Multi-modal identification pedestrian positioning navigation method

Abstract

The invention relates to the technical field of positioning and navigation, in particular to a multi-modal recognition pedestrian positioning and navigation method. The method comprises the following steps: firstly, respectively acquiring data in six motion modes of walking reading, walking telephone calling, walking swinging, walking pocket, jogging telephone calling and jogging pocket, and extracting 30-dimensional motion features; thirdly, constructing a multi-mode recognition model of a PCA-GWO-KELM (principal component analysis and grey wolf algorithm optimization kernel extreme learning machine) based on the PCA-GWO-KELM; inputting the motion features into a multi-mode recognition model for training to obtain recognition accuracy rates under different motion modes; then, step number detection is carried out by combining a peak value detection method with a self-adaptive acceleration upper threshold value and a self-adaptive dynamic time threshold value, the step length is calculated by adopting a multi-feature-based self-adaptive step length estimation algorithm, and the course is calculated by adopting a method of fusing direction sensor data and gyroscope 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] The invention discloses a pedestrian positioning and navigation method of multi-modal recognition, belonging to the field of positioning and navigation. Background Art

[0002] As human society develops and progresses, people's demand for location-based services (LBS) is increasing day by day. The construction of the Global Navigation Satellite System (GNSS) has enabled high-precision outdoor location services. However, due to the obstruction of indoor buildings, satellite navigation signals are difficult to reach indoors, or the satellite navigation signals that reach indoors are extremely weak, and high-precision indoor location services cannot be provided, which has led to the research and development of indoor positioning technology.

[0003] At present, the commonly used indoor positioning technologies include Wi-Fi positioning, ultrasonic positioning, Bluetooth positioning, inertial navigation positioning, radio frequency identification positioning, and ultra-wideband positioning. Each technology has its own advantages, but also has certain shortcomings. The indoor environment is changeable and complex, resulting in large signal loss during transmission and being more susceptible to multipath propagation. Therefore, the application of Wi-Fi positioning technology, Bluetooth positioning technology, and ultrasonic positioning technology for indoor positioning is affected to a certain extent, while radio frequency identification positioning technology and ultra-wideband positioning technology require the deployment of additional equipment. Among them, the inertial navigation system is a relative positioning technology that does not rely on infrastructure, and pedestrian dead reckoning (PDR) is a type of inertial navigation. It is a method that relies on mobile smart terminal sensors to calculate the movement trajectory of pedestrians, and has extremely strong autonomy and continuity.

[0004] Existing dead reckoning positioning methods often restrict the posture of the mobile phone, and always keep it pointing stably in the direction of the pedestrian's movement during the positioning process. However, people can use mobile phones in any posture, such as reading information, making and receiving calls on the mobile phone while walking, swinging the mobile phone casually or placing it in the pocket, etc. Therefore, this restriction on the posture of the mobile phone obviously cannot meet the positioning needs of the general public. In addition, pedestrians in a certain navigation mode often involve different movement modes, which will also introduce large cumulative errors. According to people's daily life and the way they use mobile phones, pedestrian movement modes can be roughly divided into two types: walking and jogging. At the same time, the postures of pedestrians holding mobile phones include: reading mode, calling mode, swinging mode and pocket mode. Different pedestrian movement modes will cause large cumulative errors in step frequency, step length and heading angle, which will affect the accuracy of the positioning results. IMU sensors can not only analyze and calculate the current location information of pedestrians, but also perceive different movement mode information. Therefore, using pedestrian movement pattern recognition technology to assist dead reckoning helps to improve the impact of IMU data under different pedestrian movement states and different mobile phone holding modes, thereby obtaining stable and convenient indoor location services at a lower cost. Summary of the invention

[0005] The present invention proposes a pedestrian positioning and navigation method based on multimodal recognition, which does not require any additional infrastructure and can achieve low-cost, device-compatible and flexible positioning in different modes.

[0006] To achieve the above objectives, the present invention provides a pedestrian positioning and navigation method based on multimodal recognition. The method comprises the following steps:

[0007] Step 1: Record data under different motion modes and pre-process the data.

[0008] Step 2: Data feature extraction under different motion modes.

[0009] Step 3: Construct a multi-pattern recognition model based on principal component analysis and grey wolf algorithm optimized kernel extreme learning machine (PCA-GWO-KELM).

[0010] Step 4: Count the number of steps using the peak detection method with dual adaptive thresholds.

[0011] Step 5: Calculate the step size using a multi-feature based adaptive step size estimation algorithm.

[0012] Step 6: Calculate the heading angle by fusing the direction sensor data and the gyroscope data.

[0013] Step 7: Perform dead reckoning positioning estimation under different motion modes.

[0014] Optionally, during the execution of step 1, the data of the first few seconds and the last few seconds of the original data are first removed, and then a moving mean filtering process is performed on the original data.

[0015] Optionally, during the execution of step 2, a plurality of data features that can reflect the characteristics of different pedestrian motion patterns are extracted, including the mean, variance, maximum, median, and minimum of the acceleration modulus, and the mean, variance, maximum, and minimum of the angular velocity modulus. These data features can be obtained by splitting the acceleration sequence and the angular velocity modulus sequence using a sliding time window. The present invention sets a data length of 2 seconds as the sliding window length, and each step length is 0.5 seconds, so the data overlap is 75%. When the acceleration or angular velocity modulus sequence meets the requirements, the segment of the sequence is cut out, and then the specified window length is slid backward, and the above steps are repeated to segment the IMU data to obtain more sample data containing deep features.

[0016] Optionally, during the execution of step 3, a multi-mode recognition model based on PCA-GWO-KELM is proposed. First, unimportant information in the data is removed by PCA, thereby achieving data dimensionality reduction. Then, the input weights and biases of KELM, as well as the population size and position parameters of the gray wolf are initialized according to the number of features after PCA dimensionality reduction; the root mean square error value of the training set of the KELM neural network is used as the individual fitness of the entire wolf pack, and the fitness of each gray wolf is calculated. The position of each gray wolf is updated according to the iterative formula of GWO, and the parameters in the KELM network are continuously adjusted to output the gray wolf position with the best performance and the corresponding KELM network parameters. Finally, the optimized KELM model is used to identify the movement pattern.

[0017] Optionally, during the execution of step 4, a peak detection method with dual adaptive thresholds is proposed. First, the candidate peak and trough values ​​of the combined acceleration are obtained, and then the adaptive acceleration upper threshold is calculated to update the peak value, and then the difference between the number of steps corresponding to the adaptive dynamic time threshold and the predicted number of steps is gradually determined until the limiting condition is met, and the peak value at this time is output, and finally the trough value is obtained according to the peak-to-valley constraint method to complete the step detection.

[0018] Optionally, during the execution of step 5, a multi-feature based adaptive step length estimation algorithm is used. The algorithm calculates the variance, amplitude difference, average value, and walking frequency of the acceleration in each step, and uses the improved Weinberg step length estimation model to first calculate the initial step length as the characteristic variable for predicting the step length result; then, the LASSO regression algorithm is combined with the regularization constraints of these characteristic variables, and the characteristic coefficient is solved by the coordinate descent method, and finally multiplied with the above characteristic variables to obtain the final step length estimation result.

[0019] Optionally, during the execution of step 6, a method of fusing direction sensor data and gyroscope data is used to calculate the heading angle.

[0020] Optionally, during the execution of step 7, the initial position of the target to be measured is first obtained, and then the target walking distance and moving direction are calculated through steps 4, 5, and 6, and finally the position coordinates of the pedestrian are continuously estimated.

[0021] The present invention provides a pedestrian positioning and navigation method for multimodal recognition. First, data is collected under six motion modes of walking reading, walking phone, walking swing, walking pocket, jogging phone, and jogging pocket, and motion features of different pedestrian motion modes are extracted, including 30-dimensional motion features such as the mean, variance, maximum, median, and minimum of the acceleration modulus, and the mean and variance of the angular velocity modulus. A multimodal recognition model based on principal component analysis and gray wolf algorithm optimized kernel extreme learning machine (PCA-GWO-KELM) is constructed. The extracted 30-dimensional motion features are input into the multimodal recognition model for training to obtain the recognition accuracy under different motion modes; then, a peak detection method with dual adaptive thresholds having an adaptive acceleration upper threshold and an adaptive dynamic time threshold is combined to perform step detection, an adaptive step length estimation algorithm based on multiple features is used to calculate the step length, and a method of fusing direction sensor data and gyroscope data is used to calculate the heading; finally, the pedestrian displacement is continuously dead reckoned to achieve low-cost and highly robust pedestrian navigation and positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a framework diagram of the pedestrian positioning and navigation method of multimodal recognition of the present invention.

[0023] Figure 2 It is a plan view of a test site in a specific embodiment of the present invention.

[0024] Figure 3 It is a comparison chart of motion pattern recognition accuracy of different classification models in a specific embodiment of the present invention.

[0025] Figures 4(a) to 4(f) It is a comparison diagram of the positioning trajectories of the traditional PDR and the proposed improved PDR method in a specific embodiment of the present invention.

[0026] Figure 5 It is a flow chart of the overall scheme in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0027] The implementation of the present invention is described in detail below, and the implementation examples are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The implementation described below with reference to the accompanying drawings is exemplary and is intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0028] See also Figures 1 to 5 The present invention provides a pedestrian positioning and navigation method based on multi-modal recognition, comprising the following steps:

[0029] S1: Record data under different motion modes and pre-process the data. When pedestrians hold mobile phones for navigation, the present invention defines six motion modes: walking reading, walking phone, walking swing, walking pocket, jogging phone, and jogging pocket. The present invention collects multiple sets of IMU data in different motion modes and then pre-processes them. First, the first few seconds and the last few seconds of the original data are eliminated, because the mobile phone is not stable during this period and the data will be disturbed; then a sliding mean filter is performed on it. This not only makes the pre-processed IMU data more accurate and reliable, but also removes some pseudo peaks and pseudo troughs, retaining good curve smoothness.

[0030] S2: Data feature extraction under different motion modes. By observing multiple sets of IMU data, it can be found that acceleration is closely related to the posture of the mobile phone, with a strong correlation and relatively regular data changes; angular velocity also shows a certain correlation. Therefore, this paper uses the output data of each axis of the filtered accelerometer and gyroscope and the modulus of the three-axis data for feature extraction. The modulus calculation formula for the three-axis data is as follows:

[0031]

[0032] where a x 、a y 、a z They are the three-axis component outputs of the accelerometer.

[0033]

[0034] where g x , g y , g z They are the three-axis component outputs of the gyroscope.

[0035] In order to improve the recognition accuracy of the model, the present invention extracts 30 data features that can reflect the characteristics of different pedestrian motion patterns, including the mean, variance, maximum, median, and minimum of the acceleration modulus, and the mean, variance, maximum, and minimum of the angular velocity modulus. These data features can be obtained by splitting the acceleration sequence and the angular velocity modulus sequence using a sliding time window. The present invention sets a data length of 2 seconds as the sliding window length, and each step length is 0.5 seconds, so the data overlap is 75%. When the acceleration or angular velocity modulus sequence meets the requirements, the sequence is intercepted, and then the specified window length is slid backward, and the above steps are repeated to segment the IMU data to obtain more sample data containing deep features.

[0036] S3: Construct a multi-mode recognition model based on PCA-GWO-KELM. In motion pattern recognition, there are many influencing factors, and there will inevitably be duplications between these variables. PCA is used to preprocess the data features, and the re-extracted indicators can reflect as much information as possible about different motion patterns. The extracted principal components are used as the input data of GWO-KELM, and then the sample data containing only the main control factors are divided into a training set and a test set, corresponding to the training and testing of the model respectively. KELM introduces the kernel function based on the ELM network to make it have a more powerful generalization ability. It can improve the prediction performance of the model while retaining the advantages of ELM. The present invention sets the number of input layer nodes to n, that is, the number of data feature types; the number of hidden layer nodes is set to l; the number of output layer nodes is set to 4, that is, four categories of prediction results. Given a sample (x data ,T label ), x data is the input data set, T label is the true label value, then:

[0037]

[0038] Where n represents the number of samples, represents the sth data feature of the nth sample, t i Indicates the label corresponding to the nth data feature. In the present invention, the number of samples is the total number of steps taken by pedestrians, and the data feature is the most representative main feature selected by PCA.

[0039] Next, randomly generate the positions of the wolves, calculate the fitness of the individuals in the wolf pack, whose value reflects the distance between the location of the gray wolf and the location of the prey, and find the first three gray wolves with the smallest fitness values. wolf Wolf is the optimal solution of the algorithm, β wolf Wolf is the suboptimal solution of the algorithm, δ wolf Wolf is the third best solution of the algorithm, and the other solutions are collectively referred to as ω wolf Wolves. The present invention sets the number of wolves to be m wolf , the dimension of the search space is d wolf , set the maximum number of iterations to μ max , the position of the gray wolf individual in the d-dimensional space can be expressed as:

[0040]

[0041] The root mean square error value of the training set of the KELM neural network is used as the fitness of the entire wolf pack. The smaller the root mean square error value, the smaller the error between the motion pattern recognition result predicted by the KELM model and the actual result.

[0042]

[0043] where f(x) kelm is the output function of KELM.

[0044] The ELM algorithm model can be expressed as:

[0045] T label = Ηβ (7)

[0046] Where H represents the output matrix of the hidden layer, and β represents the connection weight vector between the hidden layer and the output layer.

[0047] By using the least squares method, we can obtain β.

[0048] β=H + Y label =H T [H T H] -1 Y label (8)

[0049] Among them, H + Represents the generalized inverse matrix of the output matrix H of the hidden layer.

[0050] In order to enhance the stability of the network, the regularization coefficient C and the unit matrix I are introduced, and the least squares solution of the output weight can also be expressed as:

[0051]

[0052] In order to handle the nonlinearity of motion pattern data, KELM uses a kernel function to implicitly map the data into a high-dimensional feature space. The kernel function is introduced into ELM, and the kernel matrix is:

[0053] Ω elm =HH T =h(x i )h(x j )=K(x i ,x j ) (10)

[0054] Where h(x i )=H is the output of the hidden layer, K(x i ,x j ) is the kernel function. Because the radial basis kernel function has few parameters and good nonlinear processing ability, it is better adapted to the discrimination of motion modes. This paper uses the radial basis kernel function:

[0055]

[0056] At this time, the output function of KELM is:

[0057]

[0058] In summary, the kernel function parameter σ and the regularization coefficient C are important factors affecting the prediction performance of KELM.

[0059] Then, in the process of optimizing the GWO algorithm, the current parameters of the gray wolf and the spatial position of the wolf pack are continuously updated, and the fitness values ​​of all gray wolves are calculated. wolf Wolf, beta wolf Wolf and Delta wolf The wolf optimizes the wolf pack hunting and outputs the optimized optimal function parameter σ and regularization coefficient C when the maximum number of iterations of the wolf pack is reached.

[0060] The present invention proposes a multi-mode recognition model based on PCA-GWO-KELM, and adopts a method combining PCA and GWO-KELM to classify and identify the motion patterns of pedestrians. First, PCA retains the most important data features and removes unimportant information in the data, thereby achieving data dimension reduction. Then, the input weight and bias of KELM, as well as the population number and position parameters of gray wolves are initialized according to the number of features and the number of hidden layer neurons after PCA dimension reduction; the root mean square error value of the training set of the KELM neural network is used as the individual fitness of the entire wolf pack, and the fitness of each gray wolf is calculated. The position of each gray wolf is updated according to the iterative formula of GWO, and the new position and state of each gray wolf are determined. The parameters in the KELM network are adjusted, and the gray wolf position with the best performance and the corresponding KELM network parameters are selected as the final solution. Finally, the optimized KELM model is used to identify the motion pattern.

[0061] S4: Step counting is performed through the peak detection method with dual adaptive thresholds. When using the traditional peak detection algorithm to detect the number of steps a person walks, a large number of "pseudo peaks" and "pseudo troughs" will appear, resulting in step counting errors. In order to improve the accuracy of step counting, the present invention proposes a dual adaptive threshold peak detection method for step counting. The intensity of the three-axis acceleration reaction varies when pedestrians are in different mobile phone holding modes. Therefore, in the flat end mode, the combined acceleration of the Z axis and the Y axis is used; in the swing mode, the combined acceleration of the X axis and the Y axis is used; in the phone mode and pocket mode, the combined acceleration of the X axis and the Y axis is used. The specific steps of the algorithm are:

[0062] (1) Candidate peak / trough detection: In order to reduce the impact of device posture changes, we use the combined acceleration Acc in the corresponding mode. cp Step detection is performed. First, the combined acceleration is processed by moving mean filtering, and then the candidate peak points and candidate trough points are obtained according to formulas (13)-(14).

[0063]

[0064] in is the candidate peak value at sampling point i.

[0065]

[0066] in is the candidate peak value at sampling point j.

[0067] (2) Updating the peak value: During normal walking of pedestrians, the peak value of acceleration tends to be stable. Therefore, the present invention sets an adaptive acceleration upper threshold to eliminate some pseudo peaks with large differences. As shown in formula (16), the adaptive acceleration upper threshold is achieved by calculating the average value of the positive acceleration signal within a fixed time period.

[0068]

[0069] in is the adaptive acceleration upper threshold, and c is the number of candidate peaks in step (1).

[0070]

[0071] in is the updated candidate peak value.

[0072] (3) Peak value verification: After completing the above steps, the time interval between two consecutive candidate peaks must also be considered. When pedestrians walk normally, the step time between two adjacent steps will not change much, and the difference range between the two steps on the time axis is limited, so the time value between adjacent peaks should be within a certain interval. The present invention sets an adaptive dynamic time threshold to verify the effective peak. Record the total number of steps predicted by the multi-mode recognition model based on PCA-GWO-KELM, and then gradually determine the difference between the number of steps corresponding to the threshold and the predicted number of steps. Limitation condition: The difference between the number of steps predicted by the adaptive dynamic time threshold and the PCA-GWO-KELM model is the smallest. If the difference is greater than the limitation condition, a certain value is added to the initial threshold, and the above judgment steps are iterated until the limitation condition is met, the number of steps at this time is output, and the loop is stopped.

[0073]

[0074] is the peak value after verification. is an adaptive dynamic time interval, with an initial time threshold of 0.1s. The time interval between the peak value of the current wave and the peak value of the previous moment is greater than Then the values ​​of the two peaks are compared. If the peak value of the current peak is greater than the peak value of the previous peak, it can be determined as a real peak. If the condition is not met, the current peak value is regarded as a false peak and filtered out.

[0075] (4) Searching for troughs: A complete acceleration increase and decrease process is counted as one step. There must be a trough between two valid peaks. This method can be called the peak-to-valley pair constraint method. In the above steps, we have detected the valid peak value. Then we can process the same acceleration data sequence according to the peak-to-valley pair constraint method, remove the pseudo trough value, and finally the number of steps is determined by the smaller value between the number of peaks and the number of valleys.

[0076] The peak detection method of dual adaptive thresholds proposed in the present invention first obtains candidate peak and trough values, then calculates the adaptive acceleration upper threshold to update the peak value, then gradually determines the difference between the number of steps corresponding to the adaptive dynamic time threshold and the predicted number of steps, until the limiting condition is met, outputs the peak value at this time, and finally obtains the trough value according to the peak-to-trough constraint method to complete the step detection. In this method, the two adaptive thresholds set effectively improve the accuracy of gait detection.

[0077] S5: Calculate the step length through an adaptive step length estimation algorithm based on multiple features. The peak value of the acceleration signal is not the same in different motion modes, and the walking frequency and acceleration variance vary. If only a single step length estimation model is used, it will lead to large errors. The adaptive step length estimation algorithm based on multiple features adopted in the present invention calculates the variance, amplitude difference, average value, walking frequency, peak median, and trough median of the acceleration in each step, and uses the improved Weinberg step length estimation model to first calculate the initial prediction result of the step length; then, the LASSO regression algorithm is combined with the regularization constraints of these characteristic variables, and the coordinate descent method is used to solve the characteristic coefficient, and finally multiplied by the above-mentioned characteristic variables to obtain the final step length estimation result.

[0078] S6: Calculate the heading angle by fusing the direction sensor data and the gyroscope data. Since the direction of the smartphone posture cannot be completely consistent with the actual movement direction of the pedestrian, if the uncorrected yaw angle is used as the pedestrian's current heading angle, a large heading deviation is bound to occur. Therefore, the present invention adopts a method based on Kalman filtering to combine sensor data and gyroscope data to solve the heading angle. The core idea of ​​the algorithm is to calculate the optimal heading angle estimate at the current moment based on the heading angle measurement value at the current moment and the heading angle prediction at the previous moment, and the result is used to predict the heading angle estimate at the next moment. During the fusion process, the heading angle collected by the direction sensor is used as the system observation value, and the heading angle collected by the gyroscope is used as the state value. The algorithm is looped until the heading angle estimate value of each step is calculated, and it is used for the heading information of the pedestrian position update.

[0079] S7: Perform dead reckoning positioning estimation under different motion modes. First, obtain the initial position of the target to be measured, then calculate the target walking distance and moving direction through S4, S5, and S6, and finally estimate the position coordinates of the pedestrian after moving; loop through steps S3, S4, S5, and S6 to achieve continuous estimation of the pedestrian's position, and finally obtain the trajectory of the pedestrian under different motion modes.

[0080] Furthermore, in order to evaluate the performance of the proposed positioning method, the present invention also proposes a specific embodiment, and conducts multiple experiments in an indoor corridor with a size of 115×1.8×3m 3 The blue solid line indicates the movement trajectory of the pedestrian, and the blue solid arrow indicates the movement direction. Figure 2 In the experiment, the starting and ending points of the experimental paths are different. Pedestrians hold smartphones and walk along the test paths at a speed of 0.6m / step in different pedestrian motion modes, and the smartphones automatically collect IMU data. In this experiment, the Vivo X30 smart device is used to collect data.

[0081] In order to evaluate the performance of the multi-pattern recognition model based on PCA-GWO-KELM in the method of the present invention, the embodiment is compared with several classification methods such as support vector machine (SVM), random forest (RF), extreme learning machine (ELM), and grey wolf optimized extreme learning machine (GWO-ELM).

[0082] Based on the data collected from various pedestrian motion patterns for different motion patterns, the accuracy of the motion pattern recognition model is tested using the test data to obtain the motion pattern recognition accuracy under different classification models. In the walking reading, walking phone, walking swing, walking pocket mode, jogging phone, jogging pocket and other motion modes, such as Figure 3As shown in the figure, the multi-mode recognition model based on PCA-GWO-KELM generally has the highest recognition accuracy and can achieve high-precision motion pattern recognition. This is because the PCA-GWO-KELM model effectively reduces the randomness of the model and significantly improves the stability of recognition accuracy by comprehensively using PCA feature dimensionality reduction, GWO parameter optimization and KELM kernel method.

[0083] Table 1 lists the average recognition accuracy of the recognition model proposed in the present invention and the SVM, RF, ELM, and GWO-ELM models for all pedestrian motion modes. The recognition model proposed in the present invention has an average recognition accuracy of 97.63% for walking reading, walking phone, walking swing, walking pocket mode, jogging phone, and jogging pocket. Compared with the SVM, RF, ELM, and GWO-ELM models, the average recognition accuracy is 9.67%, 7.09%, 5.35%, and 3.29% higher, respectively. This shows that the recognition model proposed in the present invention can accurately identify the current motion mode and provide a basis for pedestrian dead reckoning technology.

[0084] Table 1. Average recognition accuracy of different classification models for motion patterns (%)

[0085]

[0086] Table 2 lists the recognition accuracy of the recognition model proposed by the present invention and the SVM, RF, ELM, and GWO-ELM models after running 5 times. SVM has good stability but poor accuracy; the maximum difference in RF recognition accuracy reaches 5.74%; the random initialization input weight and bias characteristics of the ELM model lead to its unstable performance. We can see that its maximum difference in recognition accuracy reaches 6.27%; and although the GWO-ELM model uses GWO to optimize the random parameters of ELM, GWO is also randomly set when initializing the position of the gray wolf. GWO alone cannot reduce the randomness of ELM well. It can be seen that the recognition accuracy difference of the GWO-ELM model reaches 2.57%; the PCA-GWO-KELM model not only introduces the kernel function parameters and regularization coefficients optimized by GWO to improve the performance of ELM, but also uses PCA to perform feature dimensionality reduction before inputting data, which effectively improves the recognition accuracy of ELM. It can be seen that the recognition accuracy difference of the PCA-GWO-KELM model is only 0.1%, so it can be shown that the PCA-GWO-KELM model has high stability.

[0087] Table 2. Model recognition accuracy of 5 tests (%)

[0088]

[0089] In addition, in order to verify the positioning performance of the method proposed in the present invention, pedestrians were asked to collect data along the corridor in walking reading, walking phone, walking swing, walking pocket mode, jogging phone, and jogging pocket, and comparative experiments were carried out with the traditional PDR algorithm.

[0090] As shown in Figure 4 (a), (b), (c), (d), (e), and (f), they are the positioning trajectories of the traditional PDR algorithm and the DR method proposed in the present invention under walking reading, walking phone, walking swing, walking pocket mode, jogging phone, and jogging pocket. We can see that the positioning trajectory of the improved DR method of the present invention shows a high degree of overlap with the real trajectory under different motion modes. This is because before DR positioning, multi-mode recognition based on PCA-GWO-KELM is first performed, and then the step frequency detection is performed using the proposed dual adaptive threshold peak detection method, the step length estimation is performed using the multi-feature based adaptive step length estimation algorithm, and the dead position angle estimation is performed using the method of fusing the direction sensor data and the gyroscope data. These methods have high estimation accuracy, which makes the positioning accuracy of the proposed DR method higher and the robustness better.

[0091] In order to solve the problem of cumulative errors in PDR, the present invention proposes a multi-mode recognition model of PCA-GWO-KELM, uses principal component analysis to reduce data dimension, uses the gray wolf algorithm to optimize the kernel extreme learning machine network parameters, and constructs a PCA-GWO-KELM model to identify motion patterns. The model has high recognition accuracy and can provide a basis for pedestrian dead reckoning technology. Then a dual adaptive threshold peak detection method is proposed. The algorithm is based on the recognition results of the motion pattern, combined with the adaptive acceleration upper threshold and the adaptive dynamic time threshold to gradually update the peak value, and uses an adaptive step length estimation algorithm based on multiple features to calculate the step length, and uses a method of fusing direction sensor data and gyroscope data to calculate the heading angle. In the embodiment, multiple groups of experiments are conducted on a planned path at an experimental site. The experimental scene is a corridor with dimensions of 115×1.8×3m. 3 . Experimental results show that this method can effectively reduce the cumulative error of dead reckoning under various pedestrian motion patterns and achieve high positioning accuracy. It shows that the use of motion pattern recognition technology and efficient peak detection method to assist dead reckoning can help improve the impact of IMU data under different pedestrian motion states and different mobile phone holding modes, and achieve low-cost, stable and convenient indoor positioning.

[0092] In summary, the present invention solves the following technical problems:

[0093] 1. The present invention proposes a multi-mode recognition model based on PCA-GWO-KELM, which classifies and recognizes the motion patterns of pedestrians by combining PCA and GWO-KELM. First, PCA is used to remove unimportant information in the data, thereby achieving data dimensionality reduction. Then, the input weights and biases of KELM, as well as the population size and position parameters of gray wolves are initialized according to the feature number after PCA dimensionality reduction; the root mean square error value of the training set of the KELM neural network is used as the individual fitness of the entire wolf pack, and the fitness of each gray wolf is calculated. The position of each gray wolf is updated according to the iterative formula of GWO, and the gray wolf position with the best performance and the corresponding KELM network parameters are output. Finally, the optimized KELM model is used to identify the motion pattern. Experimental results show that compared with RF, SVM, ELM, and GWO-ELM, the multi-mode recognition model based on PCA-GWO-KELM has a higher recognition accuracy.

[0094] 2. The present invention proposes a pedestrian positioning and navigation method based on multimodal recognition. In this method, the motion pattern of pedestrians is firstly effectively identified by a multimodal recognition model based on PCA-GWO-KELM, and then a dual adaptive threshold peak detection method is proposed for step frequency detection. Compared with the existing peak detection method, it not only has higher recognition accuracy, but also greatly improves the algorithm speed; an adaptive step length estimation algorithm based on multiple features is used to calculate the step length; and a method of fusing direction sensor data and gyroscope data is used to calculate the heading angle. Experimental results show that compared with the traditional PDR algorithm, this method has higher positioning accuracy and better robustness under different pedestrian motion modes.

[0095] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A pedestrian positioning and navigation method based on multi-modal recognition, It is characterized in that The following steps are involved: Step 1: Record data under different motion modes and pre-process the data; Step 2: Data feature extraction under different motion modes; Step 3: Construct a multi-pattern recognition model based on principal component analysis and gray wolf algorithm to optimize the kernel extreme learning machine; Step 4: Count the number of steps using the peak detection method with dual adaptive thresholds; Step 5: Calculate the step length using a multi-feature based adaptive step length estimation algorithm; Step 6: Calculate the heading angle by fusing the direction sensor data and the gyroscope data; Step 7: Perform dead reckoning positioning estimation under different motion modes.

2. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 1, It is characterized in that During the execution of step 1, the data of the first few seconds and the last few seconds of the original data are first removed, and then the sliding mean filtering is performed on the data.

3. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 2, It is characterized in that During the execution of step 2, 30 time domain features that can reflect the characteristics of different pedestrian motion patterns are extracted. These data features can be obtained by splitting the acceleration sequence and angular velocity sequence using a sliding time window. The data length is set to 2 seconds as the sliding window length, and the length of each step is 0.5 seconds, so the data overlap is 75%. When the acceleration or angular velocity modulus sequence meets the requirements, the sequence is cut out, and then the specified window length is slid backward, and the above steps are repeated to segment the IMU data to obtain more sample data with deep features.

4. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 3, It is characterized in that In the execution process of step 3, firstly, unimportant information in the data is removed through PCA to achieve dimensionality reduction of data features; Then, the input weights and biases of KELM, as well as the population size and position parameters of gray wolves, are initialized according to the number of features after PCA dimensionality reduction. The root mean square error value of the training set of the KELM neural network is used as the individual fitness of the entire wolf pack, and the fitness of each gray wolf is calculated. The position of each gray wolf is updated according to the iterative formula of GWO, and the parameters in the KELM network are continuously adjusted to output the gray wolf position with the best performance and the corresponding KELM network parameters. Finally, the optimized KELM model is used to recognize the movement pattern.

5. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 4, It is characterized in that During the execution of step 4, first obtain the candidate peak and trough values ​​of the combined acceleration, then calculate the adaptive acceleration upper threshold to update the peak value, and then gradually determine the difference between the number of steps corresponding to the adaptive dynamic time threshold and the predicted number of steps until the limit condition is met, output the peak value at this time, and finally obtain the trough value according to the peak-to-valley pair constraint method to complete the step detection.

6. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 5, It is characterized in that During the execution of step 5, the algorithm calculates the variance, amplitude difference, average value, and walking frequency of the acceleration in each step, and uses the improved Weinberg step length estimation model to first calculate the initial step length as the characteristic variable for predicting the step length result; then, the LASSO regression algorithm is combined with the regularization constraints of these characteristic variables, and the coordinate descent method is used to solve the characteristic coefficients, and finally multiplied with the above characteristic variables to obtain the final step length estimation result.

7. A pedestrian positioning and navigation method based on multimodal recognition as claimed in claim 6, It is characterized in that During the execution of step 6, the heading angle is estimated by combining the sensor data and the gyroscope data through the extended Kalman filter algorithm, wherein the sensor data is used as the state quantity and the gyroscope data is used as the observation quantity.

8. The pedestrian positioning and navigation method based on multi-modal recognition as claimed in claim 7, It is characterized in that During the execution of step 7, the initial position of the target to be measured is first obtained, and then the target walking distance and moving direction are calculated through steps 4, 5, and 6, and finally the position coordinates of the pedestrian are continuously estimated.

Images