A foot-bound inertial navigation zero speed detection threshold adaptive method

Abstract

The application discloses a foot binding type inertial navigation zero speed detection threshold self-adaptive method and relates to the technical field of pedestrian inertial navigation positioning. The application utilizes a Gaussian-Cauchy hybrid correlation entropy to construct a robust motion classification model, adopts a whale optimization algorithm with a nonlinear convergence factor and an adaptive inertia weight to jointly optimize model parameters and classification decision thresholds, obtains a stable motion category sequence by combining a sliding window majority voting and time series smoothing, and accordingly adaptively configures a zero speed detection threshold. The application can realize high-precision foot binding type inertial navigation positioning under complex scenes such as multi-gait switching, heavy tail noise and abnormal disturbance, reduces the influence of motion misclassification and mismatched fixed thresholds on zero speed detection, improves the stability of zero speed detection and positioning precision, and guarantees the autonomous navigation ability of pedestrians in a satellite signal-free environment.

Description

Technical Field

[0001] This invention relates to the field of navigation and positioning technology, and specifically to an adaptive method and system for zero-velocity detection threshold of foot-strap inertial navigation based on robust motion classification using GCM-LSSVM. Background Technology

[0002] In applications such as emergency rescue, public safety, mine tunnels, and military operations, Global Navigation Satellite Systems (GNSS) are prone to problems such as signal blockage, multipath interference, and signal loss, making it impossible to continuously output reliable positioning information. While indoor positioning technologies such as Wi-Fi, Bluetooth, Ultra-Wideband (UWB), and visual positioning can supplement positioning capabilities, they generally rely on external infrastructure, prior environmental information, or stable lighting conditions, resulting in drawbacks such as high deployment costs, limited coverage, and weak autonomous operation capabilities.

[0003] Foot-mounted inertial navigation systems rely on micro inertial measurement units (MIMUs) mounted on the pedestrian's feet for autonomous positioning, eliminating the need for external equipment and making them a mainstream technology for pedestrian navigation in scenarios without satellite signals. Zero-velocity detection and correction are the core components of foot-mounted inertial navigation: during walking, running, or climbing stairs, the feet periodically enter a stationary state. By identifying these zero-velocity states and combining them with Kalman filtering for error correction, drift of the inertial components can be effectively suppressed, improving positioning accuracy.

[0004] Current technologies have several shortcomings: Chinese patent CN115950426A discloses a capacitive Kalman filter attitude estimation method based on motion acceleration compensation. This method estimates attitude by determining the motion acceleration fitting interval and performing predictive compensation, combined with capacitive Kalman filtering, which can reduce the impact of motion acceleration on attitude estimation to some extent. However, this method mainly focuses on improving the accuracy of attitude estimation and does not establish robust motion classification models for multi-gait scenarios such as walking, running, climbing stairs, and gait switching. It also does not adaptively adjust the zero-velocity detection threshold according to the motion category, making it difficult to solve the problem of dynamic adaptation of the zero-velocity threshold under complex motion states.

[0005] Chinese patent CN116734858A discloses an indoor pedestrian navigation method and system based on behavioral probability analysis. This method extracts the highest probability zero-velocity point through gait detection and combines it with stair step probability detection for height correction, thereby improving the accuracy of indoor pedestrian navigation. However, this method mainly relies on probability zero-velocity point search and height-assisted correction, and does not robustly identify multi-gait motion states at the classification model level, nor does it jointly optimize classifier parameters, kernel function parameters, and classification decision thresholds.

[0006] Currently, most mainstream zero-velocity detection solutions use fixed detection thresholds, which cannot adapt to the differences in foot movement characteristics under different movement states of pedestrians. At the same time, micro inertial measurement units are prone to heavy tail noise and impact disturbances during operation. Combined with the complex working conditions of frequent gait switching, they are very likely to cause misclassification of movement state and failure of zero-velocity detection, ultimately leading to the continuous accumulation of navigation and positioning errors. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies, such as poor adaptability of fixed zero velocity detection threshold, easy misclassification caused by complex noise and gait switching, and low positioning accuracy. It provides a zero velocity detection threshold adaptive method and system based on GCM-LSSVM for foot-strap inertial navigation, which improves motion classification stability, zero velocity detection reliability and pedestrian navigation and positioning accuracy in scenarios with heavy tail noise, abnormal disturbances and multi-gait switching.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an adaptive method and system for zero-velocity detection threshold of foot-attached inertial navigation based on GCM-LSSVM, comprising: Step 1, data acquisition and sample construction: using a micro inertial measurement unit fixed to the foot of a pedestrian to acquire triaxial acceleration and triaxial angular velocity data, and constructing motion classification input samples according to a preset time window.

[0009] Step 2, GCM-LSSVM Robust Motion Classification: A GCM-LSSVM robust motion classification model is constructed based on Gaussian-Cauchy mixed correlation entropy. The model is iteratively solved under a semi-quadratic optimization framework combined with a residual threshold segmented weighting mechanism to obtain the motion type classification results.

[0010] Step 3: Joint optimization of parameters and decision thresholds: The whale optimization algorithm, which introduces a nonlinear convergence factor and adaptive inertia weight, is used to jointly optimize the free parameters and classification decision thresholds of the GCM-LSSVM model to obtain the optimal parameter combination and the optimal decision threshold.

[0011] Step 4: Motion Category Stabilization and Adaptive Zero-Speed ​​Detection: A stable motion category sequence is obtained through sliding window majority voting and temporal smoothing. Based on the stable motion category sequence, a zero-speed detection threshold is adaptively configured, and a zero-speed state decision is made according to the zero-speed detection threshold. Zero-speed correction and navigation state update are completed under the error state Kalman filter framework.

[0012] Preferably, the specific process of data acquisition and sample construction is as follows: S11, collect continuous inertial data during walking, running, climbing stairs and gait switching, stitch together the six-channel data of accelerometer and gyroscope, and slice according to the preset window length and sliding step length.

[0013] S12, the window samples obtained from the slice are normalized and simulated by random rotation to form input samples for robust motion classification.

[0014] Preferably, the construction and solution of the GCM-LSSVM robust motion classification model specifically includes the following steps: S21, converting the Gaussian kernel related entropy term... Entropy terms related to the Cauchy kernel The weighted mixture is used as the loss function, and the optimization problem is: ;in, For regularization parameters, Indicates the number of training samples. Represents the weight vector. For bias terms, To indicate the first Individual sample classification residuals, For kernel bandwidth parameters, Indicates the mixed weighting coefficient. It is the Cauchy nucleus heavy-tail factor.

[0015] S22, in the In the next iteration, based on the current classification residual Calculate local curvature weights and constant compensation terms Based on this, a weighted least squares support vector machine subproblem is constructed, and the bias term is solved by combining the KKT conditions and kernel tricks. and Lagrange multiplier vectors And construct the classification decision function for GCM-LSSVM: ;in, The selected kernel function.

[0016] S23, when the sample residuals Greater than the residual threshold At that time, the corresponding local curvature weights and compensation terms are multiplied by the attenuation factor. When the sample residuals Not greater than the residual threshold At the same time, the original weights and compensation terms are retained to reduce the impact of outliers on the classification boundary.

[0017] Preferably, the joint optimization of the free parameters and classification decision threshold of the GCM-LSSVM classifier is carried out as follows: S31, introducing a nonlinear convergence factor. This allows the algorithm to maintain strong global exploration capabilities in the early stages of iteration and accelerate local convergence in the later stages. Adaptive inertia weights are introduced. The search stride is dynamically adjusted to establish a dynamic balance between global exploration and local development; among which... This represents the current iteration number. This represents the maximum number of iterations.

[0018] S32, the model parameters for GCM-LSSVM based on the whale optimization algorithm with the introduction of nonlinear convergence factors and adaptive inertia weights include: regularization parameters. Mixed weighting coefficients Core bandwidth parameters and Cauchy nucleus heavy-tail factor Conduct joint optimization.

[0019] S33. For each set of candidate parameter vectors, train the GCM-LSSVM model based on steps S21 to S23 and obtain the classification output on the validation set; take the decision threshold corresponding to the highest classification accuracy on the validation set as the optimal decision threshold, and use the validation accuracy under this threshold as the fitness function value for iterative optimization.

[0020] Preferably, the process of obtaining a stable motion category sequence through sliding window majority voting and temporal smoothing is as follows: S41, the original predicted label sequence output by the motion classification model is processed... Construct a local sliding window time-by-time: ;in For the window radius, The length of the original predicted label sequence; Indicates the first The local sliding window corresponding to each moment; This indicates the label number within the window.

[0021] S42, Statistics for each sports category In the window Number of occurrences within The category that appears most frequently will be used as the first category. Moment smoothing labels: ;in, This represents the total number of sports categories. If multiple categories have the same maximum frequency, the original label at the current time is retained. This is to avoid over-correction when statistical evidence is insufficient.

[0022] Preferably, the motion type corresponding to the current IMU window is determined based on the smoothed motion category, and the corresponding SHOE detection threshold is determined by the mapping relationship between the preset motion category and the zero-speed detection threshold; a SHOE test statistic is constructed using the specific force and angular velocity within the current window, and when the test statistic is less than the detection threshold, it is determined to be a zero-speed state; otherwise, it is determined to be a non-zero-speed state; in the zero-speed state, the navigation solution speed is used as the zero-speed pseudo-measurement input error state Kalman filter for measurement update, and in the non-zero-speed state, only time update is performed.

[0023] Secondly, the present invention provides a foot-attached inertial navigation system that performs the above-described method, including a motion information acquisition module, a sample construction module, a robust motion classification and optimization module, a classification stability and zero-velocity detection module, and a zero-velocity correction and navigation solution module.

[0024] The motion information acquisition module uses a micro-inertial measurement unit fixed to the pedestrian's feet to collect triaxial acceleration and triaxial angular velocity data; the sample construction module performs windowing processing on the raw inertial data according to a preset time window; the robust motion classification and optimization module trains the GCM-LSSVM robust motion classification model and uses an improved whale optimization algorithm to jointly optimize model parameters and decision thresholds; the classification stability and zero-velocity detection module performs temporal smoothing on the motion classification results and adaptively configures the SHOE detection threshold according to the motion category; the zero-velocity correction and navigation solution module performs error state Kalman filter updates and outputs the pedestrian's position, velocity, and attitude information.

[0025] The robust motion classification and optimization module uses radial basis functions as the kernel function of the GCM-LSSVM model.

[0026] Preferably, it also includes a data storage module for storing the original data of the micro inertial measurement unit, model parameters, motion classification results, zero-velocity detection results, and navigation solution data.

[0027] Preferably, it also includes a barometer module, which is communicatively connected to the zero-speed correction and navigation calculation module, for collecting barometer data and assisting in correcting pedestrian height information.

[0028] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. The present invention uses Gaussian-Cauchy mixed correlation entropy to construct the GCM-LSSVM model, and combines the residual threshold segmented weighting mechanism to give full play to the local fitting advantage of Gaussian kernel and the heavy-tailed noise suppression advantage of Cauchy kernel. It can effectively suppress the heavy-tailed noise and impact disturbance generated by the micro inertial measurement unit, reduce the influence of outlier samples on motion classification results, and improve the robustness of the model.

[0029] 2. This invention improves the whale optimization algorithm by introducing a nonlinear convergence factor and adaptive inertia weight to balance the algorithm's global exploration and local development capabilities. At the same time, it achieves joint optimization of model parameters and classification decision thresholds, which greatly improves the model's classification accuracy and the rationality of the decision thresholds, and avoids the performance loss caused by independent optimization of parameters and thresholds in traditional methods.

[0030] 3. This invention suppresses the high-frequency jitter and frequent switching of classification results by using sliding window majority voting and temporal smoothing to process the motion label sequence, thus ensuring the stability of the motion category output. It abandons the traditional fixed threshold scheme and adaptively configures the zero-speed detection threshold according to the motion category, perfectly adapting to various working conditions such as walking, running, climbing stairs, and gait switching, thus solving the problem of high false detection rate of fixed threshold under different motion states.

[0031] 4. The overall solution of this invention realizes a closed loop of the entire process of data acquisition, sample construction, motion classification, parameter optimization, adaptive zero velocity detection, and navigation calculation. In complex scenarios without satellite signals, it effectively reduces the accumulation of navigation errors and significantly improves the positioning accuracy and environmental adaptability of the foot-mounted inertial navigation system. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0033] Figure 2 The indoor corridor site plan and reference trajectory diagram corresponding to the experimental dataset of this invention.

[0034] Figure 3 The image shows a comparison of navigation experimental trajectories (from top to bottom) on a mixed motion dataset of corridor walking and running, using the method proposed in this invention, the standard posture assumption optimal estimation method, the standard support vector machine method, and the existing support vector machine method based on maximum correlation entropy. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figure 1 As shown, in the first aspect, the present invention provides an adaptive method for zero-velocity detection threshold of foot-attached inertial navigation based on GCM-LSSVM, which specifically includes the following steps: Step 1, data acquisition and sample construction: using a micro inertial measurement unit fixed to the foot of a pedestrian to acquire triaxial acceleration and triaxial angular velocity data, and construct motion classification input samples according to a preset time window.

[0037] In a specific embodiment, step one, data acquisition and sample construction, is specifically implemented as follows: S11, continuous inertial data during walking, running, climbing stairs and gait switching are collected, the six-channel data from the accelerometer and gyroscope are stitched together, and sliced ​​according to the preset window length and sliding step length.

[0038] S12, the window samples obtained from the slice are normalized and simulated by random rotation to form input samples for robust motion classification.

[0039] Step 2, GCM-LSSVM Robust Motion Classification: A GCM-LSSVM robust motion classification model is constructed based on Gaussian-Cauchy mixed correlation entropy. The model is iteratively solved under a semi-quadratic optimization framework combined with a residual threshold segmented weighting mechanism to obtain the motion type classification results.

[0040] In a specific embodiment, step two, GCM-LSSVM robust motion classification, is specifically implemented as follows: S21, the Gaussian kernel related entropy term... Entropy terms related to the Cauchy kernel The weighted mixture is used as the loss function, and the optimization problem is: ;in, For regularization parameters, Indicates the number of training samples. Represents the weight vector. For bias terms, To indicate the first Individual sample classification residuals, For kernel bandwidth parameters, Indicates the mixed weighting coefficient. It is the Cauchy nucleus heavy-tail factor.

[0041] S22, in the In the next iteration, based on the current classification residual Calculate local curvature weights and constant compensation terms Based on this, a weighted least squares support vector machine subproblem is constructed, and the bias term is solved by combining the KKT conditions and kernel tricks. and Lagrange multiplier vectors And construct the classification decision function for GCM-LSSVM: ;in, The selected kernel function.

[0042] S23, when the sample residuals Greater than the residual threshold At that time, the corresponding local curvature weights and compensation terms are multiplied by the attenuation factor. When the sample residuals Not greater than the residual threshold At the same time, the original weights and compensation terms are retained to reduce the impact of outliers on the classification boundary.

[0043] Step 3: Joint optimization of parameters and decision thresholds: The whale optimization algorithm, which introduces a nonlinear convergence factor and adaptive inertia weight, is used to jointly optimize the free parameters and classification decision thresholds of the GCM-LSSVM model to obtain the optimal parameter combination and the optimal decision threshold.

[0044] In one specific embodiment, step three, joint optimization of parameters and decision threshold, is specifically implemented as follows: S31, introducing a nonlinear convergence factor. This allows the algorithm to maintain strong global exploration capabilities in the early stages of iteration and accelerate local convergence in the later stages. Adaptive inertia weights are introduced. The search stride is dynamically adjusted to establish a dynamic balance between global exploration and local development; among which... This represents the current iteration number. This represents the maximum number of iterations.

[0045] S32, the model parameters for GCM-LSSVM based on the whale optimization algorithm with the introduction of nonlinear convergence factors and adaptive inertia weights include: regularization parameters. Mixed weighting coefficients Core bandwidth parameters and Cauchy nucleus heavy-tail factor Conduct joint optimization.

[0046] S33. For each set of candidate parameter vectors, train the GCM-LSSVM model based on steps S21 to S23 and obtain the classification output on the validation set; take the decision threshold corresponding to the highest classification accuracy on the validation set as the optimal decision threshold, and use the validation accuracy under this threshold as the fitness function value for iterative optimization.

[0047] Step 4, Motion Category Stabilization and Adaptive Zero Velocity Detection: A stable motion category sequence is obtained through sliding window majority voting and temporal smoothing, and then zero velocity state decision is made. Zero velocity correction and navigation state update are completed under the error state Kalman filter framework.

[0048] In one specific embodiment, step four, motion category stabilization and adaptive zero-velocity detection, is specifically implemented as follows: S41, the original predicted label sequence output by the motion classification model is processed. Construct a local sliding window time-by-time: ;in For the window radius, The length of the original predicted label sequence; Indicates the first The local sliding window corresponding to each moment; This indicates the label number within the window.

[0049] S42, Statistics for each sports category In the window Number of occurrences within The category that appears most frequently will be used as the first category. Moment smoothing labels: ;in, This represents the total number of sports categories. If multiple categories have the same maximum frequency, the original label at the current time is retained. This is to avoid over-correction when statistical evidence is insufficient.

[0050] The motion type corresponding to the current IMU window is determined based on the smoothed motion category, and the corresponding SHOE detection threshold is determined by the mapping relationship between the preset motion category and the zero-velocity detection threshold. The SHOE test statistic is constructed using the specific force and angular velocity within the current window. When the test statistic is less than the detection threshold, it is determined to be a zero-velocity state; otherwise, it is determined to be a non-zero-velocity state. In the zero-velocity state, the navigation solution velocity is used as the zero-velocity pseudo-measurement input error state Kalman filter for measurement update. In the non-zero-velocity state, only time update is performed.

[0051] Secondly, the present invention provides a foot-attached inertial navigation system that performs the above-described method, including a motion information acquisition module, a sample construction module, a robust motion classification and optimization module, a classification stability and zero-velocity detection module, and a zero-velocity correction and navigation solution module.

[0052] The motion information acquisition module uses a micro-inertial measurement unit fixed to the pedestrian's feet to collect triaxial acceleration and triaxial angular velocity data; the sample construction module performs windowing processing on the raw inertial data according to a preset time window; the robust motion classification and optimization module trains the GCM-LSSVM robust motion classification model and uses an improved whale optimization algorithm to jointly optimize model parameters and decision thresholds; the classification stability and zero-velocity detection module performs temporal smoothing on the motion classification results and adaptively configures the SHOE detection threshold according to the motion category; the zero-velocity correction and navigation solution module performs error state Kalman filter updates and outputs the pedestrian's position, velocity, and attitude information.

[0053] This invention proposes an adaptive zero-velocity detection threshold method for foot-mounted inertial navigation systems based on GCM-LSSVM. This method achieves stable motion state recognition and reliable zero-velocity detection under complex conditions such as multi-gait switching, heavy-tailed noise, and abnormal disturbances. The method improves the robustness and parameter optimization ability of the motion classification model through Gaussian-Cauchy hybrid correlation entropy, residual threshold segmentation weighting mechanism, and improved whale optimization algorithm. Furthermore, it combines motion category temporal smoothing with SHOE zero-velocity detection threshold adaptive configuration to complete zero-velocity correction and navigation state update within the error state Kalman filter framework. This effectively improves the positioning accuracy, stability, and applicability of foot-mounted inertial navigation systems in complex motion environments.

[0054] Inertial data collected by a foot-based micro-inertial measurement unit was used to verify the effectiveness of the proposed adaptive zero-velocity detection threshold method for foot-strap inertial navigation based on GCM-LSSVM. The inertial data included triaxial acceleration and triaxial angular velocity data, and the motion states included typical gaits such as walking on flat ground, running, and climbing stairs, to verify the applicability of the invention under conditions of multi-gait switching, sensor noise, and abnormal disturbances.

[0055] Free parameters of the GCM-LSSVM model in this invention A joint optimization was performed using a whale optimization algorithm that incorporates a nonlinear convergence factor and adaptive weights. The classification decision threshold was determined by searching within the 10%–90% quantile range of the continuous output results on the validation set. The minimum search range of the free parameters was set to [value missing]. And set the maximum search range of the free parameters to The whale population size is set to 20, and the maximum number of iterations is set to 30; the penalty factor in the semi-quadratic optimization is set to a maximum of 10 iterations, and the convergence tolerance is set to... The window length for sliding majority voting is set to 5.

[0056] The initial motion category sequence is obtained based on the optimized GCM-LSSVM model, and high-frequency jitter and unreasonable frequent switching in the category sequence are suppressed by sliding window majority voting and temporal smoothing. Subsequently, the SHOE zero-velocity detection threshold is adaptively configured according to the smoothed motion category, and zero-velocity detection, zero-velocity correction, and navigation state update are completed under the error state Kalman filter framework.

[0057] The following further verifies the effectiveness of the present invention: The average root mean square error (ARMSE) is used to calculate the positioning error between the method of the present invention and the comparison method. The comparison method includes standard attitude assumption optimal estimation, standard support vector machine, existing multi-kernel correlation entropy support vector machine, and the method proposed in this invention. The formula for calculating the average root mean square error of position is: In the formula, and The first The true and estimated locations of each sampling point This represents the number of samples.

[0058] Please see Figure 2 , Figure 2This is a navigation experiment trajectory diagram based on a mixed walking and running motion dataset in a corridor, according to the present invention. The diagram shows the experimental path from top to bottom, where red dots represent reference trajectory sampling points, and star-shaped markers indicate the start and end points. The dataset was collected from five test subjects. In each trial, the subjects traversed three corridors, walked approximately 110m, then turned around and returned to the starting point along the original path. Ground conditions were obtained by recording the positions of markers on the corridor floor. When a subject passed a marker, they pressed a handheld trigger, thus comparing the estimated navigation position at that moment with the known marker position. The experiment compared four zero-velocity detection methods, including the method of the present invention, and evaluated the positioning performance using the navigation trajectory and root mean square error of the position.

[0059] Please see Figure 3 , Figure 3 Different zero-velocity detection or motion classification methods are shown in Figure 2 The figure shows a comparison of navigation trajectories in an indoor environment. Using a reference trajectory as a benchmark, this figure compares the navigation trajectories of the optimal estimation of pose assumptions with a fixed threshold standard, the standard support vector machine method, the existing support vector machine method based on maximum correlation entropy, and the method proposed in this invention. This comparison illustrates the trajectory following effect and positioning stability of this invention in scenarios such as corridor turns and switching between walking and running.

[0060] The above description is only for illustrating the concept and specific implementation of the present invention. Those skilled in the art can modify, supplement, or substitute the specific embodiments in a similar manner without departing from the concept or scope of the claims, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. An adaptive method for zero-velocity detection threshold in foot-strap inertial navigation, applied to a pedestrian foot-strap inertial navigation system, characterized in that, include: Step 1: Data Acquisition and Sample Construction: Collect triaxial acceleration and triaxial angular velocity data using a micro-inertial measurement unit fixed to the pedestrian's foot, and construct motion classification input samples according to a preset time window; Step 2, GCM-LSSVM Robust Motion Classification: Based on Gaussian-Cauchy hybrid correlation entropy, a GCM-LSSVM robust motion classification model is constructed. Under the semi-quadratic optimization framework, the model is iteratively solved by combining residual threshold piecewise weighting mechanism to obtain motion type classification results. Step 3: Joint optimization of parameters and decision thresholds: The whale optimization algorithm, which introduces a nonlinear convergence factor and adaptive inertia weight, is used to jointly optimize the free parameters and classification decision thresholds of the GCM-LSSVM robust motion classification model to obtain the optimal parameter combination and the optimal decision threshold. Step 4, Motion Category Stabilization and Adaptive Zero Speed ​​Detection: A stable motion category sequence is obtained through sliding window majority voting and temporal smoothing. Based on the stable motion category sequence, a zero speed detection threshold is adaptively configured, and a zero speed state decision is made according to the zero speed detection threshold. Zero speed correction and navigation state update are completed under the error state Kalman filter framework. Step two, GCM-LSSVM robust motion classification, specifically includes: S21, the Gaussian kernel related entropy term Entropy terms related to the Cauchy kernel The weighted mixture is used as the loss function, and the optimization problem is: ； in, For regularization parameters, Indicates the number of training samples. Represents the weight vector. For bias terms, To indicate the first Individual sample classification residuals, For kernel bandwidth parameters, Indicates the mixed weighting coefficient. The Cauchy nucleus heavy-tail factor; S22, in the In the next iteration, based on the current classification residual Calculate local curvature weights and constant compensation terms Based on this, a weighted least squares support vector machine subproblem is constructed, and the bias term is solved by combining the KKT conditions and kernel tricks. and Lagrange multiplier vectors And construct the classification decision function for GCM-LSSVM: ； in, The selected kernel function; S23, when the sample residuals Greater than the residual threshold At that time, the corresponding local curvature weights and compensation terms are multiplied by the attenuation factor. When the sample residuals Not greater than the residual threshold At the same time, the original weights and compensation terms are retained to reduce the impact of outliers on the classification boundary.

2. The adaptive method for zero-velocity detection threshold of foot-strap inertial navigation according to claim 1, characterized in that, Step one, data acquisition and sample construction, specifically includes: S11 collects continuous inertial data during walking, running, climbing stairs and gait switching, stitches together the six-channel data from the accelerometer and gyroscope, and slices them according to the preset window length and sliding step length. S12, the window samples obtained from the slice are normalized and simulated by random rotation to form input samples for robust motion classification.

3. The adaptive method for zero-velocity detection threshold of foot-strap inertial navigation according to claim 1, characterized in that, Step three, the joint optimization of parameters and decision thresholds, specifically includes: S31, Introducing a nonlinear convergence factor This allows the algorithm to maintain strong global exploration capabilities in the early stages of iteration and accelerate local convergence in the later stages of iteration by introducing adaptive inertia weights. The search stride is dynamically adjusted to establish a dynamic balance between global exploration and local development; among which... This represents the current iteration number. This represents the maximum number of iterations. S32, the model parameters for GCM-LSSVM based on the whale optimization algorithm with the introduction of nonlinear convergence factors and adaptive inertia weights include: regularization parameters. Mixed weighting coefficients Core bandwidth parameters and Cauchy nucleus heavy-tail factor Conduct joint optimization; S33. For each set of candidate parameter vectors, train the GCM-LSSVM robust motion classification model based on steps S21 to S23 and obtain the classification output on the validation set; take the decision threshold corresponding to the highest classification accuracy on the validation set as the optimal decision threshold, and use the validation accuracy under this threshold as the fitness function value for iterative optimization.

4. The adaptive method for zero-velocity detection threshold of foot-strap inertial navigation according to claim 1, characterized in that, Step four, obtaining a stable motion category sequence through sliding window majority voting and temporal smoothing, specifically includes: S41, the original predicted label sequence output by the motion classification model. Construct a local sliding window time-by-time: ； in For the window radius, The length of the original predicted label sequence; Indicates the first The local sliding window corresponding to each moment; Indicates the label number within the window; S42, Statistics for each sports category , The window displays the total number of sports categories. Number of occurrences within The category that appears most frequently will be used as the first category. Moment smoothing labels: ； in, This represents the total number of sports categories. If multiple categories have the same maximum frequency, the original label at the current time is retained. This is to avoid over-correction when statistical evidence is insufficient.

5. The adaptive method for zero-velocity detection threshold of foot-strap inertial navigation according to claim 4, characterized in that, Step four, adaptively configuring the zero-speed detection threshold, includes: The motion type corresponding to the current IMU window is determined based on the smoothed motion category, and the corresponding SHOE detection threshold is determined by the mapping relationship between the preset motion category and the zero-velocity detection threshold. The SHOE test statistic is constructed using the specific force and angular velocity within the current window. When the test statistic is less than the detection threshold, it is determined to be a zero-velocity state; otherwise, it is determined to be a non-zero-velocity state. In the zero-velocity state, the navigation solution velocity is used as the zero-velocity pseudo-measurement input error state Kalman filter for measurement update. In the non-zero-velocity state, only time update is performed.

6. A foot-strap inertial navigation system for executing the foot-strap inertial navigation zero-velocity detection threshold adaptive method according to any one of claims 1 to 5, characterized in that, include: The motion information acquisition module is used to acquire triaxial acceleration and triaxial angular velocity data using a micro-inertial measurement unit fixed to the pedestrian's feet; The sample construction module is used to perform windowing processing on the raw inertial data according to a preset time window to construct motion classification input samples; The robust motion classification and optimization module is used to train the GCM-LSSVM robust motion classification model based on Gauss-Cauchy mixed correlation entropy and semi-quadratic optimization method, and to jointly optimize the model's free parameters and classification decision threshold by introducing a nonlinear convergence factor and adaptive inertia weight. The classification stability and zero-velocity detection module is used to perform sliding window majority voting and temporal smoothing on the motion classification results to obtain a stable motion category sequence, and adaptively configure the SHOE detection threshold according to the mapping relationship between the preset motion category and the zero-velocity detection threshold to complete the zero-velocity state decision. The zero-velocity correction and navigation calculation module is used to perform error state Kalman filter update based on the zero-velocity state decision result, and combine inertial data to complete navigation calculation and error feedback correction, and output the corrected pedestrian position, velocity and attitude information.

7. The foot-attached inertial navigation system according to claim 6, characterized in that, The robust motion classification and optimization module uses radial basis functions as the kernel function of the GCM-LSSVM robust motion classification model.

8. The foot-strap inertial navigation system according to claim 6, characterized in that, It also includes a data storage module for storing the raw data of the micro inertial measurement unit, model parameters, motion classification results, zero-velocity detection results, and navigation solution data.

9. The foot-strap inertial navigation system according to claim 6, characterized in that, It also includes a barometer module, which is communicatively connected to the zero-speed correction and navigation calculation module, and is used to collect barometric pressure data and assist in correcting pedestrian height information.

Images