Neural network geomagnetic positioning method based on multi-feature driving and computer equipment

Abstract

The application is suitable for the field of indoor positioning, and provides a neural network geomagnetic positioning method based on multi-feature driving, a computer readable storage medium and computer equipment. The method comprises the following steps: receiving geomagnetic signals and sensor information collected by a mobile terminal, and creating a multi-dimensional data set for geomagnetic positioning and step estimation; performing geomagnetic positioning estimation based on a ResNet-GRU-LSTM neural network model to obtain a predicted geomagnetic positioning result; using a hierarchical GRU neural network model to obtain a step, a step frequency and a direction angle for predicting a pedestrian dead reckoning (PDR) positioning result; and performing a geomagnetic-assisted multi-feature-driven PDR positioning through particle filtering according to the step, the step frequency, the direction angle and the geomagnetic positioning result to obtain a final position of a target. The application can effectively eliminate the fuzzy positioning error and abnormal value caused by geomagnetic positioning and PDR positioning, thereby significantly improving the positioning performance, effectively alleviating the heterogeneity of the equipment, and having strong robustness.

Description

Technical Field

[0001] This application belongs to the field of indoor positioning, and in particular relates to a neural network geomagnetic positioning method based on multi-feature driven, a computer-readable storage medium, and a computer device. Background Technology

[0002] With the development of microelectromechanical systems (MEMS), the installation of MEMS-based sensors in mobile terminals has become standard, enabling the wider application of location-based services (LBS), such as pedestrian and target navigation, smart advertising, and location-based tweeting. In outdoor environments, Global Navigation Satellite Systems (GNSS) can provide efficient and stable positioning and navigation services to users worldwide, achieving accurate real-time positioning. However, indoor environments are more complex than open outdoor scenarios; weak satellite signals are easily blocked by buildings or other obstacles, preventing them from functioning properly and meeting users' location service requirements. Therefore, achieving ubiquitous, inexpensive, and more accurate indoor positioning technology is of great significance.

[0003] Researchers have conducted a series of studies on WiFi positioning, Bluetooth positioning, ultrasonic positioning, and visual positioning. WiFi and Bluetooth-based positioning methods are low-cost but susceptible to environmental factors such as multipath propagation and signal absorption. Ultrasonic positioning offers high accuracy but requires additional equipment to collect sound signals. Visual positioning has wide applications but is easily affected by environmental obstructions and lighting conditions, and privacy is difficult to guarantee. In Pedestrian Track Deduction (PDR) positioning, IMU sensors embedded in mobile terminals can assist in positioning by estimating pedestrian walking characteristics in the indoor environment (including step frequency detection, step length estimation, and orientation angle estimation). PDR has relatively high positioning accuracy over short distances and is low-cost, small in size, and highly autonomous; however, PDR positioning can generate cumulative and drift errors over long periods.

[0004] The temporal stability and universality of geomagnetic signals enable geomagnetic fingerprint-based positioning. A key factor in geomagnetic fingerprint-based positioning is identifying more unique features related to location information within the dynamically limited geomagnetic sequence, thereby obtaining stable and reliable positioning results. Therefore, this paper considers using geomagnetic sequences to learn the relationship between time-correlated location tags and the geomagnetic sequence to achieve indoor positioning. Geomagnetic fingerprint-based positioning can overcome the cumulative and drift errors of long-term positioning caused by PDR positioning and is independent of the initial position, thus enabling independent positioning. However, using discrete geomagnetic signals may lead to blurred positioning features due to random noise and similar magnetic field readings. Summary of the Invention

[0005] The purpose of this application is to provide a neural network geomagnetic positioning method based on multi-feature driven, a computer-readable storage medium, and a computer device, which aims to solve the problems of long-term positioning cumulative error and drift error generated by PDR positioning, and the problem that positioning features may be blurred due to random noise and similar magnetic field readings when using discrete geomagnetic signals.

[0006] In a first aspect, this application provides a neural network-based geomagnetic positioning method driven by multiple features, the method comprising: S101. Receive the geomagnetic signals and sensor information collected by the mobile terminal, and perform preprocessing to create a multidimensional dataset for geomagnetic positioning and a multidimensional dataset for step size estimation. S102. Geomagnetic positioning estimation based on ResNet-GRU-LSTM neural network model: The constructed ResNet-GRU-LSTM neural network model is used to extract geomagnetic features from the multidimensional dataset used for geomagnetic positioning, and the relationship between the geomagnetic sequence of each geomagnetic component and the location label is learned. The multidimensional dataset used for geomagnetic positioning is put into the trained ResNet-GRU-LSTM neural network model to predict the location and obtain the predicted geomagnetic positioning result. S103. Using a hierarchical GRU neural network model, extract step length information from the multidimensional dataset used for step length estimation, perform multi-feature driven step length estimation, and obtain the step length, step frequency, and orientation angle used to predict pedestrian trajectory and calculate PDR positioning. S104. Based on the step size, step frequency, and orientation angle used to predict the pedestrian trajectory for PDR positioning, as well as the predicted geomagnetic positioning results, geomagnetically assisted multi-feature driven pedestrian trajectory PDR positioning is performed through particle filtering to obtain the final position of the target.

[0007] Secondly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the described multi-feature-driven neural network geomagnetic positioning method.

[0008] Thirdly, this application provides a computer device, comprising: One or more processors; Memory; and One or more computer programs, wherein the processor and the memory are connected via a bus, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, wherein the processors, when executing the computer programs, implement the steps of the multi-feature-driven neural network geomagnetic positioning method as described above.

[0009] In this application, geomagnetic positioning estimation is based on a ResNet-GRU-LSTM neural network model. ResNet has excellent feature extraction capabilities, and LSTM can extract global geomagnetic features to obtain more accurate information, thereby further improving the predictive ability of the ResNet-GRU-LSTM neural network model, resulting in small positioning errors. Furthermore, by utilizing a hierarchical GRU neural network model to extract step length information from the multidimensional dataset used for step length estimation and performing multi-feature-driven step length estimation, more accurate features can be learned to more accurately and reliably estimate the step length of pedestrians of different heights. Moreover, based on the step length, step frequency, and orientation angle used to predict pedestrian trajectories for PDR positioning, as well as the predicted geomagnetic positioning results, particle filtering is used to perform geomagnetically assisted multi-feature-driven pedestrian trajectories for PDR positioning, obtaining the final target position. Therefore, it effectively eliminates fuzzy positioning errors and outliers caused by geomagnetic positioning and PDR positioning, significantly improving positioning performance, effectively mitigating device heterogeneity, and exhibiting strong robustness. This application offers great practicality and flexibility for different devices, scenarios, and pathways. Attached Figure Description

[0010] Figure 1 This is a flowchart of a neural network geomagnetic positioning method based on multi-feature driven by an embodiment of this application.

[0011] Figure 2 This is a structural block diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0013] To illustrate the technical solution described in this application, specific embodiments are provided below.

[0014] Please see Figure 1 This is a flowchart of a multi-feature-driven neural network geomagnetic positioning method provided in an embodiment of this application. This embodiment mainly uses the application of the multi-feature-driven neural network geomagnetic positioning method to computer equipment as an example for illustration. The multi-feature-driven neural network geomagnetic positioning method provided in an embodiment of this application includes the following steps: S101. Receive geomagnetic signals and sensor information collected by the mobile terminal, and preprocess them to create a multidimensional dataset for geomagnetic positioning and a multidimensional dataset for step size estimation in order to extract more features.

[0015] In one embodiment of this application, the sensor information includes acceleration, gyroscope data, and orientation data.

[0016] The mobile device needs to have a data acquisition application installed beforehand. As pedestrians walk along the planned trajectory with their mobile devices in hand, geomagnetic signals and sensor information will be automatically collected and sent to computer equipment (i.e., the server) for preprocessing.

[0017] In one embodiment of this application, creating a multidimensional dataset for geomagnetic positioning can specifically be: constructing a geomagnetic dataset containing 5-dimensional geomagnetic components, as shown in equation (1), wherein the geomagnetic dataset has high distinguishability and positioning reliability.

[0018] (1) in, This represents the three-axis magnetic field obtained by the magnetometer in the mobile terminal. and Let the amplitude and horizontal components represent the amplitude and horizontal components, respectively, obtained from formula (2): (2) Then, the median averaging filter with a predetermined window size is used to reduce noise during the data acquisition process. Specifically, the maximum and minimum values ​​of the geomagnetic sequence of each geomagnetic component in the window are removed, and then the remaining data is averaged.

[0019] To mitigate the impact of different magnetic feature distributions on the ResNet-GRU-LSTM neural network model, the geomagnetic dataset is Z-score normalized, as shown in formula (3): (3) in, , These represent the mean and standard deviation of the geomagnetic dataset, respectively. It is a standardized geomagnetic dataset.

[0020] During the model training and prediction phases, a fixed-size sliding window is used to segment the geomagnetic sequence for each location label.

[0021] In one embodiment of this application, creating a multidimensional dataset for step size estimation can specifically involve: constructing a multi-feature dimension... As a step size dataset used for step size estimation, the step size dataset is direction-independent and contains sufficient step size information for model training. The acceleration and gyroscope measurements can be expressed as shown in Equation (4): (4) in, , , , , , These are the x, y, and z components of the acceleration and gyroscope, respectively. , The acceleration and the magnitude of the gyroscope are obtained from equation (5): (5) Normalize the step size dataset. (6) in, , These are the minimum and maximum values ​​of the dataset.

[0022] The step length dataset is segmented according to the step frequency, with each segment corresponding to a step length. Segments are generated using a sliding window, as shown in formula (7): (7) Where i is the size of the sliding window, and N is a natural number greater than 1.

[0023] S102. Geomagnetic positioning estimation is performed based on the ResNet-GRU-LSTM neural network model. The constructed ResNet-GRU-LSTM neural network model is used to extract geomagnetic features from the multidimensional dataset used for geomagnetic positioning. The relationship between the geomagnetic sequence and the location label of each geomagnetic component is learned. The multidimensional dataset used for geomagnetic positioning is put into the trained ResNet-GRU-LSTM neural network model to predict the location and obtain the predicted geomagnetic positioning result.

[0024] In this invention, ResNet34 is used as the first layer neural network model. It employs a shortcut connection architecture to learn the required mappings, thus solving the degradation problem of neural networks. Therefore, it can efficiently capture the local correlation features of the input data and extract and learn the hidden features in the magnetic subsequence. ResNet34 includes Conv1, Conv2_x, Conv3_x, Conv4_x, and Conv5_x. Conv1 consists of a 7×7 convolutional layer and a max-pooling layer. Conv2_x, Conv3_x, Conv4_x, and Conv5_x consist of 3, 4, 6, and 3 residual blocks, respectively, each containing two convolutional layers. The residual blocks consist of multiple levels of residual mappings and identity mappings. The segmented magnetic sequence... Send to ResNet34.

[0025] In residual mapping, for layer The output is: (8) in, It is the first Magnetic sequence input of the layer, and It is the weight of each layer. is the activation function for the residual mapping network.

[0026] Therefore, the output of the residual block The implicit features are: (9) Then, the extracted hidden features The data is sent to the GRU layer for further extraction of the temporal features of the magnetic subsequence. The GRU's update gate controls the ratio of past magnetic field information to current information, while the reset gate controls how much of the previous geomagnetic state needs to be stored. The latent features are input at time t. In this case, the equations for GRU are defined as follows: (10) in, , These are the update door and the reset door, respectively. It is based on the hidden state calculated by resetting the door. It is the output of the neuron at time t-1. It is the output of the neuron at time t. , , It is a weight matrix. , , Represents the deviation vector. It is the sigmoid function.

[0027] LSTM layers are used to extract global temporal features, thereby achieving accurate predictions. The basic unit of an LSTM has three gate structures: a forget gate, an input gate, and an output gate. The input magnetic feature data at time t... Unit status and the output of the previous neuron Together, they determine the forgotten portion of the state memory unit. The formula for LSTM is as follows: (11) in, , , , and These represent the input gate, forget gate, output gate, storage unit state, and current neuron output, respectively. , and These represent the weights of the input gate, forget gate, and output gate, respectively. , , These represent the biases of the input gate, forget gate, and output gate, respectively. It is a bias in the cell state. It is an activation function.

[0028] To prevent overfitting, a Dropout layer is used between the GRU and LSTM layers to randomly drop a predetermined number of neurons, thereby improving the network's generalization ability. The final regression layer consists of two fully connected layers, used to output the predicted pedestrian location. During model training, the proposed ResNet-GRU-LSTM neural network model is trained using a multidimensional dataset for geomagnetic positioning and relevant location labels. After training, the predicted geomagnetic positioning result at time t+1 can be obtained by inputting test data. , .

[0029] S103. Using a hierarchical GRU neural network model, step length information is extracted from the multidimensional dataset used for step length estimation. Multi-feature driven step length estimation is performed to obtain the step length, step frequency, and orientation angle used to predict pedestrian trajectory and calculate PDR positioning, thereby achieving more accurate step length estimation.

[0030] The hierarchical GRU neural network model comprises one GRU layer, two LSTM layers, and one regression layer. To prevent overfitting, a dropout layer is applied after each layer, with a dropout rate of 0.2. In the hierarchical GRU neural network model, the segmented multi-feature inertial data sequence is first input into the LSTM layer for local feature extraction. As the first layer, the LSTM can selectively extract implicit leading information from the input inertial data, thus preserving rich local relevant information. Next, GRU is used to further extract features from the inertial data. Then, the extracted features are passed to the final LSTM layer to extract global time-series information. Finally, these extracted features are randomly discarded and fed into the regression layer. The regression layer contains two fully connected layers with sigmoid activation functions. The regression layer maps the feature vectors of the inertial data sequence to the relevant pedestrian stride length, thus establishing a mapping relationship between inertial data and stride length. During the training phase, the training dataset and corresponding stride length labels are input into the hierarchical GRU neural network model for training. After training, stride length can be predicted by inputting unlabeled test data.

[0031] Therefore, multi-feature driven step size estimation can be characterized as: (12) in, This is a segmented test step size dataset; the calculation method is described in S101. It is a step-size estimation model, the H-GRU neural network.

[0032] Finally, the step size, step frequency, and orientation angle are obtained from the H-GRU model and used to predict the pedestrian trajectory and calculate PDR positioning.

[0033] S104. Based on the step size, step frequency, and orientation angle used to predict the pedestrian trajectory for PDR positioning, as well as the predicted geomagnetic positioning results, geomagnetically assisted multi-feature driven pedestrian trajectory PDR positioning is performed through particle filtering to obtain the final position of the target.

[0034] In one embodiment of this application, S104 specifically includes the following steps: S1041, Initialize particles.

[0035] The initial particle set is generated at time t. In this system, each particle has an equal weight, and the initial particle set is generated by magnetic positioning. For particles The estimated geomagnetic location at time t.

[0036] S1042. Use pedestrian tracks to estimate PDR positioning, establish a particle motion model, and update particle positions.

[0037] After obtaining the initial particle set, the position of the particles at time t+1. Update according to formula (13).

[0038] (13) in, , These represent the step size and heading at time t+1, respectively. The noise follows a Gaussian distribution with a mean of zero and a variance of 1.

[0039] S1043. Reassess the particle weight based on the geomagnetic positioning results.

[0040] Assume the geomagnetic estimation of the system observations at time t+1 As shown in equation (14): (14) in, , This is the geomagnetic positioning result at time t+1. To observe Gaussian noise.

[0041] Based on the observations, the particle weight is reassessed as follows: (15) in, It is the covariance matrix of the geomagnetic estimation.

[0042] S1044. Normalize the particle weights based on the re-evaluated particle weights, and resample the particles based on the weight values.

[0043] In one embodiment of this application, S1044 specifically includes the following steps: The particle weights are normalized using formula (16): (16) Resample the normalized particles, focusing on particles with high weights that are close to the true state, and replicate them; reset the particle weights to ensure that the sum of all weights is 1.

[0044] S1045. Estimate the fused localization based on the updated particle positions and normalized particle weights.

[0045] The fusion localization at time t+1 is estimated using formula (17): (17) in, These are the normalized particle weights. It represents the position of the updated particle at time t+1.

[0046] By repeatedly executing S102, S103, and S104, continuous estimation of pedestrian positions can be achieved.

[0047] In this embodiment, since geomagnetic positioning estimation is based on a ResNet-GRU-LSTM neural network model, it can detect and effectively eliminate fuzzy positioning caused by unclear magnetic field signals, thereby improving feature extraction capability and positioning accuracy. This not only achieves better performance and effectively extracts geomagnetic feature location information, but also exhibits good robustness in heterogeneous devices. Furthermore, by utilizing a hierarchical GRU neural network model to extract step-size information from the multidimensional dataset used for step-size estimation, multi-feature-driven step-size estimation is performed to obtain the predicted step-size, thus achieving more accurate step-size estimation and solving the problems of accumulated error, drift error, and device heterogeneity. Moreover, since this application creates multidimensional datasets for geomagnetic positioning and step-size estimation, more features can be extracted. By using particle filtering for geomagnetically assisted multi-feature-driven PDR positioning, more reliable and robust positioning results are obtained. Therefore, the multi-feature-driven neural network geomagnetic positioning method of this application has more stable and reliable positioning performance, can achieve higher positioning accuracy, can effectively alleviate device heterogeneity, has strong robustness, and relatively low computational complexity. Furthermore, since the acquisition of geomagnetic signals and sensor information can be achieved solely through a mobile terminal without requiring additional infrastructure, the cost is low. Therefore, the multi-feature-driven neural network geomagnetic positioning method proposed in this application is feasible in indoor LBS and has great practicality and flexibility for different mobile terminals, scenarios, and paths.

[0048] One embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multi-feature-driven neural network geomagnetic positioning method provided in one embodiment of this application.

[0049] Figure 2 The diagram illustrates a specific structural block diagram of a computer device provided in an embodiment of this application. The computer device 100 includes: one or more processors 101, a memory 102, and one or more computer programs. The processors 101 and the memory 102 are connected via a bus. The one or more computer programs are stored in the memory 102 and configured to be executed by the one or more processors 101. When the processors 101 execute the computer programs, they implement the steps of the multi-feature-driven neural network geomagnetic positioning method provided in an embodiment of this application.

[0050] In this application, geomagnetic positioning estimation is based on a ResNet-GRU-LSTM neural network model. ResNet has excellent feature extraction capabilities, and LSTM can extract global geomagnetic features to obtain more accurate information, thereby further improving the predictive ability of the ResNet-GRU-LSTM neural network model, resulting in small positioning errors. Furthermore, by utilizing a hierarchical GRU neural network model to extract step length information from the multidimensional dataset used for step length estimation and performing multi-feature-driven step length estimation, more accurate features can be learned to more accurately and reliably estimate the step length of pedestrians of different heights. Moreover, based on the step length, step frequency, and orientation angle used to predict pedestrian trajectories for PDR positioning, as well as the predicted geomagnetic positioning results, particle filtering is used to perform geomagnetically assisted multi-feature-driven pedestrian trajectories for PDR positioning, obtaining the final target position. Therefore, it effectively eliminates fuzzy positioning errors and outliers caused by geomagnetic positioning and PDR positioning, significantly improving positioning performance, effectively mitigating device heterogeneity, and exhibiting strong robustness. This application offers great practicality and flexibility for different devices, scenarios, and pathways.

[0051] It should be understood that the steps in the various embodiments of this application are not necessarily executed sequentially according to the order indicated by the step numbers. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0052] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A neural network-based geomagnetic positioning method driven by multiple features, characterized in that, The method includes: S101. Receive the geomagnetic signals and sensor information collected by the mobile terminal, and perform preprocessing to create a multidimensional dataset for geomagnetic positioning and a multidimensional dataset for step size estimation. S102. Geomagnetic positioning estimation based on the ResNet-GRU-LSTM neural network model: The constructed ResNet-GRU-LSTM neural network model is used to extract geomagnetic features from the multidimensional dataset used for geomagnetic positioning, learn the relationship between the geomagnetic sequence of each geomagnetic component and the location label, and put the multidimensional dataset used for geomagnetic positioning into the trained ResNet-GRU-LSTM neural network model to predict the location and obtain the predicted geomagnetic positioning result. Among them, ResNet34 is used as the first layer of the neural network model to capture the local correlation features of the input data, extract and learn the hidden features in the magnetic subsequence, send the extracted hidden features to the GRU layer to further extract the temporal features of the magnetic subsequence, and the LSTM layer is used to extract the global temporal features. S103. A hierarchical GRU neural network model is used to extract step size information from the multidimensional dataset used for step size estimation, performing multi-feature-driven step size estimation to obtain the step size, step frequency, and orientation angle used for predicting pedestrian trajectory estimation and PDR positioning. The hierarchical GRU neural network model includes one GRU layer, two LSTM layers, and one regression layer. In the hierarchical GRU neural network model, the segmented multi-feature inertial data sequence is first input into the LSTM layer for local feature extraction. Secondly, GRU is used to further extract features from the inertial data. Then, the extracted features are passed to the last LSTM layer to extract global time-series information. S104. Based on the step size, step frequency, and orientation angle used to predict the pedestrian trajectory for PDR positioning, as well as the predicted geomagnetic positioning results, geomagnetically assisted multi-feature driven pedestrian trajectory PDR positioning is performed through particle filtering to obtain the final position of the target.

2. The method as described in claim 1, characterized in that, The creation of the multidimensional dataset for geomagnetic positioning specifically includes: A geomagnetic dataset containing 5-dimensional geomagnetic components is constructed, as shown in Equation (1): （1） in, This represents the three-axis magnetic field obtained by the magnetometer in the mobile terminal. and Let the amplitude and horizontal components represent the amplitude and horizontal components, respectively, obtained from formula (2): （2）。 3. The method as described in claim 2, characterized in that, After constructing the geomagnetic dataset containing 5-dimensional geomagnetic components, the method further includes: The noise during data acquisition is reduced by using a median average filtering method with a predetermined window size. Specifically, the maximum and minimum values ​​of the geomagnetic sequence of each geomagnetic component in the window are removed, and then the remaining data is averaged. The geomagnetic dataset is Z-score standardized as shown in formula (3): （3） in, , These represent the mean and standard deviation of the geomagnetic dataset, respectively. It is a standardized geomagnetic dataset.

4. The method as described in claim 1, characterized in that, Sensor information includes acceleration, gyroscope data, and orientation data; The creation of the multidimensional dataset for step size estimation specifically involves: Construct a multi-feature dimension As the step size dataset used for step size estimation, the acceleration and gyroscope measurements are shown in Equation (4): （4） in, , , , , , These are the x, y, and z components of the acceleration and gyroscope, respectively. , The acceleration and the magnitude of the gyroscope are obtained from equation (5): （5） Normalize the step size dataset. （6） in, , These are the minimum and maximum values ​​of the dataset; The step length dataset is segmented according to the step frequency, with each segment corresponding to a step length. The segments are generated by a sliding window, as shown in formula (7): （7） Where i is the size of the sliding window, and N is a natural number greater than 1.

5. The method as described in claim 1, characterized in that, In S102, ResNet34 uses a shortcut connection architecture to learn the required mappings. ResNet34 includes Conv1, Conv2_x, Conv3_x, Conv4_x, and Conv5_x. Conv1 consists of a 7×7 convolutional layer and a max-pooling layer. Conv2_x, Conv3_x, Conv4_x, and Conv5_x consist of 3, 4, 6, and 3 residual blocks, respectively. Each residual block contains two convolutional layers and consists of multiple levels of residual mappings and identity mappings. The segmented magnetic sequence... Send to ResNet34; In residual mapping, for layer The output is: （8） in, It is the first Magnetic sequence input of the layer, and It refers to the weights of each layer. The activation function for the residual mapping network; Output of residual block The implicit features are: （9） Then, the extracted hidden features The data is sent to the GRU layer for further extraction of the temporal features of the magnetic subsequence. The GRU's update gate controls the ratio of past magnetic field information to current information, while the reset gate controls how much of the previous geomagnetic state needs to be stored. The latent features are input at time t. In this case, the equations for GRU are defined as follows: （10） in, , These are the update door and the reset door, respectively. It is based on the hidden state calculated by resetting the gate. It is the output of the neuron at time t-1. It is the output of the neuron at time t. , , It is a weight matrix. , , Represents the deviation vector. It is the sigmoid function; The basic unit of LSTM has three gate structures: forget gate, input gate, and output gate. The input magnetic signature data at time t... Unit status and the output of the previous neuron Together, they determine the forgotten portion of the state memory unit. The formula for LSTM is as follows: （11） in, , , , and These represent the input gate, forget gate, output gate, storage unit state, and current neuron output, respectively. , and These represent the weights of the input gate, forget gate, and output gate, respectively. , , These represent the biases of the input gate, forget gate, and output gate, respectively. It is a bias in the cell state. It is an activation function; Between the GRU layer and the LSTM layer, a Dropout layer is used to randomly drop a preset number of neurons. Finally, the regression layer consists of two fully connected layers to output the predicted pedestrian location. During model training, the proposed ResNet-GRU-LSTM neural network model is trained using a multidimensional dataset for geomagnetic positioning and relevant location labels. After training, the predicted geomagnetic positioning result at time t+1 can be obtained by inputting test data. , .

6. The method as described in claim 1, characterized in that, In S103, there is a Dropout layer after each layer; In the hierarchical GRU neural network model, the segmented multi-feature inertial data sequence is first input into the LSTM layer for local feature extraction. As the first layer, the LSTM selectively extracts implicit leading information from the input inertial data. Further extract features from inertial data using GRU; The extracted features are passed to the last LSTM layer to extract global time series information; The extracted features are randomly discarded and fed into the regression layer, which contains two fully connected layers with sigmoid activation functions. The regression layer maps the feature vectors of the inertial data sequence to the relevant pedestrian stride length, thereby establishing a mapping relationship between inertial data and stride length. During the training phase, the training dataset and the corresponding step size labels are input into the hierarchical GRU neural network model for training. After training is complete, the step size can be predicted by inputting unlabeled test data; Multi-feature driven step size estimation is characterized as follows: （12） in, It is a segmented test step size dataset. It is a step-size estimation model, the H-GRU neural network; The step length, step frequency, and orientation angle are obtained from the H-GRU model and used to predict pedestrian trajectory and calculate PDR positioning.

7. The method as described in claim 1, characterized in that, S104 specifically includes the following steps: S1041, Initialize particles; S1042. Utilize pedestrian trajectory to estimate PDR positioning, establish a particle motion model, and update particle positions; S1043. Reassess the particle weight based on the geomagnetic positioning results; S1044. Normalize the particle weights based on the re-evaluated particle weights, and resample the particles based on the weight values. S1045. Estimate the fused localization based on the updated particle positions and normalized particle weights; By repeatedly executing S102, S103, and S104, continuous estimation of pedestrian positions can be achieved.

8. The method as described in claim 7, characterized in that, S1041 specifically refers to: generating the initial particle set at time t. In this system, each particle has an equal weight, and the initial particle set is generated by magnetic positioning. For particles The estimated geomagnetic location at time t; S1042 specifically refers to: the position of the particles at time t+1 after obtaining the initial particle set. Update according to formula (13): （13） in, , These represent the step size and heading at time t+1, respectively. The noise follows a Gaussian distribution with a mean of zero and a variance of 1. S1043 specifically refers to: assuming the geomagnetic estimation of the system observation values ​​at time t+1. As shown in equation (14): （14） in, , This is the geomagnetic positioning result at time t+1. To observe Gaussian noise; Based on the observations, the particle weight is reassessed as follows: （15） in, It is the covariance matrix of the geomagnetic estimation; S1044 specifically refers to: particle weights normalized using formula (16): （16） Resample the normalized particles, focusing on particles with high weights that are close to the true state, and replicate them; reset the particle weights to ensure that the sum of all weights is 1; S1045 specifically refers to: the fusion localization at time t+1 is estimated using formula (17): （17） in, These are the normalized particle weights. It represents the position of the updated particle at time t+1.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the neural network geomagnetic positioning method based on multiple features as described in any one of claims 1 to 8.

10. A computer device, comprising: One or more processors; Memory; as well as One or more computer programs, wherein the processor and the memory are connected via a bus, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, characterized in that, when the processor executes the computer programs, it implements the steps of the multi-feature-driven neural network geomagnetic positioning method as described in any one of claims 1 to 8.

Images