Fusion positioning method of HCBiGRU hybrid neural network

Abstract

The invention provides a positioning method of an HCBiGRU hybrid neural network. The method comprises the steps of collecting sensor data by holding a smart phone by a pedestrian, and preprocessing the data. The method comprises the following steps of: acquiring acceleration data, performing positioning estimation by using an HCBiGRU neural network model, performing step length estimation by using acceleration data and a nonlinear adaptive step length estimation model, performing course estimation by using a quaternion method, and obtaining an estimated position through track plotting. Finally, geomagnetic estimation serves as an observation model, track plotting estimation serves as a measurement model, and positioning estimation is achieved through particle filtering.

Description

Technical Field

[0001] The present invention relates to the field of positioning, and in particular to a fusion positioning method of a HCBiGRU hybrid neural network. Background Art

[0002] With the development of global satellite navigation systems, outdoor navigation and positioning systems have become increasingly mature. However, most indoor environments block GNSS signals. In this case, it is difficult to use GNSS signals to meet the needs of indoor positioning, or even impossible to locate. At the same time, the indoor environment is complex and changeable, and the study of robust and universal indoor positioning methods is more important for location services.

[0003] Common indoor positioning methods include but are not limited to: Bluetooth positioning, RFID positioning, Wi-Fi positioning, and ultra-wideband positioning. The principle of these positioning methods is to use the propagation of wireless signals to determine the distance to each signal transmitter based on the strength of the received signal, thereby estimating the location. The advantage of the above-mentioned indoor positioning methods is that they can be positioned using existing indoor wireless networks. The disadvantage is that the positioning accuracy of these methods alone is insufficient and it is difficult to meet the needs of indoor positioning applications.

[0004] Geomagnetic positioning technology is an indoor positioning technology based on fingerprint matching. Its advantages are that it does not require additional infrastructure and can maintain stable geomagnetic signals for a long time. Therefore, geomagnetic positioning technology has become a hot topic in the field of positioning research. At the same time, existing smart devices are equipped with various sensors, such as accelerometers, magnetometers, gyroscopes, etc. In the positioning algorithm, track calculation has high accuracy in the short term, and geomagnetic positioning can maintain relatively stable accuracy in the long term. The present invention uses particle filtering to fuse geomagnetism and track calculation to achieve complementary advantages. Summary of the invention

[0005] The present invention proposes a fusion positioning method of HCBiGRU hybrid neural network. The method does not require additional infrastructure and can achieve low-cost and high-precision positioning.

[0006] The fusion positioning method of the HCBiGRU hybrid neural network of the present invention comprises the following steps:

[0007] Step 1: Use a mobile terminal to collect geomagnetic data, acceleration, gyroscope and azimuth data, and pre-process the data;

[0008] Step 2: Frame the preprocessed geomagnetic data;

[0009] Step 3: Construct the deep learning neural network of HCBiGRU;

[0010] Step 4: Input the framed geomagnetic data into the HCBiGRU network model for geomagnetic positioning estimation;

[0011] Step 5: Use a nonlinear adaptive step length estimation model to estimate the step length;

[0012] Step 6: Use quaternion method to estimate heading;

[0013] Step 7: After obtaining the heading estimation and step length estimation results, perform positioning estimation through dead reckoning;

[0014] Step 8: Use particle filtering to achieve fusion positioning of geomagnetic estimation and dead reckoning estimation.

[0015] During the execution of step 1, geomagnetic data and acceleration, gyroscope, azimuth and other data are collected using a smartphone with data collection software installed, and these data are pre-processed.

[0016] During the execution of step 2, the geomagnetic data is extracted and framed using a sliding window. The framed geomagnetic data is input into the constructed HCBiGRU network model for positioning estimation.

[0017] During the execution of step 3, a HCBiGRU neural network model is constructed, wherein the model includes one convolution layer, one pooling layer, two BiGRU layers, and one LSTM layer.

[0018] Step 4, using the network structure configuration set in step 3, perform indoor positioning estimation based on geomagnetism. Further, the geomagnetic positioning estimation includes: firstly, framing the data, constructing a HCBiGRU neural network, the network model includes a layer of CNN, two layers of BiGRU and a layer of LSTM, and inputting the data into the constructed network model for prediction and estimation, thereby realizing target positioning.

[0019] During the execution of step 5, a nonlinear step length model is used for step length estimation, and the model can adaptively update the step length according to the change of acceleration.

[0020] During the execution of step 6, the heading is estimated using the quaternion method, which can obtain the heading value through coordinate conversion.

[0021] 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 5, 7, and 8, and finally the estimated position is obtained.

[0022] During the execution of step 8, the fusion positioning method based on particle filtering integrates geomagnetic positioning and dead reckoning positioning. In this positioning method, HCBiGRU is constructed for geomagnetic positioning, the geomagnetic positioning result is used as the observation value of the particle filter, and the dead reckoning positioning result is used as the state value.

[0023] The present invention provides a fusion positioning of a HCBiGRU hybrid neural network. First, acceleration, gyroscope and azimuth data are collected and preprocessed. Then, the HCBiGRU network model is used for geomagnetic positioning, which can extract spatial and temporal features and reduce geomagnetic positioning errors. Then, dead reckoning positioning estimation is performed. Finally, a particle filter is used to fuse geomagnetic positioning and dead reckoning positioning to obtain a final position, thereby achieving low-cost and high-precision positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the overall technical framework diagram of the present invention;

[0025] Figure 2 It is a test plane diagram of the positioning method of the present invention that combines the HCBiGRU neural network and dead reckoning;

[0026] Figure 3 This is a comparison chart of the cumulative positioning error probability using different positioning algorithms on the vivo Y93s smartphone;

[0027] Figure 4 This is a comparison chart of the cumulative positioning error probability of different positioning algorithms using the Xiaomi 10 smartphone;

[0028] Figure 5 It is a flow chart of a positioning method of a HCBiGRU network provided by the present invention. DETAILED DESCRIPTION

[0029] 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 intended to explain the present invention and should not be construed as limiting the present invention.

[0030] See also Figures 1 to 5 The present invention provides a fusion positioning method of a HCBiGRU hybrid neural network, comprising the following steps:

[0031] S1: Data acquisition and preprocessing. Use the data acquisition software installed on the smart device to collect geomagnetic signals and IMU data, and preprocess the data; use the empirical mode decomposition (EMD) method to reduce noise.

[0032] S2: The geomagnetic data is framed and processed using a sliding window, as shown in (1).

[0033] D i ={mag i ,i=1,....,N} (1)

[0034] Where i is the sliding window size and mag is the geomagnetic data.

[0035] S3: Construct a HCBiGRU hybrid neural network model. The hybrid neural network model has eight layers. The first layer is a CNN layer, which is used to extract local features of geomagnetic data. The second and fourth layers are BiGRU layers, which mine the forward and reverse information of geomagnetic sequences. The third, fifth and seventh layers are Dropout layers with a dropout rate of 0.2 to avoid overfitting. The sixth layer is an LSTM layer, and the last layer is a Dense layer for output.

[0036] S4: Construct a HCBiGRU hybrid neural network model for geomagnetic positioning estimation.

[0037] First, the geomagnetic data is placed in the CNN layer for convolution to extract local features in the geomagnetic data. The extracted features are then input into the BiGRU part of the lower layer as the input of the BiGRU, which is used to obtain the long-term dependent features in the entire geomagnetic data. The Dropout layer is used to combat overfitting by randomly deleting a portion of neurons. It simplifies the neural network structure to some extent, forcing it to learn data features in a more efficient way. Finally, it is placed in the LSTM layer to learn the relationship between the geomagnetic sequence and the location label of each geomagnetic component, and the test data set used for geomagnetic positioning is placed in the trained HCBiGRU neural network model to predict the position and obtain the predicted geomagnetic positioning result.

[0038] In the present invention, the CNN structure is mainly composed of a data input layer, a convolution layer, and a pooling layer. The convolution layer is used to extract the features of the input geomagnetic data; the pooling layer is used to reduce the number of parameters and shorten the training time. The formula used by the convolution layer to extract geomagnetic features is:

[0039] c i =f(Σw i *mag i +b i ) (2)

[0040] Among them, c i is the output of the convolutional layer, f is the activation function, mag is the input geomagnetic value, w i is the convolution kernel weight, b i is the bias. Pooling retains the most significant features in the local area and has good translation invariance. Its formula is

[0041] y i =max(mag i:i+k ) (3)

[0042] where y i Represents the output feature value.

[0043] In order to further extract the feature dependency relationship between the previous and next moments of geomagnetic data, this paper uses BiGRU (Bi-directional Gated Recurrent Unit, BiGRU) to mine the forward and reverse information of the geomagnetic sequence. The BiGRU network model can use the historical data and future time series data in the positioning data to feedback the parameters of the correction model, providing historical position information, direction information, etc. for the current moment position prediction, thereby outputting more accurate position information. Structurally, BiGRU is a bidirectional recurrent network composed of two GRUs with opposite propagation directions; the output of each BiGRU unit at time t is determined by the bias at time t and the forward propagation BiGRU output. And the back-propagation BiGRU output The three parts have a joint influence, and the specific calculation process is as shown in the formula.

[0044]

[0045] α and β are the corresponding hidden layer output weights; b is the corresponding hidden layer bias.

[0046] S5: Perform step length estimation. The nonlinear adaptive step length estimation model is used to reflect the continuity and similarity of the walking process. The model is not only related to the peak acceleration estimation of the current step, but also takes into account the step lengths of the previous two steps. The step length is adaptively updated according to the change of acceleration. Its characteristics are as follows:

[0047]

[0048] Where K is the model parameter, l i-2 , l i-1 are the step lengths at the i-2 and i-1 steps respectively, i refers to the current step number, is the weight vector, bias is the offset error measured by different mobile terminals under static conditions, obtained from empirical values, and A i peak and A i valley are the peak value and valley value of acceleration in the i-th step respectively.

[0049] S6: Estimating the heading and using the quaternion method to perform coordinate transformation. The specific method is to rotate the data in the carrier coordinate system along the three coordinate axes in a certain order, and finally obtain the corresponding data in the navigation coordinate system. The angles of rotation along the X-axis, Y-axis and Z-axis are called the roll angle γ, pitch angle θ and heading angle φ respectively.

[0050] S7: Pedestrian dead reckoning positioning estimation. Dead reckoning uses the data obtained by the inertial sensor combined with the position information at the previous moment to calculate the relative position of the pedestrian to achieve positioning;

[0051] S8: Positioning estimation. Particle filtering is used to implement fusion filtering of geomagnetic positioning and dead reckoning positioning results. The geomagnetic positioning result is used as the observation value of the particle filter, and the dead reckoning positioning result is used as the state value.

[0052] (1) State initialization: Before positioning, a certain number of particles P need to be generated in the positioning area. i (x i ,y i ,w i ), where x i ,y i ,w i are the (x, y) coordinates of the generated particles and the corresponding weights in this state. The particle weights are averaged during initialization, i.e., w i =1 / n (n is the number of particles generated).

[0053] (2) Particle state update: After the particle is initialized, the particle state is updated when a new positioning request is received. i (x i ,y i ,w i ) Update P i (x i ,y i ,w i ). During the update process, the position coordinates and weights of the particles are mainly recalculated. The position update mainly uses PDR to infer the position of each particle based on the motion trajectory of the pedestrian. The calculation formula is shown as follows, where θ is the angle of the forward direction and β is the noise in the position update process.

[0054]

[0055] The particle update process not only updates the particle position, but also requires recalculating the weight carried by the particle.

[0056]

[0057] where v is the variance of the observations.

[0058] (3) Normalization: After the state update equation is processed, the particle position and weight have changed. In order to ensure that the positioning result does not exceed the limit, it is necessary to normalize the particles and calculate the new weight value of the particles, that is:

[0059]

[0060] (4) Particle evaluation and resampling: After multiple state updates, some particles will leave the positioning area, causing a significant reduction in the number of particles. In order to ensure the diversity of particles during the update process, it is necessary to resample the particles to ensure the accuracy of the positioning results. The most common method for particle resampling is to copy particles, that is, copy particles with larger weights. The larger the weight, the more times they are copied, and the smaller the weight, the fewer times they are copied. Particles with too small weights are directly eliminated.

[0061] (5) Estimated positioning results: Perform weighted summation on the resampled particles and calculate the positioning result using the formula.

[0062]

[0063] By cyclically executing steps S2-S8, continuous estimation of the position of the pedestrian can be achieved.

[0064] As described in the above technical solution, the present invention has the following beneficial effects:

[0065] The present invention realizes a positioning method integrating geomagnetic and inertial sensors. The present invention proposes a HCBiGRU algorithm to realize geomagnetic positioning, thereby improving the accuracy of indoor positioning. Pedestrian dead reckoning has low complexity and can output reliable location information in a short time without relying on the construction of any external infrastructure. However, long-term cumulative errors will lead to reduced positioning accuracy. Geomagnetic positioning can maintain relatively stable accuracy for a long time. Therefore, geomagnetic positioning can be combined with dead reckoning to form complementary advantages. It can also solve the problem of geomagnetic mismatch to a large extent. Therefore, particle filtering is used in the present invention to fuse geomagnetic positioning and dead reckoning. By designing and implementing test cases in a typical indoor environment such as a laboratory corridor, and based on a pre-planned moving path, we have verified the performance of the proposed fusion positioning method. The experimental data clearly show that the method reduces the error of geomagnetic positioning, alleviates the cumulative error in the process of pedestrian dead reckoning, and weakens the influence of heterogeneity between different positioning devices on positioning accuracy. Overall, the method shows good performance and improves the accuracy and reliability of indoor positioning.

[0066] In order to verify the positioning performance of the proposed method, a test was carried out in an indoor environment of a rectangular corridor with an area of ​​34m×17.3m and a height of 4m. Figure 3 The red line in the figure represents the planned path. Figure 3 The path in the middle is a 92-meter-long rectangular corridor with fire extinguishers, steel cabinets, railings, and some steel doors. Pedestrians hold smartphones to collect geomagnetic intensity and inertial measurement unit data from acceleration and gyroscope sensors along the planned path. In addition, to ensure the reliability of the data, pedestrians

[0067] These data need to be collected multiple times. The sampling time is 20 milliseconds. In the experiment, vivo Y93s and Xiaomi10 smart devices are used to collect data.

[0068] HCBiGRU was tested and experimented with LSTM, CNN, DTW, MALOC, and GRU using different mobile phones in the same scenario.

[0069] Figure 4 and Figure 5 The probability diagrams of the cumulative positioning errors of different algorithms used by vivo Y93s and Xiaomi 10 are shown respectively. The experimental results show that the proposed algorithm has the smallest cumulative error. This is because the parameters of the model are corrected by the feedback of historical data and future time series data in the geomagnetic data of the proposed network model, which provides historical position information and direction information for the current moment position prediction. Thus, more accurate position information is output. The influence of different mobile phones on positioning accuracy is reduced, and the contradiction between positioning accuracy and low cost is solved.

[0070] Table 1 shows the mean and root mean square errors of LSTM, CNN, DTW, MALOC, GRU, and HCBiGRU in the planned scenes. The experimental results show that the proposed algorithm has the lowest error, so the proposed positioning system has better performance than several other algorithms.

[0071] Table 1. Comparison of mean and RMS errors of different algorithms using vivo Y93s and Xiaomi 10 in the experimental scenario

[0072]

[0073]

[0074] In order to overcome the position ambiguity and cumulative error of dead reckoning in geomagnetic positioning, this paper proposes a fusion positioning method of HCBiGRU hybrid neural network. First, in this positioning method, we propose a HCBiGRU model, which uses CNN to extract local features and BiGRU to extract forward and backward information of geomagnetic data, while LSTM is used to extract global information, thereby improving the indoor positioning accuracy of geomagnetism. The particle filter is used to combine geomagnetic positioning and PDR positioning to achieve fusion positioning, which overcomes the shortcomings of single positioning technology. The experimental results of the proposed HCBiGRU fusion positioning system show that the method can obtain higher positioning accuracy when users use different devices. This positioning method can effectively reduce the error, and the positioning performance has strong robustness.

[0075] The above results only show an implementation of a fusion positioning method of a HCBiGRU hybrid neural network. Of course, this 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 implemented above and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A positioning method of HCBiGRU hybrid neural network, characterized in that: The following steps are involved: Collect geomagnetic data, acceleration, gyroscope and azimuth data, and perform pre-processing operations on the data; The pre-processed geomagnetic data is framed; Construct a deep learning neural network of HCBiGRU; The framed geomagnetic data is input into the HCBiGRU network model for geomagnetic positioning estimation; A nonlinear adaptive step length estimation model is used for step length estimation; Use quaternion method to estimate heading; After obtaining the heading estimation and step length estimation results, the positioning estimation is performed through dead reckoning; Particle filtering is used to realize the fusion positioning of geomagnetic estimation and dead reckoning estimation.

2. The method according to claim 1, characterized in that Data collection is implemented on the user side, with pedestrians holding smart terminals to collect gyroscope, accelerometer, magnetometer, and geomagnetic data and send them to the server. On the server side, geomagnetic, acceleration, gyroscope, and heading angle data are extracted from the uploaded data.

3. The method according to claim 1, characterized in that The geomagnetic positioning estimation includes: firstly, framing the data, constructing a HCBiGRU neural network, the network model includes a layer of CNN, two layers of BiGRU and a layer of LSTM, inputting the data into the constructed network model for prediction and estimation, thereby realizing target positioning.

4. The method according to claim 1, characterized in that The nonlinear adaptive step length estimation model can adaptively update the step length according to the change of acceleration.

5. The method according to claim 1, characterized in that The quaternion method can obtain the heading value through coordinate conversion.

6. The method according to claim 1, characterized in that The particle filter is used to realize the fusion positioning of geomagnetic estimation and dead reckoning estimation, and its steps include: firstly, particle initialization, secondly, particle state update and weight normalization, and finally, obtaining the final positioning result.

Images