CN111062412A - Novel intelligent identification method for indoor pedestrian movement speed by intelligent shoes - Google Patents

Abstract

The invention discloses an intelligent identification method of indoor pedestrian movement speed by a novel intelligent shoe, which comprises the following steps of firstly, extracting pedestrian foot inertial sensing data by using an IMU (inertial measurement Unit); secondly, dividing the inertia data with continuous time domains step by adopting an acceleration peak value dividing method; thirdly, extracting the characteristics of the inertial data of each step, inputting the inertial data into a dictionary learning algorithm, and performing model training to obtain a speed recognition model; fourthly, performing single-step division and feature extraction on the newly input data, and identifying speed information according to a speed identification model; fifthly, integrating the characteristics and the speed of a group of newly obtained inertial data with the existing model, and updating the parameters of the model; and sixthly, transmitting the speed, the step number and other information of the pedestrian to a mobile phone app or a notebook computer terminal through a communication module, and realizing the visualization of the indoor pedestrian movement speed. The invention considers the wearing convenience of the whole set of device, integrates the IMU, the MCU and the wireless communication module on the shoe, is convenient and comfortable, and can conveniently identify the indoor pedestrian movement speed.

Description

Novel intelligent identification method for indoor pedestrian movement speed by intelligent shoes

Technical Field

The invention relates to the field of indoor positioning and machine learning, in particular to a method for recognizing indoor pedestrian movement speed by using intelligent shoes.

Background

Machine Learning (ML) has been emerging in recent years as a multi-domain cross discipline involving probability theory, statistics, approximation theory, convex analysis, and algorithmic complex theory. Machine learning mainly studies how a computer realizes the simulation and learning of human behaviors, acquires new knowledge, and improves the performance of the computer by organizing and perfecting the existing knowledge structure. Machine learning, which is the core of Artificial Intelligence (Artificial Intelligence), is a fundamental approach for computers to have Intelligence, which is applied throughout various fields of Artificial Intelligence, such as data processing and analysis and prediction.

With the improvement of living conditions and the development of science and technology, people are in the room for a longer time. Whether the positioning in a commercial place or the rescue in a fire place needs to be capable of accurately knowing information such as real-time position and speed of people. Due to the complexity of indoor environment and various electromagnetic interferences, satellite positioning signals have attenuation and interference problems indoors, and the indoor positioning requirement cannot be met. The Pedestrian Dead Reckoning (PDR) system can well make up for the shortage of satellite positioning in the indoor environment. However, due to the accumulated offset error existing in the PDR system, a Zero-Velocity correction algorithm (ZUPT) is required to eliminate the error, and a Zero-Velocity Update algorithm needs real-time Velocity information as a basis.

Based on the problems existing in indoor positioning, the moving speed of a pedestrian is acquired in real time when the pedestrian moves in an indoor environment, and the speed information is processed and stored. Different people walk or run differently, but for a particular person, features extracted from their foot inertia data reflect well their state of motion and speed. The identification method can be universally applied to different pedestrians, is convenient to wear, simple and convenient to use and high in operation rate, can acquire pedestrian movement speed information in real time, feeds the pedestrian movement speed information back to the zero-speed correction algorithm and the PDR system, and performs indoor pedestrian positioning.

Disclosure of Invention

Aiming at the problems and the defects in the existing method, the invention provides the intelligent identification method of the indoor pedestrian movement speed by the novel intelligent shoe.

The invention discloses an intelligent identification method of indoor pedestrian movement speed by using novel intelligent shoes, which specifically comprises the following steps:

firstly, extracting inertial sensing data of feet of pedestrians by using an Inertial Measurement Unit (IMU);

secondly, dividing the continuous inertia data of the time domain step by adopting an acceleration peak value division method, and detecting

The peak value of the acceleration sensor is divided step by step, and the two norms of the acceleration of the three axes of x, y and z

Is defined as shown in formula (1):

uploading and storing the divided single step data,

thirdly, extracting the characteristics of the inertial data of each step, inputting the inertial data into a dictionary learning algorithm, and performing model training to obtain a speed recognition model;

firstly, performing feature extraction on the inertia data which is finished step by step, performing increase and decrease experiments on various statistical data derived from the inertia data, comparing the performance of a speed identification model, and finally selecting 33-dimensional features;

the dictionary learning algorithm is used for acquiring more intrinsic characteristic representation so as to improve the accuracy of speed identification; use of

Represents a training sample, wherein Y₁,...,Y_CTraining samples representing a total of class C, Y_i(i ═ 1.., C) can be split into y₁,...,y_NThe total number of N training sample data is,

meaning a linear space of N x N dimensions, described by trainingThe dimensions of the sample matrix Y are trained. The method takes divided single-step inertial data as a sample unit, and is characterized in that refined 33-dimensional features are adopted, a label is a speed ground channel used for training a treadmill experiment of a speed recognition model, C is a speed category number, n is a feature dimension, and the purpose of dictionary learning is to learn a latent variable projection dictionary

And a projection coefficient matrix

Wherein K is the atomic weight of the dictionary,

the dimension describing the learning dictionary D is n × K, consisting of K n-dimensional vectors D_i(i ═ 1.., K);

it is described that the projection coefficient matrix X has dimensions K × N, consisting of N K-dimensional vectors X_i(i ═ 1.., N);

the objective function of the dictionary learning algorithm used is shown in equation (2):

s.t.||d_i||²＝1,i＝1,...,K

setting the maximum iteration number as T due to the mutual iteration relation of D, X, V and L_maxBefore reaching the maximum iteration number, taking the minimum value of the formula (2);

wherein Y is expressed as a training sample, D is a learning dictionary, X is a projection coefficient matrix, V is a coding coefficient matrix, L is a graph Laplacian matrix, Tr is a trace operation of solving the matrix, α, gamma is a regularization parameter, and the constraint condition satisfied by the first half part of the s.t. expression (2) is that all D is expressed as D_iTwo norms d_i||²Are all 1, i has a value range of 1,.. K,

iterating the ith time to obtain DⁱAnd XⁱThen, the next iteration D is obtained by giving equation (2)ⁱ⁺¹And Xⁱ⁺¹(ii) a Set a good maximum number of times T by iteration_maxThen, the final product is obtained

And

the two are jointly used as a speed classifier;

fourthly, performing single-step division and feature extraction on the newly input data, and identifying speed information according to a speed identification model;

fifthly, integrating a group of newly obtained characteristics and speed data with the existing model, updating model parameters, and improving the identification performance of the model on input speed data;

and sixthly, transmitting the speed, the step number and other information of the pedestrian to a mobile phone app or a notebook computer terminal through a communication module, and realizing the visualization of the indoor pedestrian movement speed.

Compared with the prior art, the invention considers the wearing convenience of the whole set of device, integrates the IMU, the MCU and the wireless communication module on the shoe, is convenient and comfortable, and can conveniently recognize the indoor pedestrian movement speed.

Drawings

FIG. 1 is a general flow chart of the intelligent identification method of indoor pedestrian movement speed by the novel intelligent shoe of the invention;

FIG. 2 is a basic schematic block diagram of an indoor pedestrian speed intelligent identification device;

FIG. 3 is an illustration of an embodiment of a smart shoe;

reference numerals: 1. wireless communication module, 2, IMU, 3, microprocessor MCU, 4, inertial sensor (can imbed the device in the inside of shoes with the help of the zip fastener).

FIG. 4 is a result of acceleration peak single step data division;

fig. 5 shows the speed recognition result.

Detailed Description

The technical solution of the present invention is described in detail below with reference to the accompanying drawings and examples.

As shown in fig. 1, it is an overall flow chart of the method for intelligently identifying the indoor pedestrian movement speed by using the novel intelligent shoe of the present invention. The method specifically comprises the following steps:

firstly, extracting inertial sensing data of feet of pedestrians by using an Inertial Measurement Unit (IMU);

secondly, dividing the inertia data with continuous time domains step by adopting an acceleration peak value dividing method;

by detecting

The peak value of the acceleration sensor is divided step by step, and the two norms of the acceleration of the three axes of x, y and z

Is defined by the formula (1):

the division result is shown in fig. 4, the horizontal axis is a sampling point, the vertical axis is the two-norm data of the triaxial acceleration calculated by the formula (1), the star mark represents the identified two-norm peak value of the acceleration, and the step result is marked by a dotted line; the divided single step data is uploaded to a hundred-degree network disk;

thirdly, extracting the characteristics of the inertial data of each step, inputting the inertial data into a dictionary learning algorithm, and performing model training to obtain a speed recognition model;

firstly, feature extraction is carried out on the inertia data which is finished step by step, various statistical data derived from the inertia data are subjected to increase and decrease experiments through ablation experiments (which can be regarded as a control variable method) and reading documents, the performance of a speed identification model is compared, and finally, 33-dimensional features are selected and shown in table 1.

The dictionary learning algorithm is used for acquiring more intrinsic characteristic representation so as to improve the accuracy of speed identification; use of

Represents a training sample, wherein Y₁,...,Y_CTraining samples representing a total of class C, Y_i(i ═ 1.., C) can be split into y₁,...,y_NThe total number of N training sample data is,

meaning a linear space of N x N dimensions, describing the dimensions of the training sample matrix Y. The method takes divided single-step inertial data as a sample unit, and is characterized in that refined 33-dimensional features are adopted, a label is a speed ground channel used for training a treadmill experiment of a speed recognition model, C is a speed category number, n is a feature dimension, and the purpose of dictionary learning is to learn a latent variable projection dictionary

And a projection coefficient matrix

Wherein K is the atomic weight of the dictionary,

the dimension describing the learning dictionary D is n × K, consisting of K n-dimensional vectors D_i(i ═ 1.., K);

it is described that the projection coefficient matrix X has dimensions K × N, consisting of N K-dimensional vectors X_i(i ═ 1.., N);

the objective function of the dictionary learning algorithm used is shown in equation (2):

s.t.||d_i||²＝1,i＝1,...,K

because of the mutual iterative relationship of D, X, V and L, the maximum iterative times are set to be T_maxBefore reaching the maximum iteration number, the formula (2) is optimizedSmall value, where Y is the training sample, D is the learning dictionary, X is the projection coefficient matrix, V is the coding coefficient matrix, L is the graph Laplace matrix, Tr is the trace operation of the matrix, α, gamma is the regularization parameter, and s.t. means that the first half of equation (2) satisfies the constraint condition that all D_iTwo norms d_i||²Are all 1, i has a value range of 1.

Iterating the ith time to obtain DⁱAnd XⁱThe formula (2) is given to find D of the next iterationⁱ⁺¹And Xⁱ⁺¹(ii) a Set a good maximum number of times T by iteration_maxThen, the final product is obtained

And

the two are jointly used as a speed classifier;

fourthly, performing single-step division and feature extraction on the newly input data, and identifying speed information according to a speed identification model;

fifthly, integrating a group of newly obtained characteristics and speed data with the existing model, and updating the model to achieve higher recognition rate and better robustness;

and sixthly, transmitting the speed, the step number and other information of the pedestrian to a mobile phone app or a notebook computer terminal through a communication module, and realizing the visualization of the indoor pedestrian movement speed.

Fig. 2 is a basic schematic block diagram of an indoor pedestrian speed intelligent recognition device.

The related algorithm is described in detail as follows:

1. peak detection&And single step division. The method of the invention is based on the basis that the inertial data are accurately and correctly classified into steps, so the problem to be solved is that the steps are accurate and the rest parts of the PDR system and the zero-speed updating algorithm are not influenced. The invention selects the two norms of the triaxial acceleration data

AsThe step division is based on the fact that when the data reaches the peak, the most probably is the time when the foot is static relative to the ground (namely, when the acceleration is the maximum, the most probably sole is close to the ground), so that the single step division is more accurate and does not influence the subsequent steps.

2. And (5) feature extraction. There are many statistical features that can be selected, but not every feature is suitable for the job of speed recognition. Through innovation, deletion and summary of the features, 33-dimensional statistical features (shown in table 1) are finally refined as feature expressions (dictionary learning algorithm) of each group of data.

3. Feature representation (dictionary learning algorithm). This section includes the algorithm improvement: the traditional machine learning algorithm comprises an SVM (support vector machine), Naive Bayes (Naive Bayes), a K-nearest neighbor algorithm and the like, improves an algorithm for training a recognition model and performing speed recognition, and selects an algorithm for dictionary learning. The dictionary learning algorithm can obtain stronger robustness and characteristic expression with more discriminative performance on sample data, and can improve the identification accuracy while shortening the algorithm time. The method not only selects the characteristics capable of reflecting the speed characteristics, but also removes redundant characteristics, and prevents the reduction of the recognition rate and the occurrence of overfitting. A specific 33-dimensional feature overview is shown in table 1.

TABLE 1

The first three steps are fused, and meanwhile, the step number and the identified speed information are input into the mobile phone app or the notebook computer terminal through the wireless communication module, so that a user can conveniently know the motion information of the user in the indoor environment in real time.

Fig. 3 is a diagram of an embodiment of a smart shoe. On the basis of common sports shoes, preset positions of a plurality of devices such as an inertial sensor are set, mark points are made, and only a specified data acquisition module, a data processing and storage module or a communication module needs to be installed at the specified positions during mass production, so that the cost is low. The embodiment comprises a wireless communication module 1, an IMU2, a microprocessor MCU 3 and an inertial sensor 4 to realize the intelligent identification method of the novel intelligent shoe for the indoor pedestrian movement speed. The original inertial data are collected through the foot inertial sensor 4, the characteristics are extracted, and the wearable recognition device for recognizing the speed is carried out in real time after the recognition model is trained. The required inertial sensing unit has higher portability degree, can be bound on shoes, has a wireless communication function, can detect the real-time speed of pedestrians, and then serves an indoor positioning system. The method provides convenience for people to acquire the position and the state of the people in the indoor environment in real time.

The present invention is not limited to the specific steps described above. The invention extends to any novel feature or any novel combination of features disclosed in this specification or to any novel combination of steps. In summary, this summary should not be construed to limit the present invention.

Claims (1)

1. The method for intelligently identifying the indoor pedestrian movement speed by the novel intelligent shoe is characterized by comprising the following steps:

firstly, extracting inertial sensing data of feet of pedestrians by using an Inertial Measurement Unit (IMU);

secondly, dividing the continuous inertia data of the time domain step by adopting an acceleration peak value division method, and detecting

The peak value of the acceleration sensor is divided step by step, and the two norms of the acceleration of the three axes of x, y and z

Is defined as shown in formula (1):

uploading and storing the divided single step data,

thirdly, extracting the characteristics of the inertial data of each step, inputting the inertial data into a dictionary learning algorithm, and performing model training to obtain a speed recognition model;

firstly, performing feature extraction on the inertia data which is finished step by step, performing increase and decrease experiments on various statistical data derived from the inertia data, comparing the performance of a speed identification model, and finally selecting 33-dimensional features;

the dictionary learning algorithm is used for acquiring more intrinsic characteristic representation so as to improve the accuracy of speed identification; use of

Represents a training sample, wherein Y₁,...,Y_CTraining samples representing a total of class C, Y_i(i ═ 1.., C) can be split into y₁,...,y_NThe total number of N training sample data is,

meaning a linear space of N x N dimensions, describing the dimensions of the training sample matrix Y. The method takes divided single-step inertial data as a sample unit, and is characterized in that refined 33-dimensional features are adopted, a label is a speed ground channel used for training a treadmill experiment of a speed recognition model, C is a speed category number, n is a feature dimension, and the purpose of dictionary learning is to learn a latent variable projection dictionary

And a projection coefficient matrix

Wherein K is the atomic weight of the dictionary,

the dimension describing the learning dictionary D is n × K, consisting of K n-dimensional vectors D_i(i ═ 1.., K);

it is described that the projection coefficient matrix X has dimensions K × N, consisting of N K-dimensional vectors X_i(i ═ 1.., N);

the objective function of the dictionary learning algorithm used is shown in equation (2):

s.t.||d_i||²＝1,i＝1,...,K

setting the maximum iteration number as T due to the mutual iteration relation of D, X, V and L_maxBefore reaching the maximum iteration number, taking the minimum value of the formula (2);

wherein Y is expressed as a training sample, D is a learning dictionary, X is a projection coefficient matrix, V is a coding coefficient matrix, L is a graph Laplacian matrix, Tr is a trace operation of solving the matrix, α, gamma is a regularization parameter, and the constraint condition satisfied by the first half part of the s.t. expression (2) is that all D is expressed as D_iTwo norms d_i||²Are all 1, i has a value range of 1,.. K,

iterating the ith time to obtain DⁱAnd XⁱThen, the next iteration D is obtained by giving equation (2)ⁱ⁺¹And Xⁱ⁺¹(ii) a Set a good maximum number of times T by iteration_maxThen, the final product is obtained

And

the two are jointly used as a speed classifier;

fourthly, performing single-step division and feature extraction on the newly input data, and identifying speed information according to a speed identification model;

fifthly, integrating a group of newly obtained characteristics and speed data with the existing model, updating model parameters, and improving the identification performance of the model on input speed data;

and sixthly, transmitting the speed, the step number and other information of the pedestrian to a mobile phone app or a notebook computer terminal through a communication module, and realizing the visualization of the indoor pedestrian movement speed.

Images