CN113239803A

CN113239803A - Dead reckoning positioning method based on pedestrian motion state recognition

Info

Publication number: CN113239803A
Application number: CN202110521228.5A
Authority: CN
Inventors: 邓平; 赵荣鑫; 朱飞翔; 王浩祥; 吴明辉
Original assignee: Southwest Jiaotong University
Current assignee: Southwest Jiaotong University
Priority date: 2021-05-13
Filing date: 2021-05-13
Publication date: 2021-08-10

Abstract

The invention discloses a dead reckoning and positioning method based on pedestrian motion state recognition, comprising the steps of: constructing a pedestrian motion state recognition classification model, identifying the pedestrian motion state, performing step frequency detection, step length estimation, course estimation and dead track reckoning. The beneficial effects of the present invention are: for the motion state of pedestrians in the two-dimensional space, five types of motion states are collected and only three-axis acceleration and three-axis gyroscope data are collected to realize pedestrian navigation and positioning, and the practicability of the dead reckoning method is stronger, It is more conducive to the application and development of the actual application environment. The five types of motion state recognition models are more complete and accurate in modeling and using models to recognize motion states. Gait detection and cadence detection do not need to change the detection method according to the motion state behavior, are more versatile and accurate, and minimize the coupling with the human body motion state recognition method. Dead reckoning can accept more complex types of motion states and is more applicable.

Description

Dead reckoning positioning method based on pedestrian motion state recognition

Technical Field

The invention relates to the technical field of positioning, in particular to a dead reckoning positioning method based on pedestrian motion state identification.

Background

With the rapid development of information technology, location-based services have gradually penetrated aspects of human life as a lifestyle. Indoor positioning has received much attention in the last decade as one of the most challenging technologies in Location Based Services (LBS). However, indoor positioning faces a series of challenges such as severe multipath effects, non-line-of-sight propagation, high signal attenuation and noise interference, etc., relative to an environment where global navigation satellite systems are indispensable or even dominant technologies outdoors. At present, a plurality of challenges still face to the implementation scheme of reliable and high-precision indoor navigation positioning, and the indoor navigation positioning technology of pedestrians is the key for ensuring the success of various positioning services. In recent years, due to the development of semiconductor technology, micro sensors such as acceleration sensors, gyroscopes and magnetometers have been widely integrated into various smart devices, so that the pedestrian indoor navigation positioning technology based on the inertial navigation technology has received more and more attention.

However, in the existing pedestrian inertial navigation positioning technology, there are certain requirements and limitations on the motion posture of the person to be positioned, and only basic simple motion states such as forward walking or jogging are generally considered. For various motion states such as left and right striding, backward walking and the like which may occur when people such as firefighters and the like carry out indoor disaster relief, the existing algorithm is rarely identified and correspondingly processed, so that the positioning performance is rapidly reduced, and the algorithm is difficult to be applied to accurate navigation positioning of pedestrians in such scenes.

Document 1: liu Yu, Zhou Sai, Li Yun Mei, etc. based on the three-dimensional autonomous navigation positioning algorithm [ J ] of human body multi-directional movement, China technical bulletin of inertia, 2016,24(04): 449) 453. The initial phase of different motion of human body is used to distinguish different motion states including walking, backward walking and left and right walking. However, the motion detection mode has poor reliability, and the detection capability is strongly related to the initial phase of the motion.

Document 2: the human motion state recognition [ J ] is based on a built-in sensor of the smart phone, and the communication bulletin, 2019,40(03):157-169. The method includes steps of recognizing actions including walking, running and riding, and finishing motion classification by using a Support Vector Machine (SVM).

Document 3: the application number is CN107084718A, and the invention name is an indoor positioning method based on pedestrian dead reckoning. The invention uses a smart phone as a positioning platform, and uses an Inertial Measurement Unit (IMU) element built in the smart phone as a positioning device, but the invention does not consider the problem of complex motion posture of pedestrians in terms of the motion state of the pedestrians, has strict requirements on positioning motion of the pedestrians, and cannot meet the positioning requirements of ordinary people in daily life.

Document 4: the invention has the application number of CN109827577B, and is named as a high-precision inertial navigation positioning algorithm based on motion state detection. The invention realizes a positioning algorithm based on foot-operated slow walking and slow running motion state identification, only quick and slow motion of pedestrians is considered in the aspect of motion state detection, the used positioning algorithm is based on a foot binding type, and equipment needs to be arranged at the feet of positioning personnel, so that the positioning algorithm is inconvenient in actual use.

Disclosure of Invention

The invention provides a track reckoning positioning method based on pedestrian motion state identification, which realizes accurate navigation positioning of pedestrians in various motion states by identifying and processing common motion states of the pedestrians and improving and correcting the existing inertial navigation positioning technology.

The technical scheme for realizing the purpose of the invention is as follows:

a dead reckoning positioning method based on pedestrian motion state identification comprises the following steps:

step one, constructing a pedestrian motion state identification classification model: acquiring training data of pedestrians in five motion states, and constructing a pedestrian motion state recognition classification model; the five motion states are walking, jogging, left striding, right striding and reversing; the training data comprises three-axis acceleration data;

step two, identifying the pedestrian motion state: collecting test data of pedestrians, and identifying a pedestrian motion state by using a pedestrian motion state identification classification model; the test data comprises triaxial acceleration data and triaxial gyroscope data;

step frequency detection is carried out to obtain single step frequency: step frequency detection is carried out on the triaxial acceleration data acquired in the step two, and single step frequency is obtained;

step four, estimating the step length: if the pedestrian motion state obtained by the identification in the step two is walking, left stepping, right stepping or reversing, estimating a single step by adopting a linear step model and combining single step frequency; if the pedestrian motion state obtained by the identification in the step two is jogging, estimating a single step size by combining a Weinberg nonlinear step size model with a single step frequency;

step five, course estimation: carrying out integral solution on the triaxial gyroscope data acquired in the step two by adopting angular velocity converted based on a quaternion coordinate system to calculate a course angle, and finishing course angle correction by adopting a heuristic offset elimination algorithm;

step six, dead reckoning: and setting the initial position coordinates and the initial course angle of the pedestrian, and updating the position of the pedestrian in the five motion states according to the pedestrian motion state obtained in the step two, the single step length obtained in the step four and the course angle corrected in the step five.

According to a further technical scheme, in the step one, a pedestrian motion state identification classification model is constructed, and the method comprises the following steps:

step 1-1: collecting training samples, and carrying out pretreatment: collecting triaxial acceleration data Acctrain of pedestrians in five motion states_x、AccTrain_yAnd AccTrain_zCombining to obtain combined acceleration data Acctrain; respectively carrying out acceleration filtering treatment to obtain Acctrain'_x、AccTrain′_y、AccTrain′_zAnd AccTrain';

step 1-2: training sample segmentation: acctrain 'in five motion states'_x、AccTrain′_y、AccTrain′_zAnd Acctrain' are respectively divided to form a training sample set Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]](ii) a Wherein, i is 1 to n is the number of the training sample set,

for training the number of sample sets, N_trainFor the total number of training sample samples, n_cSample points for each training sample set [ ·]Represents rounding down; using a label TrainLabels of the training sample set to represent the category of the motion state of the training sample set;

step 1-3: calculating time domain characteristic values of a training sample set;

the time domain characteristic value of the ith training sample set is as follows:

respectively by Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]As a time domain feature value F_1i、F_2i、F_3iAnd F_4iRespectively by Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]Is taken as a time domain characteristic value F_5i、F_6i、F_7iAnd F_8i(ii) a By analogy, Train is used respectively_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]The variance, the mode, the maximum value, the minimum value, the quartile distance and the third quartile are used as time domain characteristic values F_9i～F_32iRespectively by Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]The skewness, kurtosis and average absolute error of the time domain are taken as time domain characteristic values F_36i～F_47i(ii) a Respectively by Train_x[i]And Train_y[i]Cross correlation coefficient, Train_y[i]And Train_z[i]Cross correlation coefficient and Train_z[i]And Train_x[i]As a time domain feature value F_33i、F_34iAnd F_35i；

Respectively by Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]Intermediate variable M incorporating Hjorth parameter₄As time domain eigenvalues

By Train_x[i]Obtaining F (1) as a time domain feature value F as an input set object_52i，

Wherein d [ k ]]Representing the kth data, N, in the input set object d_dFor input of the size of the collection object d, i.e. N_d＝n_c；

Step 1-4: screening effective characteristics of five types of motion states: processing time domain features F₁Said time domain feature F₁For time domain eigenvalues F_1iThe marking of (2): according to the motion state of the training sample set characterized by the TrainLabels, n time domain characteristic values F of n training sample sets_1iRespectively extracting time domain characteristic values of five types of motion states; respectively combining time domain characteristic values corresponding to the five types of motion states, and drawing a characteristic curve graph of five characteristic curves corresponding to the five types of motion states after reordering in sequence; in the characteristic curve graph, the horizontal axis/the vertical axis are the serial numbers of the time domain characteristic values after the time domain characteristic values are rearranged in sequence, and the vertical axis/the horizontal axis are the time domain characteristic values; if there is more than one characteristic curve among the five characteristic curves which is not crossed with other characteristic curves, the time domain characteristic F₁Screening effective characteristics of the motion state corresponding to more than one characteristic curve which is not crossed with other characteristic curves; otherwise, time domain feature F₁Effective characteristics of five motion states are not screened;

processing time domain features F according to a similar method₂～F₅₂Screening effective characteristics of five types of motion states;

step 1-5: establishing a valid feature matrix Q_TrainThe effective feature matrix Q_TrainThe size of (a) is n x m; the values in the first row are m time domain feature values corresponding to the effective features of the five types of motion states screened by the first training sample set according to the steps 1-4, and so on;

step 1-6: establishing a pedestrian motion state identification classification model: normalizing the processed Q_TrainAs the characteristics for training the SVM model, TrainLabels is used as a training sample label for training the SVM model, and a pedestrian motion state recognition classification model SVM-OVO is established by utilizing an SVM one-to-one multi-classifier (OVO); wherein, the kernel function of the SVM is a linear kernel function.

According to a further technical scheme, in the second step, the motion state of the pedestrian is identified, and the method comprises the following steps:

step 2-1: collecting a test sample, and carrying out pretreatment: collecting pedestrian triaxial acceleration data AccTest_x、AccTest_yAnd AccTest_zCombining to obtain a combined acceleration AccTest; respectively carrying out acceleration filtering treatment to obtain AccTest'_x、AccTest′_y、AccTest′_zAnd AccTest'; the sampling frequency of the test sample is equal to that of the training sample;

step 2-2: test sample segmentation: acctest'_x、AccTest′_y、AccTest′_zAnd Acctest' are respectively divided to form a Test sample set Test_x[i']、Test_y[i']、Test_z[i']And Test [ i'](ii) a Wherein i ═ 1 to n' are numbers of the test sample sets,

for testing the number of sample sets, N_testThe total number of sampling points of the test sample; n is_cSample points for each training sample set [ ·]Represents rounding down; target using test sample setLabel TestLabels characterize the category of the motion state of the test sample set; step 2-3: calculating effective characteristic values of the test sample set;

the valid eigenvalues for the ith' test sample set are:

selecting effective characteristics of five types of motion states of the pedestrian motion state recognition classification model from the time domain characteristics of the ith' test sample set, and calculating the effective characteristics according to a method for calculating the same time domain characteristic value of the training sample set;

step 2-4: establishing effective characteristic matrix Q of test sample set_TestThe effective feature matrix Q_TestThe size of (a) is n' × m; wherein, the value of the first row is m effective characteristic values calculated according to the step 2-3 in the first test sample set, and so on;

step 2-5: recognizing the motion state of the pedestrian: normalizing the processed Q_TestInputting a pedestrian motion state recognition classification model SVM-OVO, and assigning values to the label TestLabels of the test sample set to obtain the motion state category of each test sample set.

In a further technical solution, the second step further includes correcting the category of the motion state of the test sample set, that is, the motion state of the test sample set is corrected

Step 2-6: the values of the labels TestLabels of the first, second, third and fourth test sample sets are made equal to W, respectively₁、W₂、W₃And W₄；

Step 2-7: such as W₂≠W₃&&W₂≠W₄&&W₃≠W₄&&W₁＝W₂Or W is₂≠W₃&&W₂≠W₄&&W₃≠W₄&&W₁＝W₂Not satisfying and W₂＝W₄&&W₂≠W₃&&W₃≠W₄Then W will be₃The value of the corresponding label TestLabels is corrected to W₂The value of the corresponding label TestLabels; otherwise, not correcting;

step 2-8: returning to step 2-6 and replacing itChanging to: the values of the labels TestLabels of the second, third, fourth and fifth test sample sets were made equal to W, respectively₁、W₂、W₃And W₄(ii) a Then, executing the step 2-7;

step 2-9: and finishing the correction of the value of the label TestLabels of the test sample set according to the method similar to the steps 2-8.

In a preferred technical scheme, in the third step, step frequency detection is performed to obtain a single step frequency, and the method comprises the following steps:

step 3-1: collecting a test sample, and carrying out pretreatment: collecting pedestrian triaxial acceleration data AccTest_x、AccTest_yAnd AccTest_zCombining to obtain a combined acceleration AccTest; carrying out acceleration filtering processing on the combined acceleration Acctest to obtain Acctest';

step 3-2: wave crest detection: if AccTest' j-1 is satisfied]<AccTest′[j]≤AccTest′[j+1]Then AccTest' j will be]Labeled as peak; acctest' [ j]Represents the resultant acceleration of the jth sample point after filtering processing, and j belongs to [2, N ]_Test-1]，N_testThe total number of sampling points of the test sample; setting the peak minimum threshold σ₁And inter-peak sample point spacing constraint sigma₂The reject value is less than σ₁When the sampling point interval between a plurality of peaks is less than sigma₂Only the peak value with the maximum retention value is discarded, and n is obtained_peakA peak to peak value; recording the index of AccTest' corresponding to the peak value obtained by detection into an array PeakIndex [ l ]]Is the index of AccTest' corresponding to the first peak value, and l is more than or equal to 1 and less than or equal to n_peak；

Step 3-3: searching the left zero point and the right zero point of each peak value: the first peak value PeakIndex [ l ]]Corresponding AccTest' [ PeakIndex [ l ]]]The value, the first zero found by the forward search is designated PeakIndex [ l ]]Left zero point Z of₁[l]The second zero found by the backward search is designated PeakIndex [ l ]]Right zero point Z of₂[l]；

The search range for the left zero is defined as: if it is searched forward to PeakIndex [ l ]]-σ₃The index position has not yet obtained the left zeroThen, the search is stopped, let AccTest' [ PeakIndex [ l ]]-σ₃]Is a left zero point Z₁[l]，σ₃Is a threshold value;

the search range for the right zero point is defined as: if search backward to PeakIndex [ l ]]+σ₄If the index position has not obtained the right zero point, the search is stopped, let AccTest' [ PeakIndex [ l ]]+σ₄]Is a right zero point Z₂[l]，σ₄Is a threshold value;

the zero point is as follows:

if the conditions that Acctest 'j-1 <0 and Acctest' j > is more than or equal to 0 or Acctest 'j-1 >0 and Acctest' j > is less than or equal to 0 are met,

AccTest' j is zero;

step 3-4: calculate the step frequency for each single step: the peak value of the first wave crest corresponds to the first step, and the step frequency of the first step is f_s[l]＝Z₂[l]-Z₁[l]。

In a preferred technical scheme, in the step six, dead reckoning, the method comprises: after the initial position coordinates and the initial course angle of the pedestrians are set, the positions of the pedestrians in the five motion states are updated according to the following formula,

wherein, X [ k ]]And Y [ k ]]Indicating the position of the k-th step of the pedestrian, theta, on the two-dimensional navigation plane_kDenotes the heading angle of the kth step, L k]Represents the step size of the k step; if the motion state of the kth step is reverse, q_kTaking 1, otherwise, taking-1; if the motion state of the kth step is left step or right step, p is_kTaking 1 or-1, otherwise p_kTake 0.

Compared with the prior art, the invention has the beneficial effects that:

1. aiming at the motion state of the pedestrian in a two-dimensional space, five motion states are collected, and only three-axis acceleration and three-axis gyroscope data are collected to realize the navigation and positioning of the pedestrian, so that the track calculation method has stronger practicability and is more favorable for the application and development of practical application environments.

2. The modeling of the five types of motion state recognition models and the recognition of the motion states by using the models are more complete and accurate.

3. The gait detection and the step frequency detection do not need to change the detection method according to the motion state behaviors, are more universal and accurate, and reduce the coupling with the human motion state identification method to the maximum extent.

4. And the dead reckoning can accept more complicated motion state types, and the applicability is stronger.

Drawings

FIG. 1 is a schematic diagram of the main steps of an embodiment.

Fig. 2 is a flow chart of gait detection.

Fig. 3 is a schematic diagram of step frequency detection.

Fig. 4 is a diagram illustrating the results of gait detection and step frequency detection.

FIG. 5 is a diagram illustrating the values of the acceleration x-axis f (1).

Fig. 6 is a flow chart of SVM classification.

Fig. 7 is a schematic diagram illustrating a motion state recognition error recognition result correction.

FIG. 8 is a diagram illustrating a multi-motion state location update.

Fig. 9 is a diagram of a manner in which a locater holds a smartphone.

Fig. 10 is a diagram of positioning effect in a multi-motion state.

Detailed Description

The method comprises the steps of firstly completing data acquisition and preprocessing of pedestrians to be positioned through portable waist-bound IMU positioning equipment, then firstly constructing an SVM multi-classification model of the pedestrian movement state before carrying out movement state identification, or identifying the pedestrian real-time movement state under the established classification model, then correcting individual error identification results through an adjacent gait correlation constraint method based on the human body movement law to realize accurate identification of the movement state in the positioning process of positioning personnel, and finally completing navigation positioning of the complex movement state of the pedestrians in a two-dimensional plane through improving a step length estimation algorithm, correcting a course angle based on an HDE (empirical Drift Elimination) algorithm and designing a position updating algorithm based on the multi-movement state on the basis of a traditional PDR algorithm.

The specific embodiment for realizing the purpose of the invention mainly comprises the following steps:

A. IMU positioning data acquisition: three-axis acceleration and three-axis gyroscope data are collected through IMU equipment carried by a person to be positioned, and preprocessing such as moving average filtering is carried out on the resultant acceleration data after the gravity is removed.

B. Establishing a motion state classification model: firstly, constructing motion state features, respectively extracting time domain features of acceleration x-axis, y-axis, z-axis and three-axis combined acceleration data in 5 motion states of walking, jogging, left striding, right striding and reversing, extracting feature data of 52 feature objects in total, and numbering the feature objects as F according to extraction sequence₁～F₅₂(ii) a Then, screening the effective features, filtering the ineffective features to reduce the interference of the ineffective features on the high-precision classification recognition rate, and finally, setting the number of the effective feature objects for the motion state classification recognition to be 34; then, completing the construction of a feature matrix by using the features of the effective feature objects, and carrying out normalization processing on the features to eliminate the factor of imbalance of feature weights caused by the fact that the feature values of different measurement units are different in size; finally, construction of a human motion state recognition classification model is completed by utilizing an SVM multi-classifier (OVO), and an SVM kernel function selects a linear kernel function.

C. Step frequency detection: using the combined acceleration waveform, firstly adopting a single peak value detection method and combining the minimum peak value threshold value constraint sigma₁Constraint sigma of pedestrian motion period interval₂And completing pedestrian motion detection in a multi-motion state, and then performing single-step left and right zero detection through a zero searching rule, thereby completing single-step frequency detection.

D. Human motion state identification: firstly, in the positioning process, performing feature extraction on the acceleration data preprocessed in the step A according to the effective feature objects screened in the step B and constructing a corresponding feature matrix, then, completing normalization processing of feature extraction in the positioning process by combining the features normalized in the step B, and finally, on the basis of the constructed SVM-OVO classification model, recognizing the motion state of the pedestrian in motion and obtaining a corresponding recognition result.

E. And (3) modifying the motion state identification result: on the basis of the preliminary identification of the motion state, the correction of individual error identification results is realized by an adjacent gait correlation constraint method based on the human motion law.

F. And (3) pedestrian track reckoning in a multi-motion state: firstly, estimating the step length by adopting a linear or nonlinear step length model based on the final identification result of the motion state, then carrying out integral solution on the angular speed converted by a quaternion coordinate system to calculate the course angle, finishing course angle correction by means of an HDE heuristic offset elimination algorithm, and finally finishing the dead reckoning and positioning of the pedestrian in the two-dimensional plane by adopting a dead reckoning algorithm based on the motion state identification.

The process of the present invention is further described in detail below with reference to specific examples.

The steps of establishing the motion state classification model are as follows:

A. setting the motion state of positioning personnel: the positioning personnel finish the acquisition of training data in a two-dimensional indoor scene, and the movement modes of the positioning personnel comprise walking, jogging, left striding, right striding and reversing;

B. collecting training data: the positioning personnel collects training data by binding or carrying an IMU (Inertial Measurement Unit, IMU) attitude sensor at the front part of the waist, and outputs acceleration sensor data with an uncompleted preprocessing result, including triaxial acceleration data Acctrain_x、AccTrain_y、AccTrain_z. Sampling frequency of f_cAnd may be set to 50 Hz.

C. Training data preprocessing: the obtained AccTrain_x、AccTrain_y、AccTrain_zThe resultant acceleration AccTrain is obtained through the combination calculation,

wherein Acctrain [ i ]]The resultant acceleration at time i. Acceleration filtering processing is carried out by utilizing a moving average filtering method, and an object is filtered and outputAre Acctrain respectively_x、AccTrain_y、AccTrain_zAnd Acctrain, the filtered output object is Acctrain'_x、AccTrain′_y、AccTrain′_zAnd AccTrain', the filter equation is:

wherein x [ i ] and y [ i ] respectively represent input and output at the moment i; m is the input object set size; j is the filter window size and can take a value of 3.

D. Training sample segmentation: filtering to obtain Acctrain'_x、AccTrain′_y、AccTrain′_zAnd AccTrain' every n_cRespectively cutting the sampling intervals of the training samples to form a training sample set Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]. Here, i has a value ranging from 1 to n. n is defined as:

wherein [ ·]Denotes rounding down, N_trainFor the total number of samples of the training sample, n_cFor the size of a single training sample set, n is taken in the present invention_c＝1.5f_c. In addition, the array train labels is used for representing the real motion states corresponding to different training sample sets, and the value range is 1: integers between 5, 1, 2, 3, 4 and 5 characterize walking, jogging, left striding, right striding and backstepping, respectively. TrainLabels [ i ]]And representing the real motion state corresponding to the ith training sample set.

E. Extracting effective characteristics of a training sample set: the obtained Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]As an input. For Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]Time domain feature extraction is carried out, the total number of feature data extraction objects is 52, and the feature objects are numbered as F according to the extraction sequence₁～F₅₂(ii) a Then, the effective features are screened, the invalid features are filtered to reduce the interference of the invalid features on the high-precision classification recognition rate, and the number of the effective feature objects finally used for the motion state classification recognition is 34. The specific rules of feature extraction and feature screening are as follows:

1) feature extraction

The objects for feature extraction include mean, mean of absolute value (first data absolute value is calculated and then mean is calculated), variance, mode, maximum, minimum, quarter-quanta, third quartile, cross correlation coefficient, skewness, kurtosis and mean absolute error, and are marked by a mark F₁～F₄₇Marking the different characteristics in sequence, wherein each characteristic is according to the input Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]The objects are numbered sequentially, wherein the cross-correlation coefficient numbering order follows Train_xAnd Train_y、Train_yAxle and Train_z、Train_zAnd Train_xThe cross correlation coefficient calculation order of (1). Furthermore, an intermediate variable M of the Hjorth parameter is introduced₄As feature F₄₈～F₅₁The numbering sequence of which follows Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]And inputting the sequence. Introducing a custom feature F (1) as the feature F₅₂F (1) input object is Train only_x。M₄And f (1) is calculated as:

in the formula, d [ k ]]Representing the kth data, N, in the input set object d_dIs the size of the set d. Can set N_d＝n_cThe alternative to d is Train_x[i]、Train_y[i]、Train_z[i]And Train [ i ]]And the value range of i is 1-n.

2) Feature screening

Aiming at the fact that corresponding characteristic values exist in various types of sports along with the serial numbers of the training sample set under different characteristic objects, the characteristic values are combined into characteristic curves, then the extracted characteristic objects are subjected to characteristic curve drawing to obtain characteristic curve graphs corresponding to different characteristics, and each characteristic curve graph comprises the characteristic curves corresponding to the five motion states. The characteristic curve corresponding to a certain motion state in the characteristic curves does not have obvious curve crossing phenomenon with the characteristic curves corresponding to the rest motion states, and the characteristic can be used as the effective characteristic of the motion state. The numbers of a total of 34 feature objects used for the motion state classification after screening are shown in the following table.

TABLE 1 available characteristics

F. Training sample set feature matrix normalization: forming a feature matrix Q with the size of n × m by using the obtained features of the effective feature object_TrainAnd carrying out normalization processing on the characteristics to eliminate the factor of unbalanced weight of the characteristics caused by the great difference of the characteristic values of different metering units, and carrying out normalization processing on Q_TrainEach line of data is processed into [ -1,1 [ -1 [ ]]An interval. Where m is the total number of valid feature objects.

G. Establishing a motion state classification model: input normalized Q_TrainAs the characteristics for training the SVM model, TrainLabels is input as a training sample label for training the SVM model. The construction of a human motion state recognition classification model SVM-OVO is completed by utilizing an SVM One-to-One multi-classification model (One-Versus-One, OVO), and an SVM kernel function selects a linear kernel function.

After the human motion state classification model is established, the positioning process comprises the following steps:

A. collecting test data: a positioning person collects test data by binding or carrying an IMU (Inertial Measurement Unit) attitude sensor at the front part of the waist, and outputs acceleration sensor data with an unfinished preprocessing result, including triaxial acceleration data AccTest_x、AccTest_y、AccTest_zAnd three-axis gyroscope data GyrTest_x、GyrTest_y、GyrTest_z. Sampling frequency of f_cTotal number of sampling points is N_Test，f_cThe setting should be consistent with the setting in the motion state recognition model establishing step.

B. Preprocessing test data: the obtained AccTest_x、AccTest_y、AccTest_zThe resultant acceleration AccTest is obtained through the combination calculation,

wherein AccTest [ i ]]The resultant acceleration at the ith sample point. Acceleration filtering processing is carried out by utilizing a moving average filtering method, and filtering output objects are AccTest_x、AccTest_y、AccTest_zAnd AccTest, the filtered output object is AccTest'_x、AccTest′_y、AccTest′_zAnd AccTest', the filter equation is:

wherein x [ i ] and y [ i ] respectively represent input and output at the moment i; m is the input object set size; j is the filter window size and takes the value of 3.

C. Step frequency detection: AccTest' is input. In the normal positioning process, a positioning person can have two stages of starting and falling in a single step period in the 5 motion states. By using the traditional step frequency detection algorithm for reference, the invention only detects the acceleration peak value in the multi-motion state to complete the single-step motion detection, namely the gait detection. For a certain AccTest' [ i ]]，i∈[2,N_Test-1]If it satisfies

AccTest′[i-1]<AccTest′[i]≤AccTest′[i+1]，

Acctest' i]Where is marked as the peak. Setting the peak minimum threshold σ₁Inter-peak sample point spacing constraintσ₂I.e. the reject value is less than σ₁When the sampling point interval between a plurality of peaks is less than sigma₂Only the peak mark with the largest value is retained, and the rest peaks are discarded. Obtaining the final reserved peak value number n according to the complaint detection method_peak，n_peakAlso represents the single step number contained in the motion process, and the index positions of the corresponding sampling points are recorded from small to large to n_peakIn the array PeakIndex. For a certain PeakIndex [ i ]],1≤i≤n_peak，PeakIndex[i]The peak position corresponding to the ith step motion. Using PeakIndex [ i ]]The sample point index represented by the value is the reference, which corresponds to AccTest' [ PeakIndex [ i ]]]The first zero point searched before is marked as the left zero point of the step, and the right zero point is the second zero point searched after the peak of the step. For a certain AccTest' [ i ]]，i∈[2,N_Test-1]If it satisfies

Acctest 'i-1 <0 and Acctest' i >0,

or

Acctest 'i-1 >0 and Acctest' i ≦ 0,

acctest' i]Is zero. Setting a zero point maximum search range and left and right zero point maximum search ranges sigma₃And σ₄. By the method, the left zero point Z corresponding to each single step can be obtained₁[i]And right zero point Z₂[i]I represents the ith step, i is equal to [1, N ]_Test]. By calculation of

f_s[i]＝Z₂[i]-Z₁[i]，

f_s[i]Indicating the step frequency of the ith step.

The detection results are shown in fig. 3: in the locating process of the locator, the acceleration of the locator changes periodically, each single step movement has a gait cycle, the acceleration in each single step has a wave peak, and therefore, only the wave peak needs to be found to indicate that there is a single step. The peak position first found will include all peaks, like there will be some invalid peaks due to the acceleration jitter when there is no motion, so the peak minimum threshold σ will appear₁. This occurs because multiple peaks may be detected in a single stepConstrained the interval of sampling points between peaks₂. And if the peaks which do not satisfy the condition are not required, the remaining peaks and the corresponding index positions can be calibrated to form a single step. And finally, the step frequency can be calculated by searching zero points before and after the peak position.

D. Test sample segmentation: acctest 'obtained after filtering'_x、AccTest′_y、AccTest′_zAnd AccTest' every n, respectively_cThe sampling interval of each sample is cut to form a Test sample set Test_x[i]、Test_y[i]、Test_z[i]、Test[i]. Here, i ranges from 1 to n, n being defined as:

[. cndot. ] represents rounding down. In addition, the motion states corresponding to different test sample sets are represented by an array TestLabels, and the value range is 1: integers between 5, 1, 2, 3, 4 and 5 characterize walking, jogging, left striding, right striding and backstepping, respectively. And the TestLabels [ i ] represents the motion state corresponding to the ith training sample set. The TestLabels assignment needs to be done in step G.

E. Extracting the characteristics of the test sample: input Test_x[i]、Test_y[i]、Test_z[i]、Test[i]Calculating characteristic values of effective characteristic items obtained in the process of establishing a human motion state model and sequentially forming a characteristic matrix Q with the size of n multiplied by m_Test. And m is the total number of the effective characteristic objects obtained in the step of establishing the motion state model.

F. Test sample feature normalization: the input is Q_TestBy normalizing Q_TestEach line of data is processed into [ -1,1 [ -1 [ ]]An interval. For the test samples, the preprocessing should be consistent with the training samples, i.e., the maximum and minimum values should be the maximum and minimum values of the training set.

G. Human motion state identification: input normalized Q_Test. On the basis of the established SVM-OVO classification model, the pedestrian movement is carried outAnd identifying the motion state and outputting a corresponding motion state identification result. Namely, the identification process completes the assignment work of the TestLabels.

H. And (3) correcting the identification result of partial error motion state: inputting the TestLabels completing assignment work, and realizing correction of individual motion state error recognition results in the TestLabels by an adjacent gait correlation constraint method based on human motion rules. Taking the current subscript t of the array of the TestLabels as a reference, and marking the label value sequence corresponding to the subscript t-3: t of the TestLabels as W₁、W₂、W₃、W₄The value to be corrected is W₃Fig. 7 shows the condition to be corrected, in which the state 1, the state 2, and the state 3 represent three different motion states. Need to mix W₃The label value corresponding to the TestLabels is corrected to W₂The conditions of (a) are as follows:

①W₂≠W₃&&W₂≠W₄&&W₃≠W₄&&W₁＝W₂；

② does not satisfy first and W₂＝W₄&&W₂≠W₃&&W₃≠W₄；

And thirdly, the other conditions are not corrected.

Wherein, the value sequence of t is 3 to n. In particular, W is not carried out when t is 3₁And (4) assigning and correcting conditions to skip the first step. And the whole correction process is executed according to the value sequence of t.

I. Step length estimation: inputting preprocessed AccTest', single step frequency f_sAnd modified TestLabels, based on which the step length is estimated using a linear or non-linear step model. For walking, left striding, right striding and reverse low speed motion states, a linear step size model determined by step frequency and acceleration variance is adopted as represented by:

L[k]＝A+B·f_s[k]+C·var(a)

for jogging motion states, a Weinberg nonlinear step size model, which is determined only by acceleration extremes, is used as follows:

in the above formula, L [ k ]]Denotes the kth step size, f_s[k]For the K step frequency, var (a) represents the single step acceleration variance, A, B and C are constants, obtained by training, K is a constant, a_maxAnd a_minShowing the maximum and minimum values of the single-step acceleration.

J. Course estimation: input GyrTest_x、GyrTest_y、GyrTest_zThe course angle psi is calculated by adopting the angular velocity based on the quaternion coordinate system conversion through integral solution, and course angle correction is completed by adopting a Heuristic Drift Elimination algorithm (HDE). The four-order Rungestota method is used for resolving the quaternion, and the quaternion is expressed as follows:

Q＝q₀+q₁i+q₂j+q₃k

wherein Q represents a result of a quaternion expression, and Q represents₀、q₁、q₂、q₃Is a real number, i, j, k are imaginary parts. Introduction of four elements

After counting, a rotation matrix of b-system to n-system expressed by quaternion

Can be expressed as:

and (3) obtaining a rotation matrix to complete gyroscope data conversion:

in the formula, ω_x、ω_yAnd ω_zRespectively representing the angular speeds of the x axis, the y axis and the z axis acquired by the gyroscope under the b system, namely GyrTest_x、GyrTest_y、GyrTest_z。

And

respectively, x-axis, y-axis and z-axis angular velocities converted to n-system. Then pass through the pair

The heading angle psi can be obtained by performing integral operation.

And utilizing an HDE algorithm to complete heuristic course compensation based on the turning angular speed. Based on HDE algorithm, the invention divides the main direction of the pedestrian positioning process into 8 main directions, namely the main direction delta is 45 degrees, and the system compensation factor i is_cSet to 0.01.

K. And (3) dead reckoning: and based on the motion state identification result, the single step length estimation result and the corrected course angle, completing the position updating of the pedestrian in the multi-motion state, and realizing the positioning navigation. As shown in fig. 8, the formula for updating the dead reckoning position of the pedestrian in the multi-motion state defined by the present invention is defined as follows:

in the formula, X [ k ]]And Y [ k ]]Indicating the position of the k-th step of the pedestrian, theta, on the two-dimensional navigation plane_kIndicates the heading of the kth step, L k]Denotes the step size of the k-th step, q_kAnd p_kFor distinguishing different values according to different motion states, if the step is in a reverse state, the step q is in a reverse state_kTaking 1, otherwise, taking-1; if the step status is left step, then p_kTaking 1, if it is a right stride, then p_kGet-1, otherwise p_kTake 0.

The experiments using the method of the invention were as follows:

an experimental field: the southwest university of transportation library is outside rectangular square.

Experimental equipment: a smart phone is adopted as positioning equipment. A method for fixing equipment of reference waist binding type positioning personnel, wherein the positioning personnel complete the positioning process through a handheld smart phoneData acquisition, fig. 9. The data sampling rate of the intelligent mobile phone inertial navigation device is set to be 50Hz, and the sensors for positioning comprise an accelerometer and a gyroscope; the size of a moving average filtering window N used in data filtering is set to be 3; the minimum peak value threshold value constraint parameter of gait detection and the pedestrian movement period interval constraint parameter are respectively 2m/s²The maximum search ranges of the left zero point and the right zero point of the step frequency detection are respectively 12 sampling points and 25 sampling points, the specific gait detection process is shown in figure 2, the gait detection effect is shown in figure 4, and the step frequency detection result is shown in figure 3; the step length estimation model parameters are obtained according to offline training of different positioning personnel, and the values of specific parameters A, B, C and K in the experiment are 0.1616, 0.2370, 0.0139 and 0.5885 respectively; the angular interval Δ of the main direction of the heuristic heading drift elimination algorithm (HDE) algorithm is 30 °, and the feedback coefficient is set to 0.01.

The experimental contents are as follows: after the positioning equipment is worn, in the experimental process, an experimenter can randomly change the walking posture by holding the mobile phone, and the main motion modes are walking and jogging. The experiment site selects a library in the Xinan university of transportation, Rhinocpu school district, and the library is just opposite to the square for rectangular motion experiment, and the total length of the reference path is about 262.12 m. Through the system flow of fig. 1, a specific motion state recognition algorithm model is shown in fig. 6, a characteristic value diagram in a model training module is shown in fig. 5, a human motion state error recognition result correction condition based on gait constraint according to a recognition result is shown in fig. 7, a PDR-based multi-motion state track calculation method design is shown in fig. 8, and a calculation result is shown in fig. 10.

Purpose of the experiment: the simulation pedestrian carries out unpredictable motion state switching in the positioning process of the common regular indoor positioning scene, thereby carrying out motion state identification on the pedestrian, further completing self-adaptive pedestrian track calculation, and checking the positioning effect of autonomous inertial navigation in a complex motion state.

The experimental results are as follows: through the positioning effect shown in fig. 10, the calculation result of the pedestrian dead reckoning algorithm based on multi-motion state recognition designed by the invention is basically consistent with the actual route, and the motion state recognition of the pedestrian can be accurately completed. Through repeated comparison experiments under the same experiment condition, the average identification accuracy can reach 99.37%, the average positioning error can reach 1.28%, and the adaptability and the reliability of the algorithm are superior to those of the conventional pedestrian track calculation algorithm based on the complex motion state.

Compared with the existing indoor positioning technology, the invention has the obvious advantages that:

firstly, the invention does not need to plan and arrange in advance, has low requirement on equipment and high flexibility on algorithm application, can realize autonomous navigation of pedestrians in common scenes by directly utilizing a smart phone, and has important reference value for researching cheap and applicable inertial navigation products of the pedestrians.

The invention fully considers different motion modes and laws of human bodies, divides the motion of pedestrians in a two-dimensional space into 5 states of walking, jogging, left striding, right striding and reversing, and the adopted step frequency detection method is suitable for various motion states.

And thirdly, the invention fully considers the motion state of the positioning personnel in the navigation process, identifies the motion state in the positioning process, corrects the motion state identification result by utilizing the motion rule of the human body and realizes the high-precision pedestrian motion state identification.

And fourthly, the step length estimation under the multi-motion state of the pedestrian is completed by adopting a linear and nonlinear step length combined estimation model, so that different motion state modes of the pedestrian are more accurately adapted.

And fifthly, the invention improves the traditional PDR model by utilizing course changes caused by different motion modes, thereby realizing accurate pedestrian track calculation in a multi-motion state.

Claims

1. a dead track reckoning positioning method based on pedestrian motion state identification, is characterized in that, comprises:

Step 1, build a pedestrian motion state recognition and classification model:

Collect the training data of pedestrians in five types of motion states, and build a pedestrian motion state recognition and classification model; the five types of motion states are walking, jogging, left stride, right stride and backwards; the training data includes three-axis acceleration data ;

Step 2: Identify the motion state of pedestrians:

Collect the test data of pedestrians, use the pedestrian motion state identification and classification model to identify the pedestrian motion state; the test data includes three-axis acceleration data and three-axis gyroscope data;

Step 3: Perform step frequency detection to obtain a single-step step frequency:

Perform cadence detection on the triaxial acceleration data collected in step 2 to obtain a single-step cadence;

Step 4, step size estimation:

If the pedestrian motion state identified in step 2 is walking, left stride, right stride or backwards, the linear step size model combined with the single-step stride frequency is used to estimate the single-step stride length; if the pedestrian motion state identified in step 2 is jogging , the Weinberg nonlinear step size model combined with the single-step step frequency is used to estimate the single-step step size;

Step 5, heading estimation:

The three-axis gyroscope data collected in step 2 is used to integrate the angular velocity based on the transformation of the quaternion coordinate system to calculate the heading angle, and the heuristic offset elimination algorithm is used to complete the heading angle correction;

Step 6, dead reckoning:

Set the pedestrian's initial position coordinates and initial heading angle, and update the pedestrian's position in five types of motion states according to the pedestrian motion state obtained in step 2, the single-step step length obtained in step 4, and the heading angle corrected in step 5.

2. a kind of dead reckoning positioning method based on pedestrian motion state identification as claimed in claim 1, is characterized in that, described step 1, constructs pedestrian motion state recognition classification model, and its method is:

Step 1-1: Collect training samples and perform preprocessing: collect the three-axis acceleration data AccTrain _x , AccTrain _y and AccTrain _z of pedestrians in five types of motion states, and combine them to obtain the resultant acceleration data AccTrain; respectively perform acceleration filtering to obtain AccTrain' _x , AccTrain' _y , AccTrain' _z and AccTrain';

Step 1-2: Training sample segmentation: AccTrain′ _x , AccTrain′ _y , AccTrain′ _z and AccTrain′ under five types of motion states are divided into training sample sets Train _x [i], Train _y [i], Train _z [i] and Train[i]; where i=1～n is the number of the training sample set,

is the number of training sample sets, N _train is the total number of training sample sampling points, n _c is the number of sampling points in each training sample set, [ ] means rounded down; use the label TrainLabels of the training sample set to represent the movement of the training sample set the category of the state;

Step 1-3: Calculate the time-domain eigenvalues of the training sample set;

The temporal eigenvalues of the i-th training sample set are:

Take the mean of Train _x [i], Train _y [i], Train _z [i] and Train[i] as time domain eigenvalues F _1i , F _2i , F _3i and F _4i , respectively, take Train _x [i] , Train _y [i], Train _z [i] and Train[i] absolute value mean as time domain eigenvalues F _5i , F _6i , F _7i and F _8i ; and so on, take Train _x [i], The variance, mode, maximum value, minimum value, interquartile range and third quartile of Train _y [i], Train _z [i] and Train[i] are taken as time domain eigenvalues F _9i ～F _32i , The skewness, kurtosis and mean absolute error of Train _x [i], Train _y [i], Train _z [i] and Train[ _i ] are taken as time domain eigenvalues F _36i ～F _47i respectively; The cross-correlation coefficient of i] and Train _y [i], the cross-correlation coefficient of Train _y [i] and Train _z [i], and the cross-correlation coefficient of Train _z [i] and Train _x [i] are taken as time domain eigenvalues F _33i , F _34i and F _35i ;

Take Train _x [i], Train _y [i], Train _z [i] and Train[i] to introduce the intermediate variable M ₄ of the Hjorth parameter as the time domain eigenvalues F _48i ～F _51i ,

Take Train _x [i] as the input set object to obtain f(1) as the time domain eigenvalue F _52i ,

Wherein, d[k] represents the kth data in the input set object d, and N _d is the size of the input set object d, that is, N _d =n _c ;

Steps 1-4: Screen the valid features of five categories of motion states:

Process the time domain feature F ₁ , which is the _{label of the time domain feature value F 1i} _: according to the motion state of the training sample set represented by TrainLabels, from the n time domain feature values F of the n training sample sets In _1i , the time-domain eigenvalues of the five types of motion states are extracted respectively; the time-domain eigenvalues corresponding to the five types of motion states are respectively combined and reordered, and then drawn into a characteristic curve diagram of five characteristic curves corresponding to the five types of motion states; In the characteristic curve diagram, the horizontal axis/vertical axis is the sequence number of the time domain characteristic values after reordering, and the vertical axis/horizontal axis is the time domain characteristic value; for example, among the five characteristic curves, there is more than one characteristic curve that does not intersect with other characteristic curves. characteristic curve, then the time domain feature F ₁ is screened as the valid feature of the motion state corresponding to one or more characteristic curves that do not intersect with other feature curves; otherwise, the time domain feature F ₁ is not screened as the valid feature of the five types of motion states ;

The time domain features F ₂ to F ₅₂ are processed according to the similar method, and the effective features of five types of motion states are screened;

Step 1-5: establish an effective feature matrix Q _Train , and the size of the effective feature matrix Q _Train is n×m; wherein, the value of the first row is the five types of sports screened by the first training sample set according to steps 1-4 m time-domain eigenvalues corresponding to the valid features of the state, and so on;

Step 1-6: Establish a pedestrian motion state recognition and classification model: use the normalized Q _Train as the feature used for SVM model training, use TrainLabels as the training sample label for SVM model training, and use the SVM one-to-one multi-classifier (OVO). ) to establish a pedestrian motion state recognition and classification model SVM-OVO; wherein, the kernel function of SVM is a linear kernel function.

3. a kind of dead reckoning positioning method based on pedestrian motion state identification as claimed in claim 2, is characterized in that, described step 2, identifies pedestrian motion state, and its method is:

Step 2-1: Collect test samples and perform preprocessing: collect pedestrian triaxial acceleration data AccTest _x , AccTest _y and AccTest _z , and combine them to obtain the resultant acceleration AccTest; perform acceleration filtering processing respectively to obtain AccTest′ _x , AccTest′ _y , AccTest' _z and AccTest'; the sampling frequency of the test sample is equal to that of the training sample; Step 2-2: Test sample segmentation: AccTest' _x , AccTest' _y , AccTest' _z and AccTest' are divided respectively to form a test sample set Test _x [i'], Test _y [i'], Test _z [i'] and Test[i']; where i'=1～n' is the number of the test sample set,

is the number of test sample sets, N _test is the total number of test sample sampling points; n _c is the number of sampling points in each training sample set, [ ] means rounded down; use the label TestLabels of the test sample set to characterize the motion of the test sample set the category of the state;

Step 2-3: Calculate the effective eigenvalues of the test sample set;

The valid eigenvalues of the i'th test sample set are:

From the time domain features of the i'th test sample set, select the effective features of the five types of motion states of the pedestrian motion state recognition classification model, and calculate according to the same method of calculating the time domain feature values of the training sample set;

Step 2-4: establish a valid feature matrix Q _Test of the test sample set, the size of the valid feature matrix Q _Test is n'×m; wherein, the value of the first row is the first test sample set according to step 2-3 Calculated m valid eigenvalues, and so on;

Step 2-5: Identify the pedestrian motion state: Input the normalized Q _Test , input the pedestrian motion state recognition classification model SVM-OVO, assign the label TestLabels of the test sample set, and obtain the motion state of each test sample set. category.

4. a kind of dead reckoning positioning method based on pedestrian motion state identification as claimed in claim 3, it is characterized in that, described step 2, also comprises revising the classification of the motion state of the test sample set, namely

Step 2-6: Set the values of the labels TestLabels of the first, second, third and fourth test sample sets to be equal to W ₁ , W ₂ , W ₃ and W ₄ , respectively;

Step 2-7: If W ₂ ≠W ₃ &&W ₂ ≠W ₄ &&W ₃ ≠W ₄ &&W ₁ =W ₂ , or W ₂ ≠W ₃ &&W ₂ ≠W ₄ &&W ₃ ≠W ₄ &&W ₁ =W ₂ is not satisfied And W ₂ =W ₄ && W ₂ ≠W ₃ && W ₃ ≠W ₄ , then modify the value of the label TestLabels corresponding to W ₃ to the value of the label TestLabels corresponding to W ₂ ; otherwise, do not correct;

Step 2-8: Return to step 2-6 and replace it with: Set the values of the labels TestLabels of the second, third, fourth and fifth test sample sets to be equal to W ₁ , W respectively ₂ , W ₃ and W ₄ ; after that, perform steps 2-7;

Step 2-9: According to the method similar to Step 2-8, complete the correction of the value of the label TestLabels of the test sample set.

5. a kind of dead reckoning positioning method based on pedestrian motion state identification as claimed in claim 1, is characterized in that, described step 3, carries out cadence detection, obtains single-step cadence, and its method is:

Step 3-1: Collect test samples and perform preprocessing: collect three-axis acceleration data AccTest _x , AccTest _y and AccTest _z of pedestrians, and combine them to obtain the resultant acceleration AccTest; perform acceleration filtering on the resultant acceleration AccTest to obtain AccTest′;

Step 3-2: Peak detection: if AccTest'[j-1]<AccTest'[j]≤AccTest'[j+1], mark AccTest'[j] as the peak; AccTest'[j] represents the first The resultant acceleration after filtering processing of j sample points, j∈[2,N _Test -1], N _test is the total number of test sample sampling points; set the peak minimum threshold σ ₁ and the sampling point interval constraint σ ₂ between peaks, and discard the value The peak value is less than σ ₁ , and when the sampling point interval between multiple peaks is less than σ ₂ , only the peak value with the largest value is retained and the remaining peak values are discarded to obtain n _{peak peak} -to-peak values; the AccTest' index corresponding to the detected peak-to-peak value is recorded. In the array PeakIndex, PeakIndex[l] is the index of the AccTest' corresponding to the lth peak peak value, 1≤l≤n _peak ;

Step 3-3: Search for the left zero point and right zero point of each peak-to-peak value: the AccTest'[PeakIndex[l]] value corresponding to the l-th peak-to-peak value PeakIndex[l], and the first zero point obtained by searching forward is demarcated as The left zero point Z ₁ [l] of PeakIndex[l], the second zero point obtained by searching backward is calibrated as the right zero point Z ₂ [l] of PeakIndex[l];

The search range of the left zero point is limited as follows: if the left zero point is not obtained by searching forward to the index position of PeakIndex[l]-σ ₃ , stop the search, and let AccTest'[PeakIndex[l]-σ ₃ ] be the left zero point Z ₁ [l ], σ ₃ is the threshold;

The search range of the right zero point is limited as follows: if the right zero point is not obtained by searching backward to the index position of PeakIndex[l]+σ ₄ , stop the search, and let AccTest'[PeakIndex[l]+σ ₄ ] be the right zero point Z ₂ [l ], σ ₄ is the threshold;

The zero point is:

If AccTest'[j-1]<0 and AccTest'[j]≥0, or AccTest'[j-1]>0 and AccTest'[j]≤0, then AccTest'[j] is zero;

Step 3-4: Calculate the step frequency of each single step: the lth peak value corresponds to the lth step, and the step frequency of the lth step is f _s [l]=Z ₂ [l]-Z ₁ [l].

6. a kind of dead reckoning positioning method based on pedestrian motion state identification as claimed in claim 1, is characterized in that, described step 6, dead reckoning, its method is: after setting pedestrian initial position coordinates and initial heading angle , according to the following formula to update the pedestrian's position in five motion states,

Among them, X[k] and Y[k] represent the position of the pedestrian at the kth step on the two-dimensional navigation plane, θ _k represents the heading angle of the kth step, and L[k] represents the step size of the kth step; if the kth step If the motion state is backward, q _k takes 1, otherwise it is -1; if the k-th motion state is left stride or right stride, p _k takes 1 or -1, otherwise p _k takes 0.