor is in a static state, and the variance of the data segment that the window slides over is close to 0, while the variance is larger during the movement period. Therefore, the state quantity should satisfy the following formula

[0082]

[0083] For the state quantity State, the present invention uses a sliding window again, and the window length is the same as before, which is L and 1 respectively. If within the window, the sum of the state values ​​of the first L / 2 is 0, and the sum of the state values ​​of the last L / 2 is L / 2, then the midpoint position of the sliding window is recorded as t2; conversely, if within the window, the sum of the state values ​​of the first L / 2 is L / 2, and the sum of the state values ​​of the last L / 2 is 0, then the midpoint position of the sliding window is recorded as t3. In this way, given the window length L, the valid data segmentation point of the time series data S' can be determined.

[0084] The selection of the window size directly affects the result of the segmentation point: if the window is too long, the details of the timing signal will be ignored, and the obtained segmentation point will also be greatly different from the actual segmentation point. If the window is too short, the window will overemphasize the signal details, and the segmentation point will also be affected. Therefore, the present invention sets a group of k different window lengths, the window group L = (l1, l2, ..., l k ), for the same time series signal data S', use k windows to slide the windows respectively. After the sliding window is completed, the results of each sliding window are weighted to obtain the segmentation point, such as Figure 4-5 As shown. The weight factor is the inverse of the window length, and the split points are as follows

[0085]

[0086] Step 103: Based on the attitude calculation model of the extended Kalman filter with dynamic noise feedback, the effective motion data is subjected to attitude calculation to obtain attitude information.

[0087] Step 102 obtains effective motion data S = (s1, s2, ..., s n ), where s k (k=1,2,…,n) contains the three-dimensional acceleration and three-dimensional angular velocity at the kth data point corresponding to the moment

[0088]

[0089] The data measured by the accelerometer in the inertial sensor not only includes the linear acceleration of the sensor in the object coordinate system, but also includes the component of the gravity acceleration G in the object coordinate system. To obtain the linear acceleration of the actual motion, it is necessary to filter out the gravity acceleration. If the object's posture at each moment is known, the component of the gravity acceleration in the object coordinate system can be accurately known, and the gravity acceleration component can be eliminated to obtain the linear acceleration of the actual motion. In terms of posture resolution, accelerometers and gyroscopes have corresponding methods.

[0090] When the sensor is stationary or moving at a constant speed, the data measured by the accelerometer is the acceleration of gravity, so the accelerometer can be used to determine the attitude. In actual motion, linear acceleration other than gravity acceleration is definitely included, so there will be a difference between the accelerometer data and the acceleration of gravity, which will cause errors in the output results of the attitude estimation model.

[0091] The angular velocity data measured by the gyroscope in the inertial sensor can also be used to calculate the motion posture of the object in real time. However, in the process of using the gyroscope to calculate the posture, there will be cumulative errors.

[0092] The accuracy of inertial sensors is limited, and using accelerometers and gyroscopes alone to calculate attitude will have corresponding defects. Therefore, the present invention designs an attitude calculation model that uses state information to adjust the weights of the extended Kalman filter. The model can adjust the noise parameters in the filter according to the measured values ​​of acceleration and angular velocity during the attitude calculation process to improve the dynamic accuracy of attitude calculation.

[0093] 1. Initial attitude quaternion solution

[0094] First, calculate the attitude angle φ at the initial moment 0 ,θ 0 ,ψ 0 , and then convert the attitude angle into attitude quaternion Since the accelerometer data is the acceleration of gravity in a stationary state, the initial pitch angle φ 0 , roll angle θ 0 It can be obtained based on the accelerometer data in a stationary state. The initial heading angle ψ 0 A magnetometer is needed to confirm, but trajectory restoration is different from navigation and does not require an absolute initial heading angle, so the initial heading angle can be directly set to 0.

[0095]

[0096]

[0097] ψ0 =0

[0098] The rotation matrix from the ground coordinate system to the object coordinate system is recorded as It can be represented by three attitude angles or by attitude quaternion. The present invention uses the initial attitude angle to calculate the rotation matrix at the initial moment Then calculate the initial attitude quaternion through the rotation matrix

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] 2. Design an attitude solution model that uses state information to adjust the weights of the extended Kalman filter

[0105] After determining the initial value of the attitude, the next step is to use the filtering method to fuse the state quantity and observation quantity of the attitude quaternion. Define the state quantity as the attitude quaternion The observed quantities are the pitch angle φ and the roll angle θ. The state equation and observation equation can be expressed as follows, where u and v are process noise and observation noise, respectively.

[0106]

[0107] After defining the above state quantities and observation quantities, the system matrix F can be derived using the first-order Runge-Kutta method, and the attitude transformation matrix The output matrix H(X k ).

[0108]

[0109]

[0110] In the design of the filter, the present invention considers process noise and measurement noise as design parameters to minimize the covariance of the error estimate.

[0111]

[0112]

[0113]

[0114] in Yes x, w y , w z The mean of Yes x , w y , w z The present invention provides the covariance matrix of process noise based on this assumption.

[0115]

[0116] The deviation part on the right side of the above equation can be regarded as process noise. For this discrete system, the output variance of the gyroscope is The covariance of the noise Q can be written as

[0117]

[0118] By performing Taylor expansion on the pitch angle and roll angle, the covariance of the measurement noise can be obtained.

[0119]

[0120] The present invention

[0121]

[0122] For this discrete system, the output variance of the accelerometer is The covariance of the noise R in the observation process can be written as

[0123]

[0124] In the dynamic model, according to the characteristics of the state quantity at each moment, the present invention designs the correction parameters t a and t w , and its value is positively correlated with the intensity of exercise.

[0125]

[0126]

[0127] The present invention further provides a correction coefficient λ a , w , the corrected covariance is as follows. Q0 and R0 are related to the state quantity and the observed quantity. The covariance parameters of acceleration and angular velocity can be adjusted in a stationary state to obtain satisfactory parameter values. Specifically, t represents the estimated state quantity. If t is large, the object is in an accelerated motion or attitude change state, which will cause errors in the attitude solution model. Therefore, λ is used to adjust the covariance matrix to achieve attitude correction.

[0128] Q=Q0+λ a ta

[0129] R=R0+λ w t w

[0130] After determining the state quantity, observation quantity, and noise covariance matrix of the extended Kalman filter through the above process, the attitude quaternion values ​​corresponding to n time points can be solved.

[0131] Based on the above definitions and derivations, the following is the attitude solution process of the extended Kalman filter based on dynamic noise feedback:

[0132] Prediction process: First calculate the covariance matrix of the process noise (Equation (2)-Equation (4)), and then predict the attitude quaternion and the prior estimated covariance at the current moment based on the attitude quaternion and the estimated covariance matrix at the k-1 moment (Equation (5)-Equation (6)).

[0133]

[0134]

[0135]

[0136] Q=Q0+λ a t a (4)

[0137]

[0138]

[0139] Update process: First calculate the covariance matrix of the observation process noise (Equation (7)-Equation (10)), then combine the observation matrix and prior covariance to further determine the gain matrix (Equation (12)), then combine the observations to update the state quantity (Equation (13)), normalize the attitude quaternion (Equation (14)), and then update the covariance matrix (Equation (14)). In this update process, H is Equation (11) H(X k )’s Jacobian matrix, the specific value of which can be calculated by computer during the iteration process).

[0140]

[0141]

[0142]

[0143] R=R0+λ w t w (10)

[0144]

[0145]

[0146]

[0147]

[0148]

[0149] Through the iterative operation of the above steps, the effective motion data S = (s1, s2, ..., s n ) Attitude quaternion information at n time points.

[0150] Step 104: restoring the motion trajectory based on the posture information.

[0151] The present invention uses the posture information at time k to convert the accelerometer data measured by the sensor in the object coordinate system to the ground coordinate system. After subtracting the gravity acceleration from the converted data, the linear acceleration data of the actual movement of the object can be obtained. Then the linear acceleration data is integrated twice, and the motion trajectory data of the object can be obtained by accumulation. The specific steps are as follows:

[0152] At time k, the data measured by the gravity accelerometer is transformed from the object coordinate system to the ground coordinate system with the help of the rotation matrix determined by the attitude quaternion solved at time k-1. Then, the gravity acceleration G is removed from the transformed accelerometer data to obtain the acceleration information of the object in the ground coordinate system.

[0153]

[0154]

[0155]

[0156]

[0157] Through recursive operation, the motion trajectory information of n moments can be obtained. Therefore, the present invention integrates the data to obtain a matrix of size 9*n, and the 9 rows of the matrix are 9 time series data of three-dimensional acceleration, three-dimensional angular velocity, and three-dimensional motion trajectory.

[0158] Step 105: extracting features from the effective motion data and the motion trajectory in the time domain and the frequency domain respectively to obtain a plurality of motion feature vectors.

[0159] Time domain features refer to some time-related features reflected in the process of sequence changes over time. Combined with the experimental results, this paper selects 9 time domain features, namely mean, variance, maximum value, minimum value, range, number of zero crossings, mean number of points, correlation coefficient and signal amplitude area.

[0160] In the selected window, the mean reflects the representativeness of the data and reflects the overall numerical size of the signal. The variance reflects the degree of deviation of a set of data from its average value and characterizes the discreteness of the signal data. The maximum value, minimum value and range reflect the vibration amplitude of a specific action, which is of great help to the interval positioning of the action. The number of zero crossings, that is, the number of intersections between the signal and the X-axis, reflects the frequency of directional changes in the motion feature quantity. Similar to the number of zero crossings, the mean number of points, that is, the number of mean points taken by the signal, measures the frequency of changes in the motion trend. The correlation coefficient refers to the correlation between two parameters, which can be used to identify the direction of motion. The signal amplitude area refers to the sum of the areas enclosed by the signal curve and the coordinate axis. This feature is more obvious during the motion process.

[0161] Before extracting the frequency domain features, the present invention needs to perform a fast Fourier transform on the time series signal to obtain a frequency domain signal. In addition to the DC component, the present invention mainly selects the power spectrum features, which can be divided into amplitude statistical features and shape statistical features. Similarly, combined with the experimental results, the present invention selects 9 frequency domain features, which are the DC component, amplitude, and the mean, variance, skewness, and kurtosis of the shape.

[0162] In the signal spectrum, the DC component is the first component after Fourier transform, which reflects the mean characteristics of these signals. Unlike the mean and variance in the time domain, the mean and variance extracted from the spectrum reflect the overall representativeness and discreteness of the data signal in the frequency domain. The skewness feature reflects the distribution form of the data and describes the symmetry of the overall value distribution of the signal. Similar to skewness, the kurtosis feature reflects the steepness of the distribution form of the data signal value.

[0163] For 9*n specific action data, the present invention extracts the above 9+9=18 features from 9 time series signals (three-axis acceleration, angular velocity, motion trajectory), and obtains a total of 9*18=162 feature quantities. In the subsequent identity detection model, for each time series signal collected for a specific action, the present invention converts it into these 162 feature quantities through feature processing.

[0164] Step 106: randomly generate multiple groups of interfering motion data, and extract feature vectors of the interfering motion data.

[0165] Through artificial collection and random generation, a total of b groups of interference data were generated, and the b data were given a separate label to prepare for the "rejection of recognition" of the subsequent classification model.

[0166] Step 107: training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model.

[0167] The timing of a set of specific actions can be converted into 162 features through data segmentation, posture solution, trajectory restoration, and feature extraction.

[0168] In different scenarios, the focus of identity detection is different. Let's take TV users as an example: in a family, only family members are allowed to enable the TV, and other strangers cannot enable it. At this time, identity detection can be understood as an "unlock" function. In this case, the focus of identity detection is on the accuracy of detection, and the success rate of detection is relatively ignored. Perhaps family members' attempts to unlock the TV sometimes fail, but for outsiders, their attempts to unlock the TV will inevitably fail. In addition, suppose a new scenario: after the TV is turned on, different content needs to be pushed based on the identified user results. In this case, the focus of identity detection is to find the specific user, rather than giving the result of "recommendation failed".

[0169] Assuming that m experimenters are given the command "action 1", each experimenter collects n sets of action data, and a total of mn sets of "action 1" data are obtained, and the data labels are recorded at the same time. After processing in subsequent steps, the time series data of the action can be converted into 162 features.

[0170]

[0171] Y1=(y1,y2,...,y mn ,y mn+1 ,...,y mn+b )

[0172]

[0173] Y2=(y1,y2,...,y mn )

[0174] The data set [X1, Y1] contains mn+b data, with a total of m+1 categories, of which b data are a single category and the remaining mn data are m categories. Using this data set, the present invention trains a random forest 1 model. Given the input action data features, the model can output the probability probs1 that the data belongs to each category. The present invention defines the category with the largest probability as the classification result, and records the probability as max_probs1. If the classification result is the category to which the b interference data belongs, the model output refuses to recognize, otherwise it will turn to the second classification model.

[0175] The data set [X2, Y2] contains mn data, with a total of m categories and n data in each category. Using this data set, the present invention trains a random forest 2 model. Different from random forest 1, this model does not include interference data during training, and the output results of the model are more biased towards the known m categories. Given the input action data features, the model can output the probability probs2 that the data belongs to each category. The present invention defines the category with the highest probability as the classification result, and records the probability as max_probs2. In this model, the present invention sets a threshold parameter k in the range of [0,1]. Compare max_probs2 with the threshold. If the recognition rate exceeds the threshold, the model outputs the classification result, otherwise it still outputs a rejection of recognition. The calculation process of the classification model is as follows: Figure 6 shown.

[0176] Step 108: Perform identity detection using the classification model. The overall process of identity detection is as follows: Figure 7 shown.

[0177] The above technical solutions have a certain consistency: the accurate data segmentation method ensures that the data solved in the posture solution and trajectory restoration can correspond to the real movement, and can also avoid the influence of static and interference data in feature extraction, ensuring the reliability of data during identity detection. The posture solution and trajectory restoration methods can provide multiple new feature quantities for the identity detection model, and the extended Kalman filter model based on dynamic noise feedback proposed in the present invention can make the solved feature quantities more accurate. Based on a large number of optimization projects and feature engineering, the present invention can give relatively accurate detection results in identity detection with a very small sample data set (the experimenter only needs to collect 5 specific actions in advance as a training set).

[0178] Experimental testing process:

[0179] The present invention selects three experimenters, gives the experimenters a unified command "action 1", and then records the 6-axis inertial sensor data of the experimenters' actions in response to the command. The present invention uses the principle of inertial navigation to restore the trajectory of the collected original time series data, and then extracts features from the trajectory data and sensor data, extracts a total of 18 features including time domain and frequency domain information, obtains 162 input features, and sets the signal window length to 6 seconds. The present invention inputs the 162 extracted feature data into the random forest model for training and testing to distinguish different experimenters. The experiment was repeated 10 times, and the experimental results were averaged. The final results are shown in Table 1. The following is an explanation of several indicators in the experimental results.

[0180] Threshold crossing rate: The proportion of individuals who agree to be identified and can still be identified after the threshold is set.

[0181] Precision: The proportion of individuals correctly identified as this category among all individuals identified as this category, which reflects the accuracy of the model for this task.

[0182] Recall rate: The proportion of individuals correctly identified as belonging to this category among the individuals who actually belong to this category, which reflects the degree of completeness of the model in finding individuals of this category.

[0183] Table 1 Test results

[0184]

[0185] For different experimenters A, B, and C, when the lower threshold of 0.2 is set, the accuracy of the model detection is 86.0%, 94.2%, and 98.7%, respectively, which shows the effectiveness of the model. As the recognition threshold increases, the model becomes more demanding in identifying individuals, and the individual recognition accuracy increases, but the recall rate decreases. This corresponds to the "unlocking" scenario. For the "rejection of recognition" scenario, the impact of the threshold is the opposite, that is, as the recognition threshold increases, the model's misjudgment rate increases, but the recall rate of rejection of recognition increases. This corresponds to the "personalized recommendation" scenario. Therefore, this identity detection process not only has a high recognition accuracy, but can also complete tasks in different scenarios by adjusting the threshold parameters.

[0186] The present invention also provides an identity detection system based on an inertial sensor, comprising:

[0187] A data acquisition module, used to collect motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points;

[0188] An effective motion data extraction module, used to extract effective motion data from the motion data;

[0189] An attitude information solving module is used for solving the attitude solving model of the extended Kalman filter based on dynamic noise feedback, solving the attitude of the effective motion data to obtain attitude information;

[0190] A motion trajectory restoration module, used to restore the motion trajectory based on the posture information;

[0191] A feature extraction module, used to extract features from the effective motion data and the motion trajectory in the time domain and the frequency domain respectively, to obtain a plurality of motion feature vectors;

[0192] An interference motion data feature vector generation module, used to randomly generate multiple groups of interference motion data and extract feature vectors of the interference motion data;

[0193] A training module, used for training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model;

[0194] The identity detection module is used to perform identity detection through the classification model.

[0195] Wherein, the effective motion data extraction module specifically includes:

[0196] A motion state data determining unit, configured to determine the motion state data based on the motion data by using a multi-sliding window method;

[0197] An effective data segmentation point determination unit, configured to determine effective data segmentation points based on the motion state data by using a multi-sliding window method;

[0198] The valid motion data determining unit is used to determine the valid motion data based on the valid data segmentation points.

[0199] Wherein, the posture information solving module comprises:

[0200] an initial state information calculation unit, configured to calculate initial state information based on the valid motion data;

[0201] The attitude solution model determination unit is used to adjust the weight of the extended Kalman filter based on the initial state information to obtain an attitude solution model of the extended Kalman filter based on dynamic noise feedback.

[0202] Specifically, they include:

[0203] A first input unit, used for inputting the interference motion data feature vector and the motion feature vector into a first random forest model to obtain a first output result;

[0204] A judging unit, configured to judge whether the first output result is an action category corresponding to a feature vector of the interference motion data;

[0205] A first output unit, configured to output a rejection of recognition when the first output result is an action category corresponding to the feature vector of the interference motion data;

[0206] A second input unit is used for inputting the motion feature vector into a second random forest model to obtain the probability of each action category to which the motion feature vector belongs when the first output result is not the action category corresponding to the interference motion data feature vector;

[0207] The second output unit is used to output a rejection of recognition when the maximum probability is less than the probability threshold; and to use the action category corresponding to the maximum probability as the classification result when the maximum probability is greater than the probability threshold.

[0208] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0209] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An identity detection method based on an inertial sensor, characterized in that: include: Collecting motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points; Extracting effective motion data from the motion data; An attitude calculation model of an extended Kalman filter based on dynamic noise feedback is used to calculate the attitude of the effective motion data to obtain attitude information; Restoring the motion trajectory based on the posture information; Extracting features of the effective motion data and the motion trajectory in the time domain and the frequency domain respectively to obtain a plurality of motion feature vectors; Randomly generate multiple groups of interfering motion data, and extract feature vectors of the interfering motion data; Training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model; Identity detection is performed through the classification model.

2. The identity detection method based on inertial sensor according to claim 1, characterized in that: Extracting effective motion data from the motion data specifically includes: Based on the motion data, a multi-sliding window method is used to determine motion state data; Based on the motion state data, a multi-sliding window method is used to determine effective data segmentation points; Valid motion data is determined based on the valid data segmentation points.

3. The identity detection method based on inertial sensor according to claim 1, characterized in that: The determination process of the attitude solution model of the extended Kalman filter based on dynamic noise feedback is as follows: Calculate initial state information based on the valid motion data; The weight of the extended Kalman filter is adjusted based on the initial state information to obtain an attitude solution model of the extended Kalman filter based on dynamic noise feedback.

4. The identity detection method based on inertial sensor according to claim 1, characterized in that: The random forest model is trained based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector, specifically including: Inputting the interference motion data feature vector and the motion feature vector into a first random forest model to obtain a first output result; Determining whether the first output result is an action category corresponding to the feature vector of the interference motion data; If so, the output is rejected; If not, the motion feature vector is input into a second random forest model to obtain the probability of each action category to which the motion feature vector belongs; If the maximum probability is less than the probability threshold, the output is rejected; If the maximum probability is greater than the probability threshold, the action category corresponding to the maximum probability is taken as the classification result.

5. An identity detection system based on an inertial sensor, characterized in that: include: A data acquisition module, used to collect motion data through an inertial sensor; the motion data is time series signal data including a plurality of data points; An effective motion data extraction module, used to extract effective motion data from the motion data; An attitude information solving module is used for solving the attitude solving model of the extended Kalman filter based on dynamic noise feedback, solving the attitude of the effective motion data to obtain attitude information; A motion trajectory restoration module, used to restore the motion trajectory based on the posture information; A feature extraction module, used to extract features from the effective motion data and the motion trajectory in the time domain and the frequency domain respectively, to obtain a plurality of motion feature vectors; An interference motion data feature vector generation module, used to randomly generate multiple groups of interference motion data and extract feature vectors of the interference motion data; A training module, used for training a random forest model based on the interfering motion data feature vector, the motion feature vector, the action category corresponding to the interfering motion data feature vector, and the action category corresponding to the motion feature vector to obtain a classification model; The identity detection module is used to perform identity detection through the classification model.

6. The inertial sensor-based identity detection system according to claim 5, characterized in that: The effective motion data extraction module specifically includes: A motion state data determining unit, configured to determine the motion state data based on the motion data by using a multi-sliding window method; An effective data segmentation point determination unit, configured to determine effective data segmentation points based on the motion state data by using a multi-sliding window method; The valid motion data determining unit is used to determine the valid motion data based on the valid data segmentation points.

7. The inertial sensor-based identity detection system according to claim 5, characterized in that: The posture information solving module comprises: an initial state information calculation unit, configured to calculate initial state information based on the valid motion data; The attitude solution model determination unit is used to adjust the weight of the extended Kalman filter based on the initial state information to obtain an attitude solution model of the extended Kalman filter based on dynamic noise feedback.

8. The inertial sensor-based identity detection system according to claim 5, characterized in that: Specifically include: A first input unit, used for inputting the interference motion data feature vector and the motion feature vector into a first random forest model to obtain a first output result; A judging unit, configured to judge whether the first output result is an action category corresponding to a feature vector of the interference motion data; A first output unit, configured to output a rejection of recognition when the first output result is an action category corresponding to the feature vector of the interference motion data; A second input unit is used for inputting the motion feature vector into a second random forest model to obtain the probability of each action category to which the motion feature vector belongs when the first output result is not the action category corresponding to the interference motion data feature vector; The second output unit is used to output a rejection of recognition when the maximum probability is less than the probability threshold; and to use the action category corresponding to the maximum probability as the classification result when the maximum probability is greater than the probability threshold.

Images