Pedestrian inertial navigation method based on action recognition

Abstract

The invention belongs to the technical field of pedestrian navigation and positioning, and relates to a pedestrian inertial navigation method based on action recognition. The method comprises the following steps: collecting IMU data and muscle electricity data by using a wearable device; in a static state, calculating an Euler angle and solving a coordinate rotation matrix by means of IMU data, converting an acceleration in a carrier coordinate system into a navigation coordinate system, and obtaining speed information after integration; after the muscle electricity data is subjected to preprocessing and feature extraction, an SVM model based on statistics and an action correction algorithm is used for recognizing the motion state of the human body in a mixed motion mode, and an action recognition result is obtained; the action recognition result is used for self-adaptively adjusting threshold detection of the zero speed interval, finally, the zero speed is corrected in the zero speed interval through a speed correction method based on the EKF, integration is conducted on the corrected speed, and therefore position information is obtained. According to the method, high positioning precision, small positioning error and good robustness can be realized in a mixed motion mode, and an accurate track can be drawn.

Description

Technical Field

[0001] The invention belongs to the technical field of pedestrian navigation and positioning, and relates to a pedestrian inertial navigation method based on motion recognition. Background Art

[0002] With the development of technologies such as wireless communication and sensors, people have an increasing demand for navigation and positioning. Today, people are very familiar with positioning systems, which are ubiquitous in a wide range of applications in daily life. To some extent, people's familiarity with these systems is largely due to the Global Positioning System (GPS). Although this technology provides good positioning results in outdoor applications, the performance of GPS systems in indoor applications is not satisfactory due to the weak GPS signal that cannot penetrate the walls of buildings and the existence of multipath effects. It can be seen that there is great research significance in achieving high-precision indoor pedestrian navigation. In addition, its application scenarios are also very wide, such as providing indoor navigation services in large buildings, tracking the location of items in logistics systems, and monitoring the location of patients in hospitals. In particular, indoor positioning can be used for security monitoring and emergency rescue in places such as buildings and underground parking lots to improve safety and rescue efficiency. Therefore, high-reliability and high-precision pedestrian navigation and indoor positioning technology have extremely important research significance and application value.

[0003] At present, there is a difficulty in the field of pedestrian inertial navigation based on zero-speed correction: in the zero-speed correction algorithm, the fixed threshold makes it difficult to accurately detect the zero-speed interval in the mixed motion mode. In the case of running, jumping, etc., where the foot is in contact with the ground for a short time, the detection of the zero-speed interval is more challenging. In addition to being still, human motion can be roughly divided into: ordinary motion, competitive motion, and marching. Ordinary motion mainly includes: walking forward, running, backing up, stepping on the right, stepping on the left, turning right, turning left, going uphill / up steps, going downhill / down steps, standing, sitting down, lying down, etc.; competitive motion mainly includes: running, jumping, crawling, climbing, dodging, etc.; marching mainly includes: escalators, elevators, riding bicycles, taking cars, taking trains, etc. Most people participate in daily sports, but soldiers, firefighters, and other personnel who perform special tasks have motion patterns including crawling, stepping, climbing, dodging, and other methods. Therefore, it is necessary for us to conduct more precise perception research on human motion patterns and combine them with navigation and positioning technology to improve the accuracy and efficiency of pedestrian navigation and positioning.

[0004] In recent years, with the rapid development of technologies such as system integration, microsensors, the Internet of Things, 5G, and artificial intelligence, the research and application of wearable devices and related fields have become increasingly extensive. Wearable devices are compact and lightweight in design, and have the ability to continuously sense, collect, and upload information and physiological data. They have been widely used in many fields such as virtual reality, medical treatment, health testing, and sports training. With the development of micro-electromechanical technology, the integration of inertial measurement units (IMUs) has gradually increased, and the reduction in size and weight has made it easier for the human body to wear them. Scholars use kinematic information collected based on IMUs to achieve motion recognition and parameter prediction. In addition to focusing on whole-body motion capture, they also focus on local human motion capture and lower limb motion parameter estimation. Surface electromyography (sEMG) signals are a weak bioelectric current that is highly real-time and less affected by the environment. The signals collected by sEMG sensors provide a direct and non-invasive data source for motion recognition and motion parameter estimation. After processing, data reflecting key information such as muscle contraction status, strength, and timing can be extracted.

[0005] In the lower limb motion recognition task, the correlation between the topological characteristics of the equipment and the task is considered, and the feature extraction and screening methods are combined. To address the problem of low recognition rate of the traditional support vector machine (SVM) model, the SVM model based on statistical analysis and motion correction algorithm is used to identify the motion state of the lower limbs in the mixed motion mode. It is then applied to the pedestrian inertial navigation method. The detection threshold of the zero-speed interval is adaptively adjusted according to the recognition results, thereby reducing the missed detection and false detection rates in the zero-speed interval, thereby improving the positioning accuracy.

[0006] Therefore, in response to the above problems, this application combines motion pattern recognition with pedestrian navigation on the basis of zero-speed correction to further improve the performance of pedestrian inertial navigation method in mixed motion mode. Specifically, the SVM algorithm is proposed to identify motion patterns, which is a machine learning method for classification problems, aiming to achieve linear classification by defining the maximum interval in the feature space. Summary of the invention

[0007] The purpose of the present invention is to solve the problem that zero-speed interval detection based on a fixed threshold in a hybrid motion mode has interval omission and false detection, which leads to reduced positioning accuracy. A pedestrian inertial navigation method based on motion recognition is proposed, which is adapted to different pedestrian motion modes and relies on wearable devices to collect IMU data and muscle electrical data; in a stationary state, the Euler angle is calculated based on the IMU data and the coordinate rotation matrix is ​​obtained, the acceleration in the carrier coordinate system is converted to the navigation coordinate system, and the speed information is obtained after integration; the muscle electrical data is preprocessed and feature extracted (the extracted features include time domain features: mean absolute value (MAV), root mean square (RMS), variance (VARE), integrated electromyogram (IEMG), wavelength (WL) and Willison amplitude (Willison The invention discloses a method for detecting a human body in a mixed motion mode by using a support vector machine (SVM) model based on statistics and motion correction algorithm after obtaining a human body's motion state in a mixed motion mode, and obtaining a motion recognition result; the motion recognition result is used to adaptively adjust the threshold detection of a zero-speed interval, and finally a speed correction method based on an extended Kalman filter (EKF) is used in the zero-speed interval to correct the zero speed and integrate the corrected speed, so as to obtain position information; the method can achieve higher positioning accuracy, smaller positioning error, better robustness and more accurate trajectory drawing in a mixed motion mode.

[0008] In order to achieve the above object, the present invention adopts the following technical scheme:

[0009] A pedestrian inertial navigation method based on motion recognition comprises the following steps:

[0010] S1, collect sEMG signals and IMU data through wearable devices;

[0011] S2, preprocessing the sEMG signal to obtain preprocessed sEMG data;

[0012] At the same time, in a stationary state, the IMU data is used to calculate the Euler angle and obtain the coordinate rotation matrix, the acceleration in the carrier coordinate system is converted to the navigation coordinate system, and the velocity information is obtained after integration;

[0013] S3, extracting the time domain features, time-frequency features and time-advanced features of the pre-processed sEMG data and combining them to obtain a feature data vector;

[0014] S4, repeat S1 to S3 multiple times, collect and obtain multiple sets of feature data vectors of different subjects in different motion states, and then divide these feature data vectors into training sets and test sets in proportion;

[0015] S5, feature screening is performed on the feature data of the training set and the data set is updated;

[0016] S6. Construct and train an SVM model based on statistics and motion correction, specifically: traverse the training set to run the SVM model to identify different motion patterns, traverse and obtain the Gaussian kernel bandwidth and penalty coefficient corresponding to the motion pattern with the highest recognition accuracy; statistically analyze the recognition results of the SVM model within a certain time interval, select the motion state recognition result with the mode being the time interval to update the SVM model, and obtain a statistical analysis SVM model; further correct the motion pattern of the previous window through the current estimated motion pattern and update the hyperparameters of the statistical analysis SVM model based on this recognition result, and obtain an SVM model based on statistics and motion correction;

[0017] S7, adjusting the zero-speed detection threshold of the corresponding motion mode according to the SVM model based on statistics and motion correction of the test set and the motion mode recognition result;

[0018] S8, perform zero speed detection based on the zero speed detection threshold of the motion mode and the speed information obtained in S2. If the speed is detected to be zero speed, perform zero speed correction, and then integrate the corrected speed to obtain position information. If the speed is detected to be non-zero speed, do not perform zero speed correction, and directly integrate the speed information obtained in S2 to obtain position information.

[0019] S1 The wearable device includes electrodes and a sensor group; the sensor group includes an accelerometer, a gyroscope and a magnetometer; the IMU data is obtained through human motion information collected by the accelerometer, the gyroscope and the magnetometer; the human motion information includes but is not limited to acceleration information, velocity information and coordinate information.

[0020] The velocity information is obtained by calculating the Euler angle and obtaining the coordinate rotation matrix in a stationary state, converting the acceleration in the carrier coordinate system to the navigation coordinate system, and then integrating it; the sEMG signal is multi-channel surface electromyography information of lower limb muscles collected based on electrodes in the wearable device; the number of electrodes is greater than 2 and less than 30.

[0021] S2, the preprocessing of sEMG signal specifically comprises the following sub-steps:

[0022] S21, filtering out high-frequency noise of the sEMG signal by low-pass filtering to obtain a low-pass filtered sEMG signal;

[0023] S22, performing a first-stage amplification on the sEMG signal after low-pass filtering to obtain a first-stage amplified sEMG signal;

[0024] S23, performing high-pass filtering on the first-stage amplified sEMG signal to obtain a high-pass filtered sEMG signal;

[0025] S24, performing secondary amplification on the sEMG signal after high-pass filtering to obtain a secondary amplified sEMG signal;

[0026] S25, converting the sEMG signal after secondary amplification into pre-processed sEMG data via an analog-to-digital converter.

[0027] The cutoff frequency of the low-pass filter described in S21 is greater than 10Hz and less than 200Hz; the amplification factor of the first-stage amplification described in S22 is greater than or equal to 5 and less than or equal to 100 when implemented; the cutoff frequency of the high-pass filter described in S23 is greater than or equal to 100Hz and less than or equal to 200Hz, and the high-pass filter can remove low-frequency noise mainly composed of drift and motion artifacts; the amplification factor of the second-stage amplification described in S24 is greater than or equal to 10 and less than or equal to 200 when implemented.

[0028] The combination of S3 includes but is not limited to addition, subtraction, series connection and parallel connection; the time domain features, time-frequency features and time-advance features are used to train the SVM model based on statistics and motion correction; the time domain features include but are not limited to MAV, RMS, VARE, IEMG, WL and WA; the time-frequency features include but are not limited to FFT, STFT and wavelet decomposition; the wavelet basis used in the wavelet decomposition includes Haar wavelet basis, dbN wavelet basis, symN or Bior wavelet basis; the number of layers of the wavelet decomposition is greater than or equal to 3 layers and less than or equal to 10 layers in specific implementation; the time-advance feature is extracted based on the advance time window; the value range of the advance time window is greater than or equal to 10ms and less than or equal to 300ms.

[0029] The specific extraction process of the time-advanced feature is: 1) determining the advanced time window; 2) extending the time length of an advanced time window from the current time point forward, and adding a fixed time window; 3) extracting the features of the total time length data within the advanced time window plus the fixed time window; 4) taking the time step as the interval, obtaining the features of the total time length data within the advanced time window plus the fixed time window in sequence.

[0030] The fixed time window has a value range of greater than or equal to 180 ms and less than or equal to 220 ms; the time step has a value range of greater than or equal to 10 ms and less than or equal to 50 ms.

[0031] The ratio of the training set and the test set in S4 ranges from 6:4 to 9:1; the different motion states include but are not limited to going up stairs, going down stairs, squatting, standing up, standing and walking at a constant speed.

[0032] The value range of the certain time interval in S6 is: 100ms to 800ms;

[0033] S7 is specifically as follows: S71, inputting the test set into the SVM model based on statistics and action correction for pattern recognition;

[0034] S72, drawing a T curve and selecting zero-speed detection thresholds for various motion modes based on the curve;

[0035] S73, identifying various motion modes and adjusting the zero-speed detection threshold of the corresponding motion mode according to the motion mode identification result.

[0036] Beneficial Effects

[0037] The pedestrian inertial navigation method based on motion recognition proposed in the present invention has the following beneficial effects compared with the existing method based on fixed threshold:

[0038] 1. In view of the fact that the application of complex neural network models in pedestrian inertial navigation increases the algorithm complexity and the low accuracy of the traditional SVM method in identifying motion states, the method proposes to combine statistical analysis and motion correction algorithm to optimize the action classification results of the SVM model under mixed motion states, improve the accuracy of action recognition in mixed motion modes, and provide conditions for the application of pedestrian inertial navigation methods in complex motion scenes;

[0039] 2. The method selects the zero-speed detection threshold for different motion states according to the T-curve characteristics of different motion states, and sets the corresponding zero-speed detection threshold according to the motion recognition result, thereby improving the detection rate of the zero-speed interval. After correcting the speed in the zero-speed interval detected by the adaptive threshold, a more accurate motion trajectory is obtained, reducing the positioning error. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flowchart of a pedestrian inertial navigation method based on motion recognition according to the present invention;

[0041] Figure 2 A schematic diagram of a wearable device based on a pedestrian inertial navigation method based on motion recognition of the present invention;

[0042] Figure 3 Schematic diagram of wavelet decomposition;

[0043] Figure 4 is the ranking of feature weights;

[0044] Figure 5It is a schematic diagram of the motion state correction method;

[0045] Figure 6 It is the zero-speed interval discrimination result of the fixed threshold in the prior art;

[0046] Figure 7 It is the zero speed interval discrimination result based on the adaptive threshold value of the present invention;

[0047] Figure 8 It is the zero speed correction result based on the fixed threshold in the prior art;

[0048] Fig. 9 It is the zero speed correction result based on the adaptive threshold value of the present invention;

[0049] Fig.10 are the motion trajectories and true values ​​of the fixed and adaptive thresholds. DETAILED DESCRIPTION

[0050] The following is a further explanation and detailed description of a pedestrian inertial navigation method based on motion recognition of the present invention in conjunction with the accompanying drawings and embodiments.

[0051] Example 1

[0052] The present invention proposes a pedestrian inertial navigation method based on motion recognition, the process is as follows: Figure 1 As shown, during the specific implementation: first, wear the wearable device to the foot and thigh to collect IMU data and muscle electrical data, such as Figure 2 As shown; secondly, in a stationary state, the IMU data is used to calculate the Euler angle and obtain the coordinate rotation matrix, and the acceleration in the carrier coordinate system is converted to the navigation coordinate system, and the speed information is obtained after integration. After preprocessing and feature extraction of the muscle electrical data, the SVM model based on statistics and motion correction algorithm is used to identify the motion state of the human body in the mixed motion mode to obtain the motion recognition result. Then, the motion recognition result is used to adaptively adjust the detection threshold of the zero-speed interval. Finally, the speed correction method based on EKF is used in the zero-speed interval to perform zero-speed correction, and the corrected speed is integrated to obtain the position information. Finally, the navigation system has high positioning accuracy, small positioning error, good robustness and draws more accurate trajectories in the mixed motion mode.

[0053] The method wears the wearable device to the thigh to collect sEMG data and IMU data, which is specifically embodied in S1 of the invention content. The collected sEMG data and IMU data are pre-processed, and after calculating the Euler angle and obtaining the coordinate rotation matrix in a stationary state, the acceleration in the carrier coordinate system is converted to the navigation coordinate system, and the acceleration is integrated to obtain the speed information, which is specifically embodied in S2 of the invention content. Feature extraction is performed on the pre-processed sEMG data and IMU data, which is specifically embodied in S3 of the invention content. Multiple groups of feature data vectors under different motion states of different subjects are obtained, and then the feature data vectors are divided into training sets and test sets in proportion, which is specifically embodied in S4 of the invention content. The extracted features are subjected to feature screening by calculating feature weights using the NCFS algorithm. , specifically embodied in S5 in the invention content, constructing a statistical analysis algorithm and a method for motion correction based on adjacent window gait constraints, and integrating them with the SVM machine learning model, thereby constructing an SVM model based on statistics and motion correction algorithms to identify the motion state of the human body in the mixed motion mode, specifically embodied in S6 in the invention content, and then constructing an adaptive threshold adjustment algorithm, using the motion recognition result to adaptively adjust the detection threshold of the zero speed interval, specifically embodied in S7 in the invention content, and then performing zero speed detection to determine whether it is in the zero speed interval, correcting the zero speed in the zero speed interval, and integrating the corrected speed to obtain position information, and directly integrating the non-zero speed that is not in the zero speed interval to obtain position information, specifically embodied in S8 in the invention content, and finally realizing pedestrian inertial navigation. The present invention realizes that the navigation system has high positioning accuracy, small positioning error, good robustness, and draws a more accurate trajectory in the mixed motion mode.

[0054] When the method of the present invention is implemented, a closed rectangular path test is performed, and the path for collecting data based on the wearable device is a closed rectangle with a width of 26m and a length of 18m. The four corners of the rectangular path are in a standing state, the wide side is in a running state, and the long side is in a walking state. The test experimental data collection time is about 70s.

[0055] The specific implementation includes the following steps:

[0056] S1, collect sEMG signals and IMU data through wearable devices;

[0057] The wearable device includes electrodes and a sensor group; the sensor group includes an accelerometer, a gyroscope and a magnetometer; the wearable device is specifically implemented as a leg ring and an IMU, and data is collected by wearing the leg ring on the thigh and the IMU on the shoelaces of the foot. The initial position of the leg ring needs to be aligned to ensure that the center of the first sEMG electrode of the leg ring is aligned with the center of the knee, and then the leg ring is worn to the position with the richest muscles on the thigh. The specific wearing schematic diagram is shown in Figures (2a) and (2b);

[0058] The sEMG signal is multi-channel surface electromyography information of lower limb muscles collected based on electrodes in the wearable device;

[0059] The IMU data is obtained by collecting human motion information from an accelerometer, a gyroscope and a magnetometer; the human motion information is specifically implemented as acceleration information, velocity information and coordinate information;

[0060] The velocity information is obtained by calculating the Euler angle in a stationary state and obtaining a coordinate rotation matrix, converting the acceleration in the carrier coordinate system to the navigation coordinate system, and then integrating the calculated velocity information;

[0061] The number of electrodes is 16 in the specific implementation. For subjects with a height range of 150 cm to 190 cm and a weight range of 100 to 200 catties, a leg ring with 16 channels or more is used, and the specific wearing diagram is shown in Figure (2c); for subjects with a weight of less than 100 catties and a height of less than 150 cm, a leg ring with 12 channels or less is used, and the specific wearing diagram is shown in Figure (2d);

[0062] S2, preprocessing the sEMG signal to obtain preprocessed sEMG data, specifically including the following sub-steps;

[0063] S21, filtering out high-frequency noise of the sEMG signal by low-pass filtering to obtain a low-pass filtered sEMG signal;

[0064] The cut-off frequency of the low-pass filter is 10 Hz during implementation;

[0065] S22, performing a first-stage amplification on the sEMG signal after low-pass filtering to obtain a first-stage amplified sEMG signal;

[0066] The magnification of the first-level amplification is 10 in the specific implementation;

[0067] S23, performing high-pass filtering on the first-stage amplified sEMG signal to obtain a high-pass filtered sEMG signal;

[0068] The cut-off frequency of the high-pass filter is 20 Hz when implemented;

[0069] S24, performing secondary amplification on the sEMG signal after high-pass filtering to obtain a secondary amplified sEMG signal;

[0070] The magnification of the secondary amplification is 20 in the specific implementation;

[0071] S25, converting the secondary amplified sEMG signal into pre-processed sEMG data via an analog-to-digital converter;

[0072] S3, extracting the time domain features, time-frequency features and time-advanced features of the pre-processed sEMG data and combining them to obtain a feature data vector;

[0073] The combination is specifically implemented in series;

[0074] The time domain features, time-frequency features and time-advance features are used to train a SVM model based on statistics and motion correction;

[0075] The time domain features are specifically implemented as MAV, RMS, VARE, IEMG, WL and WA;

[0076] The MAV, through Calculate, where x i represents the i-th sEMG signal sample point, K is the total length of the signal in the time window; the RMS is obtained by Calculate; the VARE, through Calculate, where represents the average value of K sEMG signal sample points; the IEMG, through Calculate; the WL, by Calculate, where x i+1 represents the i+1th sEMG signal sample point; the WA is Calculate, where th is the threshold;

[0077] The time-frequency features include but are not limited to FFT, STFT and wavelet decomposition; in specific implementation, wavelet decomposition is performed on MAV, RMS and WL features, and the wavelet basis used in the wavelet decomposition is db7 wavelet basis in specific implementation; the number of layers of the wavelet decomposition is 3 in specific implementation; the specific decomposition diagram is as follows Figure 3 As shown. Among them, cA 1 ,cA 2 and cA 3 Respectively represent the approximate components of each layer of wavelet decomposition, cD 1 、cD 2 and cD 3 They respectively represent the detail components of each layer after decomposition. The present invention extracts the three detail components obtained by decomposition and the approximate component of the last layer.

[0078] The time advance feature is extracted based on the advance time window; the advance time window is 200ms during specific implementation;

[0079] The specific extraction process of the time-advanced feature is: 1) determining the advanced time window; 2) extending the time length of an advanced time window from the current time point forward, and adding a fixed time window; 3) extracting the features of the total time length data within the advanced time window plus the fixed time window; 4) taking the time step as the interval, obtaining the features of the total time length data within the advanced time window plus the fixed time window in sequence.

[0080] The fixed time window is 200ms when implemented; the time step is 20ms when implemented;

[0081] When the present invention is implemented, all the features extracted are shown in Table 1, wherein the features prefixed with app are the features of the last layer of approximate components extracted by wavelet decomposition, the features prefixed with cd1, cd2 and cd3 are the features of the first, second and third level detail components extracted by wavelet decomposition, respectively, and the features prefixed with ad represent the features extracted based on the time advance window;

[0082] Table 1 All features extracted in this example

[0083]

[0084]

[0085] S4, repeat S1 to S3 multiple times, collect and obtain multiple sets of feature data vectors of different subjects in different motion states, and then divide these feature data vectors into training sets and test sets in proportion;

[0086] The training set and the test set are divided in a ratio of 6:4 when implemented;

[0087] The different motion states include going upstairs, going downstairs, squatting, standing up, standing and walking at a constant speed;

[0088] When S4 was implemented, in the experiments of going up and down stairs, the duration of data collection was 10 seconds. The subjects first stood for 2 seconds and then went up or down stairs in the next 8 seconds. For squatting and standing up, the data collection time was 8 seconds. The subjects stood for the first 2 seconds, squatted for the next 2-4 seconds, squatted for 4-6 seconds, and finally stood up within 6-8 seconds. The data collection of walking lasted for 55 seconds. The subjects stood for the first 2 seconds and walked at a constant speed for the next 53 seconds.

[0089] S5, feature screening is performed on the feature data of the training set and the data set is updated;

[0090] The feature screening uses the NCFS algorithm to find a weighted vector that can maximize the similarity between samples of the same type and minimize the similarity between samples of different types; specifically:

[0091] S51. Calculate two samples x i With x j The weighted distance of

[0092] The two samples x i With x j The weighted distance of w (x i ,x j ),pass Calculate, where w l is the weight of the l-th dimension feature, x il Represents x i The lth dimension feature, x jl Represents x j The lth dimension feature, d is the feature vector x i Dimensions;

[0093] S52. Query sample x i Select a sample x in the space j As the probability of the reference point; the probability reflects the relative position relationship of the sample in the feature space;

[0094] The sample x i Select a sample x in the space j As the probability of the reference point, denoted as p ij ,pass Calculate, where k(z)=exp(-z / σ) is the kernel function and σ is the kernel width;

[0095] The reference points refer to other samples used to calculate the distance between samples;

[0096] S53, calculate sample x i The probability of correct classification;

[0097] The sample x i The probability of correct classification is denoted by p i ,pass Calculate, where y i is the sample label, y i ∈{1,2,3,···,C};

[0098] S54, calculating the average correct classification probability of all samples, that is, the classification accuracy;

[0099] The accuracy of the classification, denoted as ε(w), is obtained by Calculate, where N is the number of samples in the training set;

[0100] S55. In order to alleviate overfitting, a regularization term is introduced to obtain the objective function;

[0101] The objective function, denoted as ε(w), is given by Calculate, where λ>0 is the regularization parameter, which is used to control the strength of regularization and can be adjusted by cross-validation;

[0102] S56. Derivative the above objective function ε(w) and update the feature weight w l ;

[0103] The result after derivation is recorded as pass:

[0104]

[0105] Calculate and get the gradient update equation of NCFS algorithm;

[0106] The updated feature weight w l ,pass Calculate, where η is the learning rate;

[0107] S57, calculating feature weights according to the NCFS algorithm;

[0108] The feature weight in the NCFS algorithm is calculated based on the training set. The feature weight value is denoted as w f ,pass Calculate, where K is the number of IMU and sEMG channels, w fi and w adfi are the weights of the features of each channel and the corresponding time-advanced features respectively;

[0109] The present invention uses the NCFS algorithm to obtain the weight values ​​of the features VARE, IEMG, RMS, WL, MAV, WA, DB7-MAV, DB7-RMS and DB7-WL, such as Figure 4 shown. Figure 4 Sort the weight values ​​of each feature from large to small. The weight values ​​are WL, RMS, cd from high to low. 3 -MAV, VARE, IEMG, MAV, app-WL, cd 1 -MAV, cd 1 -RMS, cd 1 -WL,app-RMS,cd 3 -RMS, cd 3 -WL, WA, cd 2 -MAV, cd 2 -RMS, app-MAV, cd 2-WL;

[0110] S58, taking the top five weighted features and their time-advanced features as the retained feature data to update the training set, validation set and test set; specifically in this embodiment, all retained features are shown in Table 2:

[0111] Table 2 Summary of features retained after feature screening in this example

[0112]

[0113] S6. Construct and train a SVM model based on statistics and action correction, specifically:

[0114] S61, traversing the training set to run the SVM model to identify different action modes, traversing and obtaining the Gaussian kernel bandwidth and penalty coefficient corresponding to the action mode with the highest recognition accuracy;

[0115] S62, statistically analyzing the recognition results of the SVM model within a certain time interval, and selecting the mode as the motion state recognition result of the time interval; the time interval is 0.5s in specific implementation;

[0116] S63, further correct the motion mode of the previous window by the current estimated motion mode and update the hyperparameters of the SVM model based on the recognition result, and obtain the SVM model based on statistics and motion correction, specifically: assuming that the motion types estimated by four consecutive adjacent windows are A, B, C and D respectively, when A=C and B≠C, it means that the motion category B is misjudged, then correct B to the motion category of A and C. When A=B and B≠C and C≠D, it means that the motion category C is misjudged, then correct C to the motion category of A and B, such as Figure 5 shown.

[0117] S7, adjusting the zero-speed detection threshold of the corresponding motion mode according to the SVM model based on statistics and motion correction of the test set and the motion mode recognition result, specifically including the following sub-steps:

[0118] S71, inputting the test set into the SVM model based on statistics and action correction to perform pattern recognition;

[0119] S72, drawing a T curve and selecting zero-speed detection thresholds for various motion modes based on the curve;

[0120] The T curve, through Calculate, where ω i is the three-axis angular velocity value measured by the gyroscope, W is the time window, ||·|| is the second norm, a i is the three-axis acceleration value measured by the accelerometer, is the mean of the three-axis acceleration in the window, g is the gravitational acceleration, and are the standard deviations of the acceleration and angular velocity measurement errors, respectively;

[0121] The zero speed thresholds corresponding to the different motion modes are shown in Table 3;

[0122] Table 3 Zero speed detection thresholds for different motion states

[0123]

[0124] S73, identifying various motion modes and adjusting the zero speed detection threshold of the corresponding motion mode according to the motion mode identification result;

[0125] S8, performing zero speed detection according to the zero speed detection threshold of the motion mode, if the speed is detected to be zero speed, performing zero speed correction, and jumping to S9, if the speed is detected to be non-zero speed, no zero speed correction is performed, the speed after zero speed correction is integrated to obtain position information, and jumping to S10;

[0126] The zero speed detection specifically calculates the T value and determines the size of the T value and the zero speed interval threshold γ. If T < γ, it is judged as zero speed, and if T ≥ γ, it is judged as non-zero speed;

[0127] S9, perform zero speed correction on the speed in the zero speed interval, specifically:

[0128] S91, calculating the a priori estimated state quantity at the current moment;

[0129] The a priori estimated state quantity at the current moment is recorded as X(k|k-1), and is calculated by X(k|k-1)=f[X(k-1|k-1)], where X(k-1|k-1) is the optimal estimate at the k-1 moment, also called the posterior estimate;

[0130] S92, calculating the covariance matrix of the prior estimate;

[0131] The covariance matrix of the prior estimate is denoted as P(k|k-1), and is given by P(k|k-1)=F·P(k-1|k-1)·F T +Q calculation, where P(k-1|k-1) is the covariance matrix of X(k-1|k-1), Q is the covariance matrix of the process noise W(k), and F is the Jacobian matrix of the state transfer function f[x] at X(k|k-1);

[0132] S93, calculating the Kalman gain, which is used to update the state by combining the prior estimate and the actual observation value;

[0133] The Kalman gain, denoted as K(k), is obtained by Calculate, where R is the covariance matrix of the observation noise V(k), and H represents the Jacobian matrix of the observation function h[x] at X(k|k-1);

[0134] S94, calculating the optimal estimated state quantity at the current moment;

[0135] The optimal estimated state quantity uses the Kalman gain to correct the prior state estimate to the optimal state estimate, denoted as X(k|k), and is calculated by X(k|k)=X(k|k-1)+K(k)·[Z(k)-h{X(k|k-1)}];

[0136] S95, calculating the covariance matrix at the current moment to reflect the uncertainty of the latest state estimation;

[0137] The current covariance matrix, denoted as P(k|k), is calculated by P(k|k)=[IK(k)H]·P(k|k-1);

[0138] S10, integrating the speed after zero speed correction to obtain position information;

[0139] The calculated position information, that is, the calculated position coordinates in the navigation coordinate system, is denoted as P n ,pass calculate;

[0140] In the hybrid motion mode, the fixed threshold and the adaptive threshold are used to detect the zero speed interval. The detection results of the zero speed interval are as follows: Figure 6 and 7 As shown. Among them, 1 and 0 represent the zero speed moment and non-zero speed moment respectively. Figure 6 It can be seen that in mixed sports, there are 28 and 30 zero-speed intervals in running and walking respectively. There are more zero-speed intervals missed in the running state, 13 zero-speed intervals missed in the running state, and the detection rate of zero-speed intervals is 77.6%. Figure 7 It can be seen that the method of adaptively adjusting the zero-speed interval detection threshold based on the motion state can greatly increase the detection rate of the zero-speed interval, and the number of missed zero-speed intervals is 1. The detection rate of the zero-speed interval is 98.3%.

[0141] In hybrid motion mode, the results of zero speed correction based on fixed threshold and adaptive threshold are as follows: Figure 8 and 9 As shown. Figure 8 It can be seen that when there is a missed detection of the zero-speed interval, zero-speed correction cannot be performed in the missed detection interval, which leads to a linear error in the speed. The passage of time causes the error to gradually accumulate, thereby reducing the accuracy of pedestrian inertial navigation. Fig. 9It shows that when an adaptive threshold is used to detect the zero-speed interval, the detection rate of the zero-speed interval is higher, the linear error of the speed is reduced, and the speed correction effect is better than that based on the fixed threshold method, which improves the overall performance of the inertial navigation algorithm.

[0142] The motion trajectory drawn based on the fixed threshold method, the motion trajectory drawn based on the adaptive threshold method, and the true value of the trajectory are as follows Fig.10 As shown in the figure, the true value of the trajectory is the black solid line, the trajectory of the fixed threshold method is the purple dotted line, and the trajectory of the adaptive threshold method is the blue dotted line. As can be seen from the figure, after correcting the speed in the zero-speed interval detected by the adaptive threshold, the obtained trajectory can more accurately reflect the actual motion trajectory than the trajectory obtained by using a fixed threshold to detect the zero-speed interval and perform zero-speed correction.

[0143] The positioning accuracy based on fixed threshold and adaptive threshold is shown in Table 4. The relative position error of the trajectory obtained by the adaptive threshold method is 1.03m, which is 64.97% less than the positioning error based on the fixed threshold method, and the relative position error is reduced by 1.17%.

[0144] Table 4 Comparison of positioning accuracy using fixed threshold and adaptive threshold

[0145]

[0146] The above results show that compared with other methods, the pedestrian inertial navigation method based on motion recognition can effectively solve the problems of missed detection, false detection and low positioning accuracy in zero-speed interval detection based on fixed thresholds in mixed motion mode. In terms of speed correction effect and positioning accuracy, its performance is significantly superior to that of the ZUPT algorithm based on no zero-speed correction and fixed threshold, with improved zero-speed interval detection rate and reduced position error.

[0147] So far, through S1 to S10, a pedestrian inertial navigation method based on motion recognition is completed in the application of this example, and it has extremely high practicality and obvious advantages.

[0148] The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A pedestrian inertial navigation method based on motion recognition, characterized in that: The steps include: S1, collect sEMG signals and IMU data through wearable devices; S2, preprocessing the sEMG signal to obtain preprocessed sEMG data; At the same time, in a stationary state, the IMU data is used to calculate the Euler angle and obtain the coordinate rotation matrix, the acceleration in the carrier coordinate system is converted to the navigation coordinate system, and the velocity information is obtained after integration; S3, extracting the time domain features, time-frequency features and time-advanced features of the pre-processed sEMG data and combining them to obtain a feature data vector; S4, repeat S1 to S3 multiple times, collect and obtain multiple sets of feature data vectors of different subjects in different motion states, and then divide these feature data vectors into training sets and test sets in proportion; S5, feature screening is performed on the feature data of the training set and the data set is updated; S6. Construct and train an SVM model based on statistics and motion correction, specifically: traverse the training set to run the SVM model to identify different motion patterns, traverse and obtain the Gaussian kernel bandwidth and penalty coefficient corresponding to the motion pattern with the highest recognition accuracy; statistically analyze the recognition results of the SVM model within a certain time interval, select the motion state recognition result with the mode being the time interval to update the SVM model, and obtain a statistical analysis SVM model; further correct the motion pattern of the previous window through the current estimated motion pattern and update the hyperparameters of the statistical analysis SVM model based on this recognition result, and obtain an SVM model based on statistics and motion correction; S7, adjusting the zero-speed detection threshold of the corresponding motion mode according to the SVM model based on statistics and motion correction of the test set and the motion mode recognition result; S8, performing zero speed detection according to the zero speed detection threshold of the motion mode and the speed information obtained in S2, and if the speed is detected to be zero speed, performing zero speed correction, and then integrating the corrected speed to obtain position information; If the detected speed is non-zero, no zero-speed correction is performed, and the speed information obtained by S2 is directly integrated to obtain the position information.

2. The pedestrian inertial navigation method based on motion recognition according to claim 1, characterized in that: S1 The wearable device includes electrodes and a sensor group; the sensor group includes an accelerometer, a gyroscope and a magnetometer; the IMU data is obtained through human motion information collected by the accelerometer, the gyroscope and the magnetometer; the human motion information includes but is not limited to acceleration information, velocity information and coordinate information.

3. The pedestrian inertial navigation method based on motion recognition according to claim 2 is characterized in that: The velocity information is obtained by calculating the Euler angle and obtaining the coordinate rotation matrix in a stationary state, converting the acceleration in the carrier coordinate system to the navigation coordinate system, and then integrating it; the sEMG signal is multi-channel surface electromyography information of lower limb muscles collected based on electrodes in the wearable device; the number of electrodes is greater than 2 and less than 30.

4. The pedestrian inertial navigation method based on motion recognition according to claim 1, characterized in that: S2, the preprocessing of sEMG signal specifically comprises the following sub-steps: S21, filtering out high-frequency noise of the sEMG signal by low-pass filtering to obtain a low-pass filtered sEMG signal; S22, performing a first-stage amplification on the sEMG signal after low-pass filtering to obtain a first-stage amplified sEMG signal; S23, performing high-pass filtering on the first-stage amplified sEMG signal to obtain a high-pass filtered sEMG signal; S24, performing secondary amplification on the sEMG signal after high-pass filtering to obtain a secondary amplified sEMG signal; S25, converting the sEMG signal after secondary amplification into pre-processed sEMG data via an analog-to-digital converter.

5. The pedestrian inertial navigation method based on motion recognition according to claim 4 is characterized in that: The cutoff frequency of the low-pass filter described in S21 is greater than 10Hz and less than 200Hz; the amplification factor of the first-stage amplification described in S22 is greater than or equal to 5 and less than or equal to 100 when implemented; the cutoff frequency of the high-pass filter described in S23 is greater than or equal to 100Hz and less than or equal to 200Hz, and the high-pass filter can remove low-frequency noise mainly composed of drift and motion artifacts; the amplification factor of the second-stage amplification described in S24 is greater than or equal to 10 and less than or equal to 200 when implemented.

6. The pedestrian inertial navigation method based on motion recognition according to claim 1, characterized in that: The combination of S3 includes but is not limited to addition, subtraction, series connection and parallel connection; the time domain features, time-frequency features and time-advance features are used to train the SVM model based on statistics and motion correction; the time domain features include but are not limited to MAV, RMS, VARE, IEMG, WL and WA; the time-frequency features include but are not limited to FFT, STFT and wavelet decomposition; the wavelet basis used in the wavelet decomposition includes Haar wavelet basis, dbN wavelet basis, symN or Bior wavelet basis; the number of layers of the wavelet decomposition is greater than or equal to 3 layers and less than or equal to 10 layers in specific implementation; the time-advance feature is extracted based on the advance time window; the value range of the advance time window is greater than or equal to 10ms and less than or equal to 300ms.

7. The pedestrian inertial navigation method based on motion recognition according to claim 6, characterized in that: The specific extraction process of the time-advanced feature is: 1) determining the advanced time window; 2) extending the time length of an advanced time window from the current time point forward, and adding a fixed time window; 3) extracting the features of the total time length data within the advanced time window plus the fixed time window; 4) taking the time step as the interval, obtaining the features of the total time length data within the advanced time window plus the fixed time window in sequence.

8. The pedestrian inertial navigation method based on motion recognition according to claim 7, characterized in that: The fixed time window has a value range of greater than or equal to 180 ms and less than or equal to 220 ms; the time step has a value range of greater than or equal to 10 ms and less than or equal to 50 ms.

9. The pedestrian inertial navigation method based on motion recognition according to claim 1, characterized in that: The ratio of the training set and the test set in S4 ranges from 6:4 to 9:1; the different motion states include but are not limited to going up stairs, going down stairs, squatting, standing up, standing and walking at a constant speed.

10. The pedestrian inertial navigation method based on motion recognition according to claim 1, characterized in that: S7 is specifically as follows: S71, inputting the test set into the SVM model based on statistics and action correction for pattern recognition; S72, drawing a T curve and selecting zero-speed detection thresholds for various motion modes based on the curve; S73, identifying various motion modes and adjusting the zero-speed detection threshold of the corresponding motion mode according to the motion mode identification result.

Images