A Step Length Estimation Method Based on MEMS Inertial Sensors and FM Broadcast Signals

Abstract

The present invention discloses a step length estimation method based on MEMS inertial sensors and FM broadcast signals. The inertial sensor features and FM broadcast signal features are combined by the support vector machine regression method to achieve step length estimation. The method includes: detecting the number of steps of a pedestrian based on the MEMS accelerometer, extracting multiple features based on the MEMS inertial sensor signals and FM broadcast signals, and using the principal component analysis method to reduce the dimension of the extracted multiple features to eliminate redundant features. Finally, SVR is used to fuse multiple features, and step length estimation is achieved after two stages of training and prediction. The present invention uses MEMS inertial sensors and FM broadcast signals for step length estimation, without the need for prior information such as height or a parameter tuning process; by introducing FM broadcast signal features related to distance change and not affected by movement speed or people, it not only reduces the negative impact of MEMS inertial device deviation and noise on performance, but also ensures the practicability of the algorithm for different speeds and different people.

Description

Technical Field

[0001] The present invention relates to the technical field of indoor navigation and positioning, and particularly relates to a step length estimation method based on MEMS (Micro-Electro-Mechanical System) inertial sensors and FM (Frequency Modulation) broadcast signals. Background Art

[0002] With the wide popularization of smart phones in recent years, the application of indoor positioning has become more and more extensive, including museum guides, shopping guides, search and rescue, mobile advertising, and locatable social networks, etc. The Global Navigation Satellite System (GNSS) has become the standard solution for outdoor positioning after decades of development. However, indoors, the movement of people is more frequent than outdoors. The frequent human activities and a large number of factors such as walls, doors, and furniture cause the GNSS signal to be severely affected by attenuation and multipath indoors, and the condition of unobstructed straight-line propagation is often difficult to achieve. Therefore, the GNSS system receives few satellites and has a low signal strength in the indoor environment, making it difficult to meet the requirements of precise positioning. Indoor positioning systems usually use other means to achieve precise positioning.

[0003] In indoor positioning systems, pedestrian dead reckoning (PDR) systems are widely used because they do not require complex equipment deployment or laborious on-site surveys. Stride length estimation is one of the key components of PDR, and its accuracy will directly affect the performance of PDR. Commonly used step length estimation methods are mainly divided into two types: based on geometric models and based on statistical regression methods. The method based on geometric models defines the analytical expression of step length according to the geometric relationship between certain sizes, angles and displacements of different parts of the human body. Although this method is simple, due to the complexity of the human body and the variability of people and scenes, these models are usually simplified and approximate, and a parameter calibration phase is required to adjust their performance to adapt to each specific user and different walking speeds. The method based on statistical regression predicts step length through certain variables. For example, there is an obvious relationship between step frequency and step length, so it can be used as a predictor of step length. The process of estimating the relationship between variables is called the training process. The data and model estimation techniques used in this task will affect the accuracy and versatility of subsequent step length estimation. This method is very flexible and basically does not require parameter adjustment. However, most existing studies require rich training set data, so the preliminary work is relatively large. In addition, existing studies only use inertial sensor features such as acceleration to predict step length. However, device measurements will introduce noise, especially when low-cost MEMS inertial sensors are selected as measurement tools. The measurement accuracy will be severely limited, and the acceleration characteristics are highly dynamic and are greatly affected by people and the environment. Therefore, without the assistance of other information, the performance of step length estimation is poor.

[0004] Frequency Modulation (FM) broadcasting is a radio broadcasting technology that uses frequency modulation technology to transmit sound signals. Frequency modulation technology changes the frequency of the carrier according to the content of the signal to be transmitted, thereby storing the information in the frequency of the carrier and transmitting it. FM broadcasting is currently the main wireless broadcast signal system and is widely used in cars and mobile phones. FM broadcast signals are sent by broadcasting stations through transmission towers and can be distinguished by frequency, achieving a high coverage rate worldwide. Its propagation conforms to the propagation model theory, that is, as the distance from the signal source increases, the signal strength index (Received Signal Strength Indication, RSSI) decreases. Therefore, the change in RSSI has a certain correspondence with the change in distance and can be used to estimate the step length. Summary of the invention

[0005] Technical problem solved by the present invention: Overcoming the deficiencies of the prior art, providing a step length estimation method based on MEMS inertial sensors and FM broadcast signals. A mechanism for combining inertial sensor features and FM features through the support vector machine regression (SVR) method to achieve step length estimation can effectively overcome the defects of existing step length estimation algorithms, improve the applicability to different speeds and different people, and compared with existing step length estimation technologies, it requires less preliminary work and has strong practicability.

[0006] The technical solution adopted by the present invention is: A step length estimation method based on MEMS inertial sensors and FM broadcast signals, applied to an indoor pedestrian positioning system, characterized in that: combining inertial sensor features and FM features through the use of the SVR method to achieve step length estimation, specifically including the following steps:

[0007] Step 1: Based on the output of the accelerometer in the MEMS inertial sensor, first perform noise reduction and filtering processing, and then combine the zero-crossing detection and duration constraint methods to achieve the detection of the number of steps of the pedestrian, and obtain the number of walking steps of the pedestrian.

[0008] Step 2: Segment the MEMS inertial sensor data and FM broadcast signal data based on the number of steps detection result in Step 1. Each segment corresponds to the data of one step of the pedestrian. Extract multiple features of each segment based on the data of each segment, including the average value, range, variance, median, power spectral centroid, frequency domain entropy feature of the signal strength index (Received Signal Strength Indication, RSSI) of the FM broadcast signal, and the standard deviation and skewness of the three-axis acceleration, the standard deviation and skewness of the three-axis angular velocity, acceleration energy, angular velocity energy, the standard deviation and range of the acceleration amplitude, the standard deviation and range of the angular velocity amplitude, the peak interval time of the acceleration amplitude; use the principal component analysis method to reduce the dimension of the above multiple features, obtain the useful features after removing redundant features, irrelevant features, and interference features, and finally use these useful features to achieve step length estimation;

[0009] Step 3: Use the support vector machine regression SVR to fuse the useful features extracted in Step 2 to achieve step length estimation, specifically including two stages: the training stage and the prediction stage. In the training stage, select data with different speeds as the training set and construct a training model. In this stage, input the true step length of the training set and the useful features obtained in Step 2 into the SVR to obtain the training model; in the test stage, use this training model to predict the step length of the test set data with unknown true step length to complete the step length estimation.

[0010] Further, in step 2, the MEMS inertial sensor data and FM broadcast signal data are segmented based on the step count detection result in step 1, and each segment corresponds to the data of one step of the pedestrian. Multiple features of each segment are extracted based on the data of each segment, including the average value, range, variance, median, power spectral centroid, frequency domain entropy feature of the signal strength index (Received Signal Strength Indication, RSSI) of the FM broadcast signal, as well as the standard deviation and skewness of the three-axis acceleration, the standard deviation and skewness of the three-axis angular velocity, acceleration energy, angular velocity energy, standard deviation and range of the acceleration amplitude, standard deviation and range of the angular velocity amplitude, and peak interval time of the acceleration amplitude; the principal component analysis method is used to reduce the dimension of the above multiple features to obtain useful features after removing redundant features, irrelevant features, and interference features, and finally these useful features are used to implement step length estimation as follows:

[0011] First, based on step 1, feature extraction is performed on the original data of the MEMS inertial sensor and FM broadcast signal for each step of the pedestrian's walking, where the FM broadcast signal data of three channels are used; the original data are the received signal strength (Received Signal Strength Indication, RSSI) of the FM broadcast signal, carrier acceleration, and angular velocity. The specific multiple features extracted are as follows:

[0012] (1) The average value of the RSSI of each FM channel;

[0013] (2) The range of the RSSI of each FM channel;

[0014] (3) The variance of the RSSI of each FM channel;

[0015] (4) The median of the RSSI of each FM channel;

[0016] (5) The power spectral centroid of the RSSI of each FM channel;

[0017] (6) The frequency domain entropy of the RSSI of each FM channel;

[0018] (7) The standard deviation of the three-axis acceleration;

[0019] (8) The standard deviation of the three-axis angular velocity;

[0020] (9) The skewness of the three-axis acceleration;

[0021] (10) The skewness of the three-axis angular velocity;

[0022] (11) The energy of the acceleration;

[0023] (12) The energy of the angular velocity;

[0024] (13) Standard deviation of acceleration amplitude;

[0025] (14) Standard deviation of angular velocity amplitude;

[0026] (15) Range of acceleration amplitude;

[0027] (16) Peak interval time of acceleration amplitude;

[0028] (17) Range of angular velocity amplitude;

[0029] Among the above multi - features, items (1) to (10) contain three feature dimensions. Among them, the feature dimensions contained in items (8) to (10) specifically refer to the x, y, and z axes directions of the carrier coordinate system, while the feature dimensions of items (1) to (7) specifically refer to the three channels of FM. The remaining items contain one feature dimension;

[0030] Secondly, use the principal component analysis method to reduce the dimension of the above - extracted multi - features, obtain the useful features after removing redundant features, irrelevant features, and interference features, and finally use these features to achieve step - length estimation. Principal component analysis is a commonly used unsupervised dimension - reduction method. It realizes this mapping through orthogonal transformation and makes the variances of the obtained principal components the largest in the projection direction. The specific algorithm is as follows:

[0031] (1) Centralize the feature matrix X composed of multi - features, that is, subtract the mean value of each column of feature values from the feature values of each column to achieve centralization;

[0032] (2) Calculate the covariance matrix of the centralized feature matrix;

[0033] (3) Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors;

[0034] (4) Sort the eigenvalues in descending order;

[0035] (5) Take the eigenvectors w1, w2,..., w corresponding to the largest eigenvalues d' Get the transformation matrix

[0036] (6) Calculate the dimension - reduced feature matrix according to Z = XW, obtain the useful features after removing redundant features, irrelevant features, and interference features, and use them to estimate the step - length.

[0037] Further, in step 3, a support vector machine regression (SVR) is used to fuse the useful features extracted in step 2 to achieve step length estimation, which specifically includes two stages: a training stage and a prediction stage. In the training stage, data containing different speeds is selected as the training set and a training model is constructed. In this stage, the true step lengths of the training set and the useful features obtained using step 2 are input into the SVR to obtain the training model. In the testing stage, this training model is used to predict the step lengths of the test set data with unknown true step lengths, and the step length estimation is completed as follows:

[0038] The invention uses a support vector machine regression to combine the selected useful features to estimate the step length. The support vector machine (SVM) is an approximation technique for classification and regression. The invention adopts a special implementation method of SVM, namely support vector regression (SVR), to establish the input-output relationship, and its use specifically includes two processes: a training process and a prediction process. Among them, the data used in the training process is the training set data, and the true step length needs to be known, while the test process uses the test set data and the true step length does not need to be known.

[0039] During the training process, first, the training set data is selected, and a combination method of a group of extremely fast data and a group of extremely slow data is used to construct the training set; then, the true step lengths of the training set data and the calculated useful features are input into the support vector machine to obtain a training model based on the training set data;

[0040] During the prediction process, the calculated useful features of the test set data are input into the support vector machine, and the step lengths of the test set data are predicted based on the training model to obtain the step length estimation results of the test set data.

[0041] The advantages of the present invention compared with the prior art are as follows:

[0042] (1) The present invention uses an intelligent method of support vector machine to achieve step length estimation, without the need for prior information such as the height and leg length of the experimenter, and without the need for a parameter tuning process such as adjusting the k value in the early stage, so it has strong practicability;

[0043] (2) The present invention not only extracts inertial sensor features but also uses FM broadcast signal features. There is a certain corresponding relationship between the change of FM signal RSSI and the change of distance, and the change degree of RSSI under different step lengths is different, which can be used to estimate the step length; in addition, different from inertial sensor features, FM signals are not easily affected by the movement speed and human movement habits. Therefore, by introducing FM broadcast signal features, on the one hand, the negative impact of MEMS device measurement noise on performance can be reduced, ensuring the performance of the step length estimation method of the present invention; on the other hand, it makes up for the defect that acceleration or angular velocity is easily affected by different movement speeds and different people, improving the practicability of the step length estimation method of the present invention;

[0044] (3) The present invention uses FM broadcast signals and MEMS sensors to achieve step length estimation, without the need to use existing radiation source signals within a building or deploy radiation source signals inside and outside the building, without the need to pre - construct a signal feature fingerprint database or use an indoor map for constraint, with low cost and strong practicability;

[0045] (4) The present invention proposes a selection mode for the training set (embodied in claim 3), which overcomes the drawback of existing step length estimation methods based on statistical regression that require a large amount of training set data. It can accurately predict the walking step lengths of different speeds and different people by only using two sets of training sets with different speeds. The preliminary work is less and the practicability is strong, so it has engineering practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the implementation flowchart of the method of the present invention;

[0047] Figure 2 is the schematic diagram of feature extraction of the present invention;

[0048] Figure 3 is the schematic diagram of using the support vector machine regression method to estimate the step length of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0050] As Figure 1 shown, a step length estimation method based on MEMS inertial sensors and FM broadcast signals of the present invention combines inertial sensor features and FM broadcast signal features by using the SVR method to achieve step length estimation. It includes the following steps:

[0051] Step 1: Based on the output of the accelerometer in the MEMS inertial sensor, first perform noise reduction and filtering processing, and then comprehensively use two methods of zero - crossing detection and duration constraint to achieve the detection of the number of steps of a pedestrian, and obtain the number of walking steps of the pedestrian.

[0052] Step 2: Segment the MEMS inertial sensor data and FM broadcast signal data based on the step detection result in Step 1, with each segment corresponding to the data of one step of the pedestrian. Extract multiple features for each segment, including the mean, range, variance, median, power spectral centroid, frequency domain entropy feature of the signal strength index (Received Signal Strength Indication, RSSI) of the FM broadcast signal, as well as the standard deviation and skewness of the three-axis acceleration, the standard deviation and skewness of the three-axis angular velocity, acceleration energy, angular velocity energy, the standard deviation and range of the acceleration amplitude, the standard deviation and range of the angular velocity amplitude, and the peak interval time of the acceleration amplitude; use the principal component analysis method to reduce the dimension of the above multiple features, obtain the useful features after removing redundant features, irrelevant features, and interference features, and finally use these useful features to achieve step length estimation;

[0053] Step 3: Use support vector machine regression SVR to fuse the useful features extracted in Step 2 to achieve step length estimation, which specifically includes two stages: the training stage and the prediction stage. In the training stage, select data with different speeds as the training set and construct a training model. In this stage, input the true step length of the training set and the useful features obtained in Step 2 into SVR to obtain the training model; in the testing stage, use this training model to predict the step length of the test set data with unknown true step length to complete step length estimation.

[0054] The specific content of Step 1 is as follows: Based on the output of the MEMS accelerometer, first perform noise reduction and filtering processing, and then use the two methods of zero-crossing detection and duration constraint to achieve the step detection of the pedestrian and obtain the walking steps of the pedestrian.

[0055] The data input for this part is the three-axis acceleration. First is the noise reduction part. It is achieved by designing a low-pass filter, a mean filter, etc., with the purpose of avoiding the influence of factors such as noise. First, use the low-pass filter to filter out the high-frequency noise in the original data, and then use the simple moving average (SMA) method to smooth the original data. The principle of SMA is to define a window of size N and replace each data value with the mean of adjacent data to reduce the influence of noise on the waveform.

[0056] The principle of step detection is that during the walking process of the pedestrian, the acceleration shows periodic peak-valley changes, and this periodicity will be reflected in the periodic peaks of the acceleration. Therefore, counting the steps can be achieved by detecting the peaks of the acceleration. Here, the acceleration amplitude after removing the gravitational acceleration is used. The acceleration peak starts from 0 and ends at 0. First, use zero-crossing detection to initially find the acceleration peak. Secondly, the duration of the acceleration peak must meet certain conditions to be determined as a valid peak. Through the above methods, each peak of the acceleration can be found, and thus the steps can be detected.

[0057] As shown in Figure 2 Figure [X], Step 2 is specifically as follows: Based on the step count detection result in Step 1, segment the MEMS inertial sensor data and FM broadcast signal data. Each segment corresponds to the data of one step of the pedestrian. Extract multiple features for each segment, including the average value, range, variance, median, power spectral centroid, frequency domain entropy feature of the Received Signal Strength Indication (RSSI) of the FM broadcast signal, as well as the standard deviation and skewness of the three-axis acceleration, the standard deviation and skewness of the three-axis angular velocity, acceleration energy, angular velocity energy, the standard deviation and range of the acceleration amplitude, the standard deviation and range of the angular velocity amplitude, and the peak interval time of the acceleration amplitude; Use the principal component analysis method to reduce the dimension of the above multiple features to obtain useful features after removing redundant features, irrelevant features, and interference features, and finally use these useful features to achieve step length estimation.

[0058] First, based on Step 1, extract features from the original data of the MEMS inertial sensor and FM broadcast signal for each step of the pedestrian's walking. Among them, the FM broadcast signal data of three channels are used. The reason is that for the FM receiver chip, it takes about 0.1 second to tune to a new FM frequency and read the RSSI, which will cause the actual sampling rate of each FM channel to decrease as the number of channels increases. Therefore, considering signal strength, stability, and actual sampling rate comprehensively, the present invention selects three FM channels to extract features. In terms of inertial sensor features, acceleration has a relatively direct relationship with step length. The acceleration feature can reflect the size of the step length. For example, the classic Weinberg formula establishes the relationship between the step length and the range of the acceleration in the z-axis of the navigation system; In addition, the angular velocity feature also has a certain connection with the step length. For example, when the step length is long, the angular velocity energy tends to be greater. In terms of FM features, the theoretical basis is the propagation model theory. The propagation model of the FM broadcast signal can be expressed by the following formula:

[0059] P(d) = P(d0) - 10nlg(d / d0)

[0060] where P(d) is the signal strength at the user's location, P(d0) is the signal strength at the reference point, d is the distance between the user and the reference point, n is the path loss exponent, and d0 is the reference point distance, and the present invention takes d0 as 1m. From the above formula, it can be seen that as the distance from the signal source increases, the signal strength decreases, that is, the signal strength at each position is different. Therefore, there is a certain corresponding relationship between the change of RSSI and the change of distance. The greater the change of distance, the more drastic the corresponding change of RSSI tends to be, that is, the change and the degree of change of RSSI corresponding to different step lengths are different. Based on this, the present invention extracts FM broadcast signal features to estimate the step length.

[0061] The multi - features extracted by the present invention include the statistical features, time - domain features, and frequency - domain features of MEMS inertial sensors and FM broadcast signals. The statistical features are the features that identify the statistical distribution of the signal within a window. It includes measures of central tendency, such as the mean and median, etc., which represent the value of the signal by deriving an average value; it also includes measures of dispersion, such as the standard deviation, variance, range, energy, skewness, etc., which indicate whether these values are close or widely distributed. The time - domain features are the features that judge the way the signal changes over time. For example: the peak - to - peak time (indicating the density of peaks in the waveform). The frequency - domain features are used to analyze the frequency components of the signal. Among them, the frequency - domain entropy represents whether the distribution of the amplitudes of the frequency components is concentrated. The smaller the frequency - domain entropy, the more uniform the frequency response distribution. The larger the value, the more concentrated the frequency response is around certain frequency components. If there are cyclic motion components during the movement, the frequency - domain entropy has a larger value. And the power - spectral centroid attempts to represent the most dominant frequency components in the signal by providing a measure of the mid - frequency components. The specific multi - features extracted are as follows:

[0062] (1) The average value of the RSSI of each FM channel;

[0063] (2) The range of the RSSI of each FM channel;

[0064] (3) The variance of the RSSI of each FM channel;

[0065] (4) The median of the RSSI of each FM channel;

[0066] (5) The power - spectral centroid of the RSSI of each FM channel;

[0067] (6) The frequency - domain entropy of the RSSI of each FM channel;

[0068] (7) The standard deviation of the three - axis acceleration;

[0069] (8) The standard deviation of the three - axis angular velocity;

[0070] (9) The skewness of the three - axis acceleration;

[0071] (10) The skewness of the three - axis angular velocity;

[0072] (11) The energy of the acceleration;

[0073] (12) The energy of the angular velocity;

[0074] (13) The standard deviation of the acceleration amplitude;

[0075] (14) The standard deviation of the angular velocity amplitude;

[0076] (15) The range of the acceleration amplitude;

[0077] (16) Peak interval time of acceleration amplitude;

[0078] (17) Range of angular velocity amplitude;

[0079] Among the above multiple features, items (1) to (10) include three feature dimensions. Among them, the feature dimensions included in items (8) to (10) specifically refer to the x, y, and z axes directions of the carrier coordinate system, while the feature dimensions of items (1) to (7) specifically refer to the three channels of the FM. The remaining items include one feature dimension;

[0080] Secondly, use the principal component analysis method to reduce the dimension of the above-extracted multiple features, obtain useful features after removing redundant features, irrelevant features, and interference features, and finally use these features to achieve step length estimation. The feature space obtained after feature extraction is often a high-dimensional space. The high-dimensional situation may cause problems such as small instance density and difficult distance calculation. At the same time, the extracted features are not necessarily all suitable as classification bases, and there may be some irrelevant features, redundant features, or interference features. These features may not only consume more computing resources but may even have a negative impact on the recognition accuracy. In order to reduce the amount of calculation and complexity, prevent overfitting, and maintain or even improve the recognition accuracy, it is necessary to process the high-dimensional features to obtain low-dimensional features.

[0081] Feature dimension reduction refers to mapping data from a high-dimensional space to a low-dimensional space through a certain mathematical method. Principal component analysis is a commonly used unsupervised dimension reduction method. It realizes this mapping through orthogonal transformation and makes the variance of each principal component the largest in the projection direction. The specific method is as follows:

[0082] (1) Centralize the feature matrix X composed of multiple features, that is, subtract the mean value of each column of feature values from the column to achieve centralization;

[0083] (2) Calculate the covariance matrix of the centralized feature matrix;

[0084] (3) Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors;

[0085] (4) Arrange the eigenvalues in descending order;

[0086] (5) Take the eigenvectors w1, w2,..., w corresponding to the largest eigenvalues d' Obtain the transformation matrix

[0087] (6) Calculate the reduced-dimensional feature matrix according to Z = XW to obtain useful features after removing redundant features, irrelevant features, and interference features for estimating the step length.

[0088] Such asFigure 3 As shown, support vector machine regression SVR is used to fuse the useful features extracted in step 2 to achieve step length estimation, which specifically includes two stages: training stage and prediction stage. In the training stage, data containing different speeds are selected as the training set and a training model is constructed. In this stage, the actual step length of the training set and the useful features obtained in step 2 are input into SVR to obtain the training model; in the testing stage, this training model is used to predict the step length of the test set data whose actual step length is unknown to complete the step length estimation.

[0089] The present invention uses support vector machine regression to combine selected useful features to estimate the step size. Support vector machine is an approximate technique for classification and regression. The support vector regression (SVR) method is used to establish the input-output relationship. Since it is not affected by local minimization and overfitting problems, the SVR method can provide high generalization ability.

[0090] In the regression process of the SVM method, the kernel function plays an important role in projecting the nonlinear model into a high-dimensional space. Therefore, the prediction accuracy is highly dependent on the choice of the kernel function. Here, the Gaussian kernel function is selected as the kernel function because it has less computational complexity and less implementation difficulty.

[0091] Due to the nonlinear regression and high generalization ability of the SVR method, it can be used to combine the extracted inertial sensor features and FM broadcast signal features for step length estimation. Its use specifically includes two processes: training process and prediction process. During training, it is necessary to select the training set data first. In order to improve practicality and reduce the amount of preliminary work, less data should be used as the training set as much as possible; in order to improve the generalization ability of the training model, the training set should try to include the numerical range of the test set data. Based on this, the present invention adopts a combination of a group of extreme fast data plus a group of extreme slow data to construct a training set, thereby improving performance while reducing the preliminary work; then the true step length of the training set data and the calculated feature list are input into the support vector machine to obtain a training model. During prediction, the test set data with unknown true step length is used, and the calculated feature list of the test set data is input into the support vector machine, and its step length is predicted based on the training model to obtain the step length estimation result of the test set data.

[0092] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments, for those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A step length estimation method based on MEMS inertial sensors and FM broadcast signals, which is applied to an indoor pedestrian positioning system, and is characterized in that: The step length estimation is achieved by combining the features of a MEMS inertial sensor including a triaxial accelerometer and a triaxial gyroscope and the features of an FM broadcast signal using the support vector machine regression (SVR) method, which includes the following steps: Step 1: Perform noise reduction filtering on the output of the accelerometer in the MEMS inertial sensor, and then comprehensively use the zero-crossing detection and duration constraint methods to detect the number of steps of the pedestrian, and obtain the step detection result, that is, the number of walking steps of the pedestrian; Step 2: Segment the MEMS inertial sensor data and the FM broadcast signal data based on the step detection result in Step 1, and each segment corresponds to the data of one step of the pedestrian; extract multiple features of each segment from the original data of each segment. The multiple features include the average value, range, variance, median, power spectrum centroid, frequency domain entropy feature of the signal strength RSSI of the FM broadcast signal, and the standard deviation and skewness of the triaxial acceleration, the standard deviation and skewness of the triaxial angular velocity, the acceleration energy, the angular velocity energy, the standard deviation and range of the acceleration amplitude, the standard deviation and range of the angular velocity amplitude, and the peak interval time of the acceleration amplitude; then use the principal component analysis method to reduce the dimension of the multiple features, and obtain the useful features after removing the redundant features, irrelevant features and interference features; Step 3: Use the SVR method to fuse the useful features extracted in Step 2 to achieve step length estimation, which specifically includes two stages: the training stage and the prediction stage. In the training stage, select the data including different speeds as the training set and construct a training model. In this stage, input the true step length of the training set and the useful features obtained in Step 2 into the SVR method to obtain the training model; in the test stage, use this training model to predict the step length of the test set data with unknown true step length, and complete the step length estimation.

2. The step length estimation method based on the MEMS inertial sensor and the FM broadcast signal according to claim 1, wherein: The specific steps of Step 2 include the following steps: First, extract features from the original data of the MEMS inertial sensor and the FM broadcast signal for each step of the pedestrian walking, and use the FM broadcast signal data of three channels; the original data is the received signal strength RSSI of the FM broadcast signal, the carrier acceleration and angular velocity, and the multiple features extracted are as follows: (1) The average value of the RSSI of each FM channel; (2) The range of the RSSI of each FM channel; (3) The variance of the RSSI of each FM channel; (4) The median of the RSSI of each FM channel; (5) The power spectrum centroid of the RSSI of each FM channel; (6) The frequency domain entropy of the RSSI of each FM channel; (7) The standard deviation of the triaxial acceleration; (8) The standard deviation of the triaxial angular velocity; (9) The skewness of the triaxial acceleration; (10) The skewness of the triaxial angular velocity (11) The energy of the acceleration; (12) The energy of the angular velocity; (13) The standard deviation of the acceleration amplitude; (14) The standard deviation of the angular velocity amplitude; (15) The range of the acceleration amplitude; (16) The peak interval time of the acceleration amplitude; (17) The range of the angular velocity amplitude; Among the above multi-features, items (1) to (10) contain three feature dimensions. Among them, the feature dimensions contained in items (8) to (10) specifically refer to the x, y, and z axis directions of the carrier coordinate system, while the feature dimensions of items (1) to (7) specifically refer to the three channels of FM, and the remaining items contain one feature dimension; Secondly, the principal component analysis method is used to reduce the dimension of the above-extracted multi-features, obtaining useful features after removing redundant features, irrelevant features, and interference features. Finally, these features are used to achieve step size estimation; the specific algorithm is as follows: (1) Centralize the feature matrix X composed of multi-features, that is, subtract the mean value of each column of feature values from the feature values of each column to achieve centralization; (2) Calculate the covariance matrix of the centralized feature matrix; (3) Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors; (4) Arrange the eigenvalues in descending order; (5) Take the eigenvector corresponding to the largest eigenvalue to obtain the transformation matrix ; (6) According to calculate the feature matrix after dimensionality reduction to obtain useful features that eliminate redundant features, irrelevant features, and interfering features.

3. The step length estimation method based on the MEMS inertial sensor and the FM broadcast signal according to claim 1, wherein: In step three, the SVR method specifically includes two processes: a training process and a prediction process. Among them, the data used in the training process is the training set data, and the true step size needs to be known, while the data used in the test process is the test set data, and the true step size does not need to be known; During the training process, first select the training set data, and use a combination of a set of extremely fast data and a set of extremely slow data to construct the training set; then input the true step size and the calculated useful features of the training set data into the support vector machine SVR to obtain a training model based on the training set data; During the prediction process, input the calculated useful features of the test set data into the support vector machine, and predict the step size of the test set data based on the training model to obtain the step size estimation result of the test set data.

Images