Multi-dimensional information constrained adaptive dead reckoning positioning method

Abstract

The invention relates to the technical field of positioning and navigation, in particular to a multi-dimensional information constrained adaptive dead reckoning positioning method. The method comprises the following steps: firstly, acquiring acceleration, geomagnetism and gyroscope data, and preprocessing the acquired data; secondly, performing step number detection by using a peak detection method with peak-valley pairing constraint of an acceleration threshold value and a time threshold value, calculating a step length by adopting an Adam optimization step length estimation algorithm based on multi-dimensional information constraint, and performing heading estimation by adopting a method of fusing Sage-Husa adaptive filtering of H infinity constraint with gyroscope and magnetometer data; finally, continuous dead reckoning is carried out on pedestrian displacement, and low-cost and high-robustness pedestrian navigation positioning is achieved.

Description

Technical Field

[0001] This invention discloses an adaptive dead reckoning positioning method with multi-dimensional information constraints, belonging to the field of positioning and navigation. Background Technology

[0002] In today's era of rapid technological advancement, smart living has quietly become a mainstream trend. The widespread adoption of mobile smart terminals such as smartphones and smartwatches has greatly promoted the development of the location services market. Application scenarios such as product location detection in large shopping malls, personnel location tracking in special locations, finding available parking spaces in underground parking garages, locating people in mine accidents, and precise advertising in shopping malls are constantly emerging. Since people's daily lives are closely intertwined with indoor environments—statistics show that people spend 70% to 90% of their time indoors—the reliability of indoor navigation technology has become a crucial factor. While GPS can provide high-precision outdoor positioning, it cannot provide high-precision indoor location services due to signal obstruction. Driven by this demand, low-cost, high-precision indoor positioning technology has naturally become a research focus in the fields of navigation and location-based services (LBS).

[0003] Currently, various indoor positioning methods have been proposed both domestically and internationally, such as Wi-Fi positioning, Bluetooth positioning, infrared positioning, radio frequency identification (RFID) positioning, ultrasonic positioning, and ultra-wideband (UWB) positioning. Each technology has its own advantages and disadvantages. The complex and variable indoor environment leads to significant signal transmission loss and susceptibility to multipath propagation. This limits the application of Wi-Fi, Bluetooth, and ultrasonic positioning technologies in indoor positioning, while RFID and UWB technologies require the deployment of additional equipment. Inertial navigation systems (INS) are relative positioning technologies that do not rely on infrastructure, and Pedestrian Dead Retrieval (PDR) is a type of INS. It is a method that uses mobile smart terminal sensors to calculate pedestrian trajectories, possessing strong autonomy and continuity.

[0004] However, PDR still faces significant challenges in practical applications. On the one hand, the inherent noise and drift problems of low-cost sensors, as well as the complex motion states of pedestrians during walking, affect the accuracy of step length and heading estimation. On the other hand, since PDR is essentially an integral algorithm, its positioning error accumulates over time, leading to a significant deviation of the positioning result from the true location after long-term operation. This problem is particularly prominent in large and complex indoor environments such as libraries and underground parking lots. Frequent turns and drastic speed changes in user walking paths cause significant errors in traditional step length models based on fixed coefficients or simple empirical formulas. At the same time, numerous metal structures and electromagnetic interference in the environment severely distort the geomagnetic field, causing inaccurate heading information calculated using magnetometers, further exacerbating positioning deviations.

[0005] To address the rapid decline in pedestrian positioning accuracy in large, complex indoor environments such as libraries and underground parking lots, targeted improvements are needed to the step size and heading estimation in PDR (Pedestrian Direct Recognition). In these scenarios, users' walking patterns are variable, and their paths frequently turn, leading to significant errors in traditional fixed step size models. Simultaneously, the metal structures cause severe interference with magnetometers, resulting in inaccurate heading estimation. Therefore, a dynamically adaptive step size model can be constructed by fusing multi-source information such as acceleration time-frequency domain features and step frequency. This model, combined with gyroscope integration to optimize heading estimation, significantly improves the robustness and accuracy of indoor pedestrian positioning without relying on external signals. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive dead reckoning positioning method with multi-dimensional information constraints, which does not require additional infrastructure, overcomes the shortcomings of cumulative dead reckoning errors, and achieves low-cost, high-precision positioning.

[0007] To achieve the above objectives, this invention provides a dead reckoning and positioning method with multi-dimensional information constraints, comprising the following steps:

[0008] Step 1: Collect inertial sensor data and preprocess the data;

[0009] Step 2: Count the number of steps using the peak detection method with peak-valley pair constraints;

[0010] Step 3: Extract feature step frequency and variance from the preprocessed acceleration data;

[0011] Step 4: Calculate the fine step size by fusing multi-dimensional information and coarse step size estimation through Adam parameter optimization;

[0012] Step 5: Estimate the heading by integrating the preprocessed gyroscope data and directly measuring the magnetometer data;

[0013] Step 6: Calculate the final fused heading angle using Sage-Husa adaptive filtering with H∞ constraints;

[0014] Step 7: Perform the final positioning estimate using the dead reckoning formula to predict the target's location.

[0015] Optionally, during the execution of step 1, the first and last few seconds of the original data are first removed, then the accelerometer and gyroscope data are processed by moving average filtering, and the geomagnetic data are processed by empirical mode decomposition (EMD) to reduce noise.

[0016] Optionally, during the execution of step 2, different acceleration thresholds are set according to individual differences. When the detection value triggers the threshold, candidate peaks / troughs are marked alternately, and adjacent marks must meet the preset time interval threshold to be determined as valid motion peaks.

[0017] Optionally, during the execution of step 3, a nonlinear expression of the walking frequency and variance is extracted from the acceleration data, and a zero-mean unit variance normalization process is applied to it.

[0018] Optionally, during the execution of step 4, a rough estimate of the step size at the current moment is obtained by using the maximum and minimum values ​​of acceleration and the step size of the previous three steps. Then, Adam parameter optimization is used to perform a weighted fusion of the rough step size with the nonlinear expression of the normalized walking frequency and variance, thereby obtaining a fine estimate.

[0019] Optionally, during the execution of step 5, the preprocessed gyroscope and magnetometer data are unified to the same coordinate system using the quaternion method. Then, the heading is calculated by direct integration of the gyroscope data and by geometric calculation of the magnetometer data.

[0020] Optionally, during the execution of step 6, the present invention uses the heading angle calculated by the gyroscope as the state vector and the heading angle calculated by the magnetometer as the measurement vector, and then uses Sage-Husa adaptive filtering to perform the final fused heading estimation. In order to suppress the impact of external disturbances on system performance, H∞ constraint is used to minimize the maximum gain of the interference signal to the controlled output, ensuring that the system can maintain stable tracking performance under disturbances, thereby obtaining a better heading angle.

[0021] Optionally, during the execution of step 7, the initial position of the target to be measured is first obtained, and then the target step length and movement direction are calculated through steps 2, 4 and 6. Finally, the position coordinates of the pedestrian are continuously estimated.

[0022] This invention provides an adaptive dead reckoning positioning method with multi-dimensional information constraints. First, acceleration, geomagnetic, and gyroscope data are collected and preprocessed. Then, a peak detection method with peak-valley pairing constraints and acceleration and time thresholds is used for step count detection. An Adam-optimized step size estimation algorithm based on multi-dimensional information constraints is employed to calculate the step size. Finally, a Sage-Husa adaptive filtering method with H∞ constraints is used to fuse gyroscope and magnetometer data for heading estimation. Finally, continuous dead reckoning of pedestrian displacement is performed, achieving low-cost and highly robust pedestrian navigation and positioning. Attached Figure Description

[0023] Figure 1 This is a framework diagram of the adaptive dead reckoning and positioning method with multi-dimensional information constraints of the present invention;

[0024] Figure 2 This is a plan view of the test site in a specific embodiment of the present invention;

[0025] Figure 3 This is a comparison chart of step size error results in a specific embodiment of the present invention;

[0026] Figure 4 This is a comparison chart of heading angle error CDF in a specific embodiment of the present invention.

[0027] Figure 5 This is a comparison diagram of the positioning results of different methods in a specific embodiment of the present invention.

[0028] Figure 6 This is a flowchart of the overall scheme in a specific embodiment of the present invention. Detailed Implementation

[0029] The embodiments of the present invention are described in detail below, and examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] Please see Figures 1 to 6 This invention provides an adaptive dead reckoning and positioning method with multi-dimensional information constraints, comprising the following steps:

[0031] S1: Data Acquisition and Preprocessing. First, the first and last few seconds of the raw data are removed. Then, the accelerometer and gyroscope data are processed by moving average filtering, and the geomagnetic data is denoised using Empirical Mode Decomposition (EMD).

[0032] S2: Step Frequency Detection. Step frequency detection is the primary step in a PDR system, its main purpose being to accurately detect the number of steps and step frequency taken by a pedestrian. Because walking is a regular movement, the accelerometer will show regular changes when a pedestrian walks, and step count can be estimated by analyzing the changes in acceleration data. This invention uses a peak detection algorithm for step count detection, the steps of which are as follows:

[0033] 1) Obtain acceleration data in the global coordinate system using a rotation matrix;

[0034] 2) Remove gravitational acceleration and perform data filtering;

[0035] 3) The local maximum values ​​in the search acceleration data are taken as peak data, and the local minimum values ​​are taken as trough data;

[0036] 4) Set a peak threshold and check if the peak value being searched is greater than the threshold. If it is, it is recorded as a valid peak value; the same applies to troughs.

[0037] 5) After the inspection is completed, count the number of steps.

[0038] S3: Step Length Estimation. Step length estimation is a crucial step in PDR localization; the accumulated errors at each step significantly impact the final positioning result. Commonly used step length estimation models include the Weinberg, Scarlet, and Kim models. A pedestrian's step length is not only related to direct motion parameters such as acceleration collected by inertial sensors but may also be coupled with various other features. Therefore, theoretically, incorporating multiple features to estimate the step length can yield more accurate results. However, introducing too many features can lead to data redundancy, increase computational costs, and potentially affect model performance due to multicollinearity among features.

[0039] Considering the dynamic characteristics of pedestrian gait, the mapping relationship between stride length and multiple feature classes is complex, requiring parameter optimization to adapt to complex scenarios. Traditional regularized regression methods have significant limitations: Ridge regression relies solely on L2 regularization to constrain parameters, failing to filter redundant features, and its fixed learning rate cannot adapt to gait variations. While LASSO regression can filter features through L1 regularization, it is prone to accidentally deleting key features and suffers from low convergence efficiency. Adam optimization, on the other hand, uses first-order momentum to smoothly update the direction and second-order momentum to adaptively adjust the learning rate, balancing convergence speed and dynamic adaptability, and can also improve stability through bias correction. Therefore, this paper chooses Adam optimization for stride length estimation to efficiently learn the dynamic relationship between features and stride length.

[0040] To address the aforementioned issues, this paper proposes a multi-dimensional information-constrained Adam optimization step size estimation model. This model introduces more dimensional feature variables to achieve accurate step size estimation. First, a coarse step size is estimated using the difference between the local maximum and minimum values ​​of the acceleration at the three previous time steps and the current time step, as shown in the equation.

[0041]

[0042] in , , The step size for the first three steps, , , It is a weighting factor. and These are the local maximum and minimum values ​​of vertical acceleration, respectively.

[0043] Then, walking frequency and acceleration variance are used as constraints to correct the coarse step size, thus obtaining the predicted step size. Of course, to improve the stability of the data and the effect of subsequent model training, the two added motion information parameters need to be normalized to zero mean and unit variance.

[0044]

[0045]

[0046] in It is the normalized walking frequency. It is the normalized variance of acceleration.

[0047] After obtaining the coarse step size, normalized walking frequency, and acceleration variance, Adam optimization obtains the optimal step size estimate by dynamically optimizing the feature coefficients. The loss function for Adam optimization is as follows:

[0048]

[0049] in It is the step size. It is a true correlation statistic. It is a characteristic independent variable.

[0050] Find the gradient of the loss function with respect to the coefficients.

[0051]

[0052] in It optimizes the weighting coefficients. It is the gradient of the coefficients. It is the first Prediction error of one step.

[0053] In each iteration, to smooth the gradient update direction, the first-order momentum is... With the help of parameters It integrates historical momentum information with the current gradient. To adaptively adjust the learning rate scale, second-order momentum... Combination The formula is as follows: [Formula omitted for brevity] and the current gradient are used to capture the magnitude of gradient change. As shown.

[0054]

[0055] in , These are the first-order and second-order momentum decay coefficients, respectively.

[0056] Due to initial momentum , The value is 0, indicating that the early iteration estimates were biased and need to be corrected.

[0057]

[0058] in Second-order momentum These are the corrected first and second order momentum. It represents the number of iterations.

[0059] Based on the corrected momentum, update the parameters:

[0060]

[0061] in It is the set learning rate. It is a constant that prevents division by zero.

[0062] By iteratively executing the gradient calculation, momentum update, and parameter adjustment processes described above, the loss function is minimized. To obtain the optimal weight parameters Once Adam's iterations converge to the optimal parameters, substituting them into the model yields the final step size estimate:

[0063]

[0064] The proposed Adam optimization step size estimation algorithm with multi-dimensional information constraints first generates an initial step size estimate for the current moment based on the acceleration extremum and the step size information of the previous three steps. Second, it extracts the nonlinear feature representation of walking frequency and variance from the acceleration sequence and performs zero-mean unit variance standardization on this feature set. Finally, it uses the parameters learned by the Adam optimization algorithm to perform weighted fusion of the initial step size estimate and the standardized nonlinear features of walking frequency and variance, and finally outputs the accurate step size estimate.

[0065] S4: Heading Angle Estimation. Heading angle estimation, as a core component of the PDR system, plays a decisive role in the analysis of trajectory direction and the reconstruction of spatial topology. However, current mainstream traditional heading angle estimation methods generally suffer from time-cumulative drift, where the measured values ​​continuously deviate from the true heading angle over time. This systematic deviation directly introduces irreversible cumulative errors into the trajectory extrapolation process, ultimately significantly degrading positioning accuracy. To address this issue, this paper proposes an H∞-constrained Sage-Husa adaptive filtering method that fuses gyroscope and magnetometer data for heading angle estimation, thereby improving the accuracy of the heading angle.

[0066] The output signals of different inertial measurement units (IMUs) naturally belong to the coordinate system of their respective attached devices. Therefore, before performing heading fusion calculations, a coordinate system transformation operation needs to be performed on the sensor data to achieve spatial alignment of multi-source sensor data in a unified reference coordinate system. Here, the quaternion method is chosen for coordinate system transformation, and the rotation formula based on quaternions is typically as follows:

[0067]

[0068] in This represents the rotation matrix from the IMU body coordinate system to the navigation coordinate system. and They are all quaternions.

[0069] The pedestrian heading angle is calculated using a gyroscope by integrating the measured angular velocity, expressed as:

[0070]

[0071] in It is the heading angle calculated by integrating the gyroscope. This is the data for the z-axis of the gyroscope. This indicates the heading angle at the previous time step.

[0072] The pedestrian heading angle is calculated using a magnetometer by integrating the measured angular velocity, expressed as:

[0073]

[0074] in It is the heading angle calculated using geomagnetism. , These are the geomagnetic x and y axis data after coordinate transformation.

[0075] Since the range of radian values ​​returned by arctan is Therefore, it needs to be adjusted to

[0076]

[0077] S5: H∞-constrained Sage-Husa adaptive filtering method for calculating fused heading. To address the gyroscope cumulative error problem, an H∞-constrained Sage-Husa adaptive filtering method is proposed to fuse gyroscope and magnetometer data for heading estimation. The Sage-Husa adaptive filtering utilizes the innovation sequence to estimate noise parameters online, solving the problem of time-varying noise modeling. Then, H∞ robust filtering is applied to suppress the worst-case disturbance effects through frequency domain energy constraints, enhancing system robustness.

[0078] In the heading fusion calculation process, the heading angle calculated by the gyroscope is set as the state vector and the heading angle calculated by the magnetometer is set as the measurement vector. Here is the state transition matrix. This is the measurement matrix. and These are the covariances of the state noise matrix and the measurement noise matrix, respectively.

[0079] State equation prediction:

[0080]

[0081] Error covariance prediction:

[0082]

[0083] When dealing with systems exhibiting time-varying noise characteristics, traditional Kalman filtering algorithms lack an online update mechanism for the statistical characteristics of time-varying system noise and measurement noise, often limiting their estimation accuracy. To optimize filtering performance and enhance the relative weight of innovation data in the filtering process to adapt to dynamic noise changes, an operator is introduced. ,satisfy:

[0084]

[0085] in It is a genetic factor, and .

[0086] Based on operators This is derived from adaptively updating measurement noise and process noise. Measurement noise covariance utilizes new information. The adaptive estimation update, the process noise covariance is obtained through the state residuals. Estimated update.

[0087]

[0088]

[0089] To address the shortcomings of Sage-Husa adaptive filtering, such as sensitivity to model mismatch and susceptibility to sudden disturbances, H∞ robust filtering is introduced to enhance robustness against model uncertainties and worst-case disturbances. H∞ filtering introduces a performance boundary γ, which adjusts the gain K... k Revised to:

[0090]

[0091] The error covariance update equation then incorporates a robust correction term:

[0092]

[0093] State equation update:

[0094]

[0095] S6: First, obtain the initial position of the target to be tested, then calculate the target step length and movement direction through S2, S3 and S5, and finally continuously estimate the position coordinates of the pedestrian.

[0096] Furthermore, to evaluate the performance of the proposed positioning method, this invention also presents a specific embodiment, conducting an experiment in a 92-meter-long rectangular corridor. In the figure, the blue line represents the pedestrian's trajectory, and the black arrows indicate the direction of movement. The pedestrian walks along the test path, and the integrated IMU automatically collects data from the inertial sensor.

[0097] To evaluate the performance of the step-size improvement model based on multi-dimensional information constraints, an example is experimentally compared with the Weinberg model, Scarlet model, Kim model, and Yan model. To evaluate the performance of the proposed heading angle estimation, an example is experimentally compared with gyroscope integration, direct geomagnetic measurement, and Sage-Husa fusion gyroscope and geomagnetism.

[0098] Figure 3 The comparison results of different step size models are shown. The results indicate that the proposed improved step size model has the smallest average step size error. Therefore, the proposed improved step size model is effective and calculates the step size with high accuracy.

[0099] Figure 4 The cumulative density function (CDF) of heading errors is shown for gyroscope integration, direct geomagnetic measurement, Sage-Husa fusion gyroscope and geomagnetic measurement, and the proposed algorithm. The results show that the proposed improved heading angle model performs best. Therefore, the proposed improved heading angle model is effective.

[0100] In addition, to verify the positioning performance of the present invention, multiple experiments were also conducted between the PDR algorithm and the method of the present invention.

[0101] Figure 5 The CDF (Corrective Distortion Factor) of the positioning error using the PDR algorithm and the improved PDR algorithm is shown in Table 1. The average error and root mean square error (RMSE) of the PDR algorithm and the improved PDR algorithm are also shown in Table 1. Experimental results show that the improved PDR algorithm improves both the average error and the RMS error compared to the traditional PDR algorithm. This indicates that the proposed improved algorithm improves positioning accuracy to a certain extent and reduces the impact of accumulated PDR error.

[0102] Table 1. Average error and mean squared error among different algorithms

[0103]

[0104] In summary, the present invention solves the following technical problems:

[0105] 1. To address the cumulative error problem in dead reckoning, this paper designs an improved step size model based on multi-dimensional information constraints. The model uses the maximum and minimum values ​​of acceleration and the step size of the previous three steps to obtain a rough estimate of the step size at the current moment. Then, it extracts the nonlinear expressions of the walking frequency and variance from the acceleration data and performs zero-mean unit variance normalization on them. Finally, it uses Adam parameter optimization to perform a weighted fusion of the rough step size value and the normalized nonlinear expressions of the walking frequency and variance to obtain a refined estimate.

[0106] 2. To mitigate the heading estimation offset error caused by gyroscope drift over time, this invention proposes an H∞-constrained Sage-Husa adaptive filtering method. The heading angle calculated by the gyroscope is used as the state vector and the heading angle calculated by the magnetometer is used as the measurement vector. Then, the final fused heading estimate is performed using Sage-Husa adaptive filtering. Furthermore, to suppress the impact of external disturbances on system performance, H∞ constraints are used to minimize the maximum gain of the disturbance signal to the controlled output, ensuring that the system can maintain stable tracking performance under disturbances, thereby obtaining a better heading angle.

[0107] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A multi-dimensional information constrained adaptive dead reckoning positioning method, characterized in that, The method comprises the following steps: Step 1: collect inertial sensor data, and pre-process the data; Step 2: count steps by using a peak-valley pair constraint peak detection method; Step 3: extract feature step frequency and variance from pre-processed acceleration data; Step 4: fuse multi-dimensional information and rough step length estimation to calculate fine step length by using Adam optimization parameters; Step 5: estimate heading by directly measuring pre-processed gyroscope data integration and magnetometer data; Step 6: calculate the finally fused heading angle by using H∞ constraint Sage-Husa adaptive filtering; Step 7: perform final positioning estimation and predict the target position by using the formula of dead reckoning.

2. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 1, the first few seconds of raw data are removed, then the accelerometer and gyroscope data are subjected to sliding mean filtering, and the geomagnetic data are subjected to empirical mode decomposition for noise reduction.

3. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 2, different acceleration thresholds are set according to individual differences, and when the detection value triggers the threshold, the candidate peak / trough is alternately marked, and the adjacent marks need to meet the preset time interval threshold to be determined as valid motion peak / trough.

4. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 3, the nonlinear expression of walking frequency and variance is extracted from the acceleration data, and zero-mean unit-variance normalization processing is performed.

5. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 4, the maximum value and minimum value of acceleration and the step length of the previous three steps are used to obtain a rough estimation value of the step length at the current time, and then Adam parameter optimization is used to perform weighted fusion of the rough step length value and the normalized nonlinear expression of walking frequency and variance, thereby obtaining a fine estimation value of the step length.

6. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 5, the pre-processed gyroscope and magnetometer data are unified to the same coordinate system by using the quaternion method, then the gyroscope data are directly integrated to calculate the heading, and the magnetometer data are geometrically calculated to obtain the heading.

7. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 6, the gyroscope calculated heading angle is used as the state vector and the magnetometer calculated heading angle is used as the measurement vector, then Sage-Husa adaptive filtering is used for final fusion of the heading estimation, and in order to suppress the influence of external disturbance on the system performance, H∞ constraint is used to minimize the maximum gain of the disturbance signal to the controlled output, so that the system can maintain stable tracking performance under disturbance, thereby obtaining a better heading angle.

8. The multi-dimensional information-constrained adaptive dead-reckoning localization method of claim 1, wherein, In the execution process of step 7, the initial position of the target to be measured is obtained, then the target step length and movement direction are calculated by using steps 2, 4 and 6, and finally the position coordinates of the pedestrian are continuously estimated.

Images