A method and system for GPS enhanced positioning in urban canyons based on deep learning

Through deep learning technology, the environment, satellite state and time data of GPS positioning in urban canyons are processed, and GPS positioning errors are predicted and corrected, which solves the problem of GPS positioning errors in urban canyons, improves positioning performance and reliability, and avoids expensive hardware and the dependency of precise 3D models.

CN114814914BActive Publication Date: 2024-11-22SHENZHEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210425764.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-22
Publication Date
2024-11-22
Estimated Expiration
2042-04-22

AI Technical Summary

Technical Problem

In urban canyon environments, GPS positioning errors are caused by NLOS signals, and existing methods rely on accurate 3D urban models or expensive hardware and are difficult to widely use.

Method used

Using a deep learning-based method, by obtaining GPS accuracy influencing factors in urban canyons, including environmental data, satellite state data and positioning time data, converting them into a matrix of the same size, inputting the trained encoder and decoder, and predicting And correct the GPS positioning error to achieve more accurate positioning.

Benefits of technology

Using deep neural network models to learn the context state of GPS positioning from a large number of labeled GPS samples, accurately predict error positioning points, improve the positioning performance and reliability of GPS in urban canyons without relying on expensive hardware or accurate 3D models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114814914B_ABST
    Figure CN114814914B_ABST
Patent Text Reader

Abstract

The present invention relates to a method and system for enhancing GPS positioning in urban canyons based on deep learning. The method comprises: taking multiple factors (environment, time, GPS satellites, etc.) that affect GPS accuracy in urban canyons as the context state of GPS positioning points; a deep neural network model predicts positioning errors through the context state of GPS positioning points; if the positioning error is greater than the acceptable error, the deep neural network model predicts the corrected GPS positioning position according to the context state of the GPS positioning point; maintaining a mobile model to estimate the user's movement speed, limit the user's reachable area, and help determine the correct GPS positioning position to assist in correction. The present invention can predict the positioning error of GPS positioning points and quickly and effectively correct the positioning position of GPS, improve the positioning performance of GPS in urban canyons, and enhance the reliability of GPS.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of navigation and positioning technology, and more specifically, to a method and system for urban canyon GPS enhanced positioning based on deep learning. Background Art

[0002] Navigation and positioning technology is widely used in people's daily lives, such as personal navigation, online taxi-hailing, and terrain surveying and mapping. The most common navigation system is the GPS navigation system (Global Positioning System).

[0003] The GPS navigation system consists of three parts: a satellite constellation, ground base stations, and the user's GPS receiver. Specifically, the satellite constellation consists of 32 satellites, each of which orbits the earth approximately every 12 hours and continuously broadcasts the satellite's own position and current time above the earth. The satellites carry stable atomic clocks that are precisely synchronized with the clocks of the ground stations. The ground base station monitors and manages the status of all satellites by sending parameters that control their orbits and trajectories to all satellites. The GPS receiver receives signals from existing satellites and calculates the propagation distance of each signal by multiplying the speed of light by the propagation time of the signal, namely the pseudorange (the approximate distance between the GPS receiver and the satellite measured by the GPS receiver). The position of the GPS receiver on the earth can be calculated by combining the position of the satellite broadcast and the pseudorange with each satellite.

[0004] Although GPS navigation systems work well in most places, they are still very unreliable in urban canyons. GPS positioning errors in urban canyons mainly come from multipath interference (Multipath Interference) and non-line of sight (NLOS) signals caused by reflections of satellite signals from high-rise buildings. Although there are a lot of hardware and software that can solve the problem of multipath interference well, the NLOS satellite signals on GPS receivers can still cause large positioning errors in urban canyons. In urban canyon scenarios, the signals of NLOS satellites may be reflected by one or more high-rise buildings, resulting in larger pseudoranges and thus positioning errors.

[0005] To address the errors caused by NLOS signals, some current methods use dedicated hardware (such as panoramic cameras) and LiDAR to reconstruct the reflective surfaces of street buildings, and apply ray tracing algorithms to satellite signals to calculate pseudorange errors (Wen, W., Zhang, G., & Hsu, LT (2019). Correcting NLOS by 3D LiDAR and building height to improve GNSS single point positioning. Navigation, 66 (4), 705-718.). In addition, there are some methods that use building geometry based on proprietary 3D city models to calculate the similarity of signal paths or satellite visibility between candidate locations around the GPS initial position and receiver observations, and output the best matching point as the correction point (Liu, X., Nath, S., & Govindan, R. (2018, June). Gnome: A practical approach to nlosmitigation for gps positioning in smartphones. In Proceedings of the 16th Annual International Conference on Mobile Systems, Applications, and Services (pp. 163-177).). However, these methods are heavily dependent on accurate 3D city models or other third-party resources, such as panoramic images, and some methods even require expensive hardware equipment. These resources are not easy to obtain, which greatly limits their practical application. Summary of the invention

[0006] The present invention provides a method and system for GPS enhanced positioning in urban canyons based on deep learning.

[0007] In order to achieve the above technical objectives, the present invention specifically adopts the following technical solutions:

[0008] A method for GPS enhanced positioning in urban canyons based on deep learning, comprising:

[0009] Obtaining data on factors affecting GPS accuracy in urban canyons, including environmental data, satellite status data, and positioning time data, and converting them into matrices of the same size; the environmental data includes building heights and road distribution data around the initial GPS position;

[0010] Inputting the converted multiple matrices of the same size into a trained encoder to obtain a feature map encoding environment, time and satellite status information; the encoder is an encoder based on a residual network block;

[0011] Input the feature map into a trained distance decoder to obtain a predicted value of the positioning error of the original GPS positioning point;

[0012] The predicted positioning error value is compared with a preset error threshold. If the predicted positioning error value is greater than the acceptable error value, the feature map is input into a trained position decoder to obtain the corrected positioning position of the original GPS positioning point.

[0013] As a preferred implementation, the method of converting the influencing factor data into a matrix form is:

[0014] a. For environmental data, a square area is constructed with the original estimated GPS position as the center, and the area is divided into a number of square grid cells. The square grid cells are assigned values ​​according to the environment they belong to, and the environmental matrix is ​​constructed;

[0015] For example, if a square grid cell corresponds to a building, the building height is assigned to it; if a square grid cell corresponds to a road, it is assigned 0; the remaining grid cells are assigned -1, indicating that they are inaccessible;

[0016] b. For satellite status data, draw a sky map based on the altitude angle, azimuth angle and signal-to-noise ratio of the GPS satellite; construct a sky map matrix of the same size as the environment matrix, map the satellites in the sky map to the sky map matrix, and assign values ​​to the matrix grid elements where there are satellites using the signal-to-noise ratio of the corresponding satellite;

[0017] c. For the positioning time data, the timestamp at the time of positioning is converted into an n-dimensional vector, and the n-dimensional vector is converted into a matrix of the same size as the environment matrix and the sky map matrix, i.e., the timestamp matrix, using a multilayer perceptron. There are many factors that affect the accuracy of GPS in urban canyons, such as the layout and height of buildings, satellite distribution, etc., and these factors have different forms, resulting in different dimensional representations. How to represent and fuse these factors in a suitable form as input for a deep neural network without losing information is a huge challenge. The present invention converts all factors into matrices of the same size in a specific way, and combines them into a whole multi-source data input. Various factors are represented in the form of a matrix, thereby realizing the application of the technical problem described in the present invention in a neural network model.

[0018] As a preferred implementation, the satellites in the sky map are mapped to the sky map matrix as follows:

[0019] The central element of the sky map matrix corresponds to the center of the sky map. The distance between the satellite element and the central element on the sky map matrix is ​​determined according to the satellite's altitude angle. The clockwise angle between the line connecting the satellite element and the central element and the "directly above" direction is determined according to the satellite's azimuth angle to map the satellite in the sky map to the sky map matrix.

[0020] As a preferred implementation, the timestamp during positioning is converted into a 7-dimensional vector, and the 7 elements are:

[0021] Element 1: an element that characterizes the progression of location time in a natural day;

[0022] Element 2: An element that characterizes the progress of the satellite in a revisit cycle;

[0023] Elements 3 to 6: Apply sine and cosine calculations to elements 1 and 2, respectively, to assist in finding periodic elements;

[0024] Element 7: an element representing that the positioning time is before or after 12:00am on the same day in the time zone of the positioning location;

[0025] The 7-dimensional vector is obtained by combining the 7 elements in any order.

[0026] The present invention takes into account the natural daily cycle and the revisit period in the 7-dimensional vector, and uses sine and cosine functions to assist in finding periodic features.

[0027] As a preferred implementation, the encoder, distance encoder, and position encoder are trained using the marked GPS positioning point data, the corresponding road network, building measurement data, and satellite status data as training samples;

[0028] The training target of the distance decoder is a one-hot encoding vector, and the element value corresponding to the correct error interval is 1; the cross entropy loss function is used to measure the distance between the output of the distance decoder and the training target;

[0029] The training target of the position decoder is the Gaussian representation matrix of the true position, and the element corresponding to the true position in the Gaussian representation matrix is ​​the maximum value 1; the mean square error loss function is used to measure the distance between the output of the position decoder and the training target.

[0030] As a preferred implementation, the output of the position decoder is a probability matrix, and the final output of the position decoder is obtained after adopting the environment matrix constraint;

[0031] Set the elements in the environment matrix that have buildings or other inaccessible elements to 0, otherwise to 1, that is, valid candidate positions. The assigned environment matrix is ​​used as a constraint masker. The final output of the position decoder is the constraint masker dot product probability matrix. The element with the largest value in the final output matrix is ​​the element corresponding to the corrected GPS positioning position, and the longitude and latitude corresponding to the center of the element is the output corrected position.

[0032] As a preferred implementation, the encoder, distance encoder, and position encoder are trained together after being connected according to their input-output relationship during training. Joint training can enable the distance decoder and the position decoder to transmit each other's feature information through the association of the encoders, thereby achieving more accurate prediction.

[0033] As a preferred implementation, when the encoder, distance encoder, and position encoder are trained together, the overall loss function of the training model is set to

[0034] L ocerall =L dis ·λ+L pos

[0035] Where L dis is the loss function of the distance encoder, L pos is the loss function of the position encoder, and λ is the regularization parameter used to balance the impact of the distance decoder and the position decoder on the encoder.

[0036] As a preferred implementation, it also includes: when it is necessary to continuously locate the mobile positioning object, estimating the moving speed of the mobile positioning object, limiting the reachable area of ​​the mobile positioning object to assist in determining the correct GPS positioning position, including:

[0037] Calculate the average moving speed of the mobile positioning object based on the latest corrected GPS positioning positions and corresponding positioning times;

[0038] Construct a square area with the original estimated GPS position as the center, and divide the area into several square grid units;

[0039] The corrected GPS positioning position of the mobile positioning object corresponding to the previous positioning time is used as the center of the circle, and the average moving speed × the time difference between adjacent positioning times is used as the radius to draw a circle to construct a circular area;

[0040] According to the overlap between the square area and the circular area, the grid cells are assigned values. If the grid cells overlap with the circular area, a value of 1 is assigned; otherwise, a value of ρ (0≤ρ≤1) is assigned. The matrix formed after the assignment is called a moving masker.

[0041] The output of the motion masker is multiplied by the output of the position decoder to obtain the final probability matrix, and the longitude and latitude corresponding to the center of the grid cell with the largest value in the probability matrix is ​​output as the corrected position for continuous positioning.

[0042] Another object of the present invention is to provide an urban canyon GPS enhanced positioning system based on deep learning, comprising:

[0043] The terminal obtains the original GPS positioning point, positioning time and satellite status data of the positioning object, presets the required error threshold, and sends it to the server;

[0044] Server, including: database and deep learning model;

[0045] The database stores environmental data and historical calculation data; the environmental data includes building height and road distribution data;

[0046] The deep learning model performs positioning correction on the original GPS positioning point, including a multi-source data fusion module, an encoder, a distance encoder and a position encoder;

[0047] The multi-source data fusion module provides input parameters for the deep learning model. The multi-source data fusion module searches for environmental data corresponding to the original GPS positioning point and positioning time from the database according to the original GPS positioning point, positioning time and satellite status data sent by the terminal, and converts the environmental data, positioning time and satellite status data into matrices of the same size;

[0048] The encoder extracts features from the matrix data output by the multi-source data fusion module and outputs a feature map encoding the environment, time and satellite status information;

[0049] The distance encoder predicts the positioning error of the original GPS positioning point based on the feature map and the error threshold. If the positioning error prediction value is greater than the error threshold, the position decoder is called, otherwise the original GPS positioning point is not corrected;

[0050] The position decoder outputs the corrected positioning position of the original GPS positioning point based on the feature map, and sends the original GPS positioning point and the corrected positioning position to a database for storage.

[0051] As a preferred implementation, the server is a cloud server.

[0052] The present invention uses readily available information to implement GPS positioning correction in urban canyons, including environmental factors (building height and road distribution around the GPS initial position), time, and satellite status, which are used as the context state of GPS positioning, and uses a deep neural network model to learn the implicit complex relationship between the context state of GPS positioning and GPS positioning estimation from a large number of labeled GPS samples. The model in the present invention is well trained and can accurately predict the correct position of each erroneous GPS estimate based on the context state of the GPS initial position, thereby improving the positioning performance of GPS in urban canyons and enhancing the reliability of GPS. In addition, the deep neural network model trained using a large number of GPS samples is universal and can therefore serve a wide range of urban areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the method of the present invention.

[0054] Figure 2 It is how the environment matrix is ​​constructed.

[0055] Figure 3 This is how the sky map matrix is ​​constructed.

[0056] Figure 4 It is the schematic diagram of continuous positioning.

[0057] Figure 5 It is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0058] Example 1

[0059] A method and system for GPS enhanced positioning in urban canyons based on deep learning (such as Figures 1 to 5 shown).

[0060] S1: Convert various factors that affect GPS accuracy in urban canyons (building height and road distribution around the initial GPS position, time, and satellite status) into a matrix as the context state of the original GPS point; including:

[0061] S1.1: Convert the surrounding environment information of the original GPS positioning point into the environment matrix M e ,like Figure 2 shown.

[0062] (1) GPS receiver provides raw estimated position The longitude and latitude of .

[0063] (2) First select a A square area with a center and a side length of 2R is then divided into square grids of size c×c. Each grid in the square area corresponds to the environment matrix Me In the calculation example, R = 50cm, c = 1m.

[0064] (3) Set the environment matrix M according to the following rules e The value of each element in:

[0065] i. If the grid in the square is part of a building, then the environment matrix M e The value of the corresponding element in is set to the building height;

[0066] ii. If the grid in the square is part of the road, then the environment matrix M e The value of the corresponding element in is set to 0;

[0067] iii. Environmental matrix M e The values ​​of the remaining elements in are set to -1 (meaning that the cells in the corresponding squares are inaccessible).

[0068] S1.2: Convert the state of the GPS satellites used by the original GPS positioning point into a sky plot matrix M s ,like Figure 3 shown.

[0069] (1) The GPS receiver provides the status of each GPS satellite, namely the altitude (the angle between the line connecting the GPS receiver and the satellite and the horizontal plane), the azimuth (the horizontal angle between the GPS receiver's true north direction and the satellite in a clockwise direction), and the signal-to-noise ratio (SNR).

[0070] (2) Draw the corresponding sky plot based on the altitude angle, azimuth angle and signal-to-noise ratio.

[0071] (3) Structure and environment matrix M e The sky map matrix M of the same size s .

[0072] (4) The sky map matrix M s The central element of corresponds to the center of the sky map, and then the satellites in the sky map are mapped to the sky map matrix M s That is, the sky map matrix M of the satellite element and the central element is determined according to the satellite's altitude angle s The distance on the satellite is determined by the azimuth of the satellite, and the clockwise angle between the line connecting the element where the satellite is located and the center element and the "directly above" direction.

[0073] (5) If the sky map matrix M sIf there is a corresponding satellite on an element of , its value is set to the signal-to-noise ratio of the corresponding satellite, otherwise it is set to 0.

[0074] S1.3: Convert the positioning time of the original GPS positioning point into the timestamp matrix M t .

[0075] (1) The timestamp t when the GPS receiver provides positioning (referring to the Unix timestamp, the total number of seconds from 00:00:00 January 1, 1970 UTC to the present, excluding leap seconds).

[0076] (2) Convert the provided timestamp t into a 7-dimensional vector V t .

[0077] (3) Set the 7-dimensional vector V according to the following rules t The value of each element in:

[0078] i. The first element: V t [0] = (t% 86400) / 86400. The total number of seconds in a natural day (i.e. 24 hours) is 86400 seconds, and this element represents the progress of positioning time in a natural day.

[0079] ii. The second element: V t [1] = (t% 86280) / 86280. The total number of seconds in a revisit period (i.e. 23 hours and 58 minutes) is 86280 seconds. This element represents the progress of the satellite in one revisit period.

[0080] iii. The third element: V t [2] = cos(2*π*V t [0]). represents the t [0] The result of applying cosine helps to quickly find periodicity.

[0081] iv. The fourth element: V t [3] = sin(2*π*V t [0]). represents the t [0] The result of applying sine, which helps to quickly find periodicity.

[0082] v. The fifth element: V t [4] = cos(2*π*V t [1]). represents the t [1] The result of applying cosine helps to quickly find periodicity.

[0083] vi. The sixth element: V t [5] = sin(2*π*V t [1]). represents the t[1] The result of applying sine helps to quickly find periodicity.

[0084] vii. The seventh element: If the time t corresponds to the morning of the day (before 12:00am) in the time zone where the positioning location is located, this element is set to 1; otherwise, it is set to 0.

[0085] (4) Use a simple multi-layer perceptron (MLP) model to transform the vector V t Converted into a feature vector. The MLP model consists of a fully connected layer, with the input layer being V t as input; there are 3 hidden layers, the number of neurons in each layer is 100, 1000, and 2000 respectively, and ReLU is used as the activation function; finally, the derived feature vector is reconstructed into e and M s Matrices M of the same size t .

[0086] S2: Input the matrix converted in S1 into the trained encoder built based on the residual network block, and the encoder extracts the feature map from it for use by the distance decoder and position encoder;

[0087] The implementation method of the encoder based on the residual network block includes the following steps:

[0088] The input of the S2.1 encoder is three matrices M e 、M s and M t combination.

[0089] The S2.2 encoder consists of 4 convolutional layers and 4 residual network blocks in the following order:

[0090] (1) The number of convolution kernels is 3, the kernel size is 7*7, the stride is 1, the padding is 3, the normalization function is BatchNorm, and the activation function is Tanh;

[0091] (2) The number of convolution kernels is 64, the size is 7*7, the step size is 1, the padding number is 3, the normalization function is BatchNorm, and the activation function is ReLU;

[0092] (3) The number of convolution kernels is 128, the size is 3*3, the step size is 2, the padding number is 1, the normalization function is BatchNorm, and the activation function is ReLU;

[0093] (4) The number of convolution kernels is 256, the size is 3*3, the step size is 2, the padding number is 1, the normalization function is BatchNorm, and the activation function is ReLU;

[0094] (5) 4 residual network blocks.

[0095] S2.3 inputs the combination of the three matrices into the encoder, and after processing through 4 convolutional layers and 4 residual network blocks, it outputs a feature map that encodes the environment, time, and satellite status information.

[0096] S3: Input the feature map in S2 into the trained distance decoder, and the distance decoder predicts the positioning error of the original GPS positioning point through the feature map;

[0097] The implementation method of the range decoder for predicting positioning error comprises the following steps:

[0098] The input of S3.1 distance decoder is the feature map output by the encoder.

[0099] The S3.2 distance decoder consists of two convolutional layers and multiple fully connected layers in the following order:

[0100] (1) The number of convolution kernels is 512, the kernel size is 3*3, the stride is 3, the padding is 0, the normalization function is BatchNorm, and the activation function is ReLU;

[0101] (2) The number of convolution kernels is 1024, the kernel size is 3*3, the stride is 3, the padding is 0, the normalization function is BatchNorm, and the activation function is ReLU;

[0102] (3) Four fully connected layers with 2048, 1024, 512, and 256 neurons respectively, all using ReLU as the activation function.

[0103] The output of the S3.3 distance decoder is a probability vector V dist , where V dist Each element of corresponds to an error interval, and the value of the element represents a probability. The present invention sets an error interval to 2 meters, and the probability vector V dist is set to a 25-dimensional vector because more than 99% of GPS errors are less than 50 meters. Therefore, V dist The i-th element of V indicates that the error is within the range of (2*i,2*(i+1)] meters. dist The larger the value of an element, the greater the possibility that the positioning error value is within the error interval corresponding to the element.

[0104] S4: Compare the predicted positioning error in S3 (the error corresponding to the maximum probability in the probability vector) with the acceptable error value. If the predicted positioning error is smaller than the acceptable error value, no correction is required; otherwise, the feature map in S2 is input into the trained position decoder, and the position decoder predicts the original GPS corrected positioning position through the feature map.

[0105] The method for implementing a position decoder for predicting a positioning position comprises the following steps:

[0106] The input of S4.1 position decoder is the feature map output by the encoder.

[0107] The S4.2 position decoder consists of 6 residual network blocks, 2 deconvolutional layers and 1 convolutional layer in the following order:

[0108] (1) 6 residual network blocks.

[0109] (2) The number of deconvolution kernels is 128, the size is 3*3, the step size is 2, the padding number is 1, the normalization function is BatchNorm, and the activation function is ReLU;

[0110] (3) The number of deconvolution kernels is 64, the size is 3*3, the step size is 2, the padding number is 1, the normalization function is BatchNorm, and the activation function is ReLU;

[0111] (4) The number of convolution kernels is 1, the size is 7*7, the step size is 1, the padding number is 3, the normalization function is BatchNorm, and the activation function is Sigmoid;

[0112] S4.3 The output of the above operation is a matrix that is equal to the input matrix (M e 、M s and M t ) has the same length and width matrix R posi , R posi Each element in is related to the environment matrix M e In order to avoid the gradient disappearance during the model training process, the present invention will R posi The corresponding target matrix is ​​set to Gaussian Peak. Therefore, R posi is a probability matrix, R posi The larger the value of an element is, the greater the possibility that the real location of GPS positioning is at the position corresponding to the element is.

[0113] S4.4 The matrix R output by the above operation posi Need to pass through the constraint masker Processing. Constraint Masker With matrix R posi The same length and width encodes information about the surrounding environment. Specifically, if a location is inaccessible due to the presence of buildings or other reasons, The corresponding element is set to 0; otherwise, the element is 1, indicating that the position corresponding to the element is a valid candidate position. Therefore, by constraining the masker, inaccessible cells can be filtered out.

[0114] The final output of the S4.5 position decoder is Where ⊙ is the dot product of the matrix. env The element with the largest median value is considered to be the corresponding element of the corrected GPS positioning position, and the longitude and latitude corresponding to the center of the element is used as the corrected position output of this method.

[0115] S5: If continuous positioning is required, the maintained mobility model is used to estimate the user's movement speed and limit the user's reachable area to assist in determining the correct GPS positioning position.

[0116] The implementation method of the mobile model for estimating the user's motion speed includes the following steps:

[0117] S5.1 If continuous positioning is required, a movement model for estimating the user's movement speed is used.

[0118] S5.2 first estimates the user's movement speed. Using the user's most recent k corrected GPS positioning positions p′, i.e., position (p′ i-1 , ..., p′ i-k ) and the corresponding positioning time (t i-1 ,……,t i-k ) to infer the user's movement. For simplicity and generality, this method only calculates the user's average movement speed v, because other movement information, such as movement direction, requires additional sensor data. For a given user, this method continuously updates the average speed v using the latest k positioning positions (k is set to the number of available GPS positioning positions) through the following formula:

[0119]

[0120] Among them, L(p′ i-j ,p′ i-j-1 ) returns the position p′ i-j and position p′ i-j The distance between.

[0121] S5.3 After obtaining the user's average moving speed v, it is necessary to estimate the position Correction of the constructed motion constraint: GPS-based estimated position Construct a As the center, the square area with a side length of 2R is then divided into square grids of size c×c. In addition, we construct a grid with p′ i-1 The circle with the center at S is S=v×Δt, where Δt is Positioning time t i and Positioning time t i-1 The circular area represents the area that the user can reach based on his / her recent moving speed.

[0122] S5.4 Combined with GPS to estimate location And the reachable area, that is: the square grid overlapping with the above circular area is the correct position p i The candidate unit that is most likely to be located, while other square grids that do not overlap with the circular area are unlikely to be the correct position p i This method constructs a matrix called a moving mask Each element corresponds to a grid in the square area. If the grid is (or partially) overlapped with the circular area, the value of its corresponding element is set to 1; otherwise, the value of its corresponding element is set to ρ (0≤ρ≤1). The value of an element is ρ, which means the correct position p i The confidence of the grid corresponding to this element is ρ. Finally, correction is performed according to the reachable area, that is, the moving mask With R env Multiply them together and get the final probability matrix Finally, R mob The longitude and latitude corresponding to the center of the element with the largest median value are output as the corrected position for continuous positioning by this method.

[0123] The MLP model, encoder, distance decoder, and position decoder all need to be trained to work. This method also includes instructions for training the encoder, distance decoder, and position decoder:

[0124] (1) Loss function of distance decoder:

[0125] The training target V of the distance decoder true is a one-hot encoding vector, where the element value corresponding to the correct error interval is 1, and the rest is 0. Since we model the problem of predicting positioning error as a classification problem, we use cross-entropy loss to measure the output V of the distance decoder dist With target V true Distance between:

[0126]

[0127] The loss function L of the distance decoder dis As shown above, b i is a binary indicator (if the i-th interval is the correct error interval, then b i =1, otherwise b i =0), p i is the Softmax probability of the i-th interval, n is V dist By default, n is set to 25.

[0128] (2) Loss function of position decoder:

[0129] The training target R of the position decoder is true is the Gaussian representation matrix of the ground truth position, where the element corresponding to the ground truth position is the maximum value 1, and the values ​​of the surrounding elements gradually decrease as the distance from the Gaussian peak increases. Therefore, the mean square error loss (also known as L2-loss) is used to measure the output R of the position decoder. env With target R true Distance between:

[0130] L pos =‖R true -R env ‖ 2

[0131] The loss function L of the position decoder pos As shown above, ‖·‖ 2 Represents the mean squared error loss.

[0132] (3) Overall loss function:

[0133] Since the MLP model, encoder, distance decoder and position decoder together form a model that can correct GPS positioning, they need to be trained together during training. The overall loss function of the model is:

[0134] L overall =L dis ·λ+L pos

[0135] The overall loss function L overall As shown above, λ is a regularization parameter used to balance the influence of the distance decoder and the position decoder on the encoder. In this method, λ defaults to 1e-3.

[0136] (4) Model training method:

[0137] When training the model, a large amount of marked GPS positioning point data, as well as the corresponding road network, building measurement data and satellite data are needed to train the model.

[0138] For each GPS positioning instance and its corresponding real location, the environment matrix M is constructed using the road network and building height information. e , the timestamp matrix M constructed using the positioning time t , using GPS satellite status (altitude angle, azimuth angle and signal-to-noise ratio) to construct the corresponding sky map matrix M s .

[0139] For the training target, construct the corresponding one-hot encoding vector V for the distance decoder according to the positioning error in the GPS positioning instance true ; Construct a matrix R for the position decoder, marking the elements of the true position as Gaussian peaks true .

[0140] In order to determine the setting of λ, the distance decoder and the position decoder are trained separately to observe the order of magnitude of their loss functions, and then the value of λ is set to balance the loss functions of the two decoders. This method finally sets λ = 1e-3, at which both decoders can achieve the best prediction performance.

[0141] (5) In addition, this method uses Adam as the optimizer, with a learning rate of 1e-5 and a batch size of 128.

[0142] After setting the above parameters, this method uses the labeled GPS samples to train the model as a whole. Because the loss of the distance decoder and the loss of the location decoder are added together to update the common encoder, the two decoders can pass each other's feature information, so more accurate predictions can be achieved.

[0143] In addition, this method uses the MLP model as a part of the encoder, so the MLP model is trained together with the encoder.

[0144] Example 2

[0145] This embodiment specifically describes the implementation method of the system of the present invention.

[0146] like Figure 5 As shown, the system comprises:

[0147] The terminal obtains the original GPS positioning point, positioning time and satellite status data of the positioning object, presets the required error threshold, and sends it to the server;

[0148] Server, including: database and deep learning model.

[0149] The database stores environmental data and historical calculation data; the environmental data includes building height and road distribution data;

[0150] The deep learning model performs positioning correction on the original GPS positioning point, including a multi-source data fusion module, an encoder based on a residual network block (ResNet Block), a distance decoder (Distance Decoder) for predicting positioning errors, and a position decoder (Position Decoder) for predicting positioning positions;

[0151] The multi-source data fusion module provides input parameters for the deep learning model. The multi-source data fusion module searches for environmental data corresponding to the original GPS positioning point and positioning time from the database according to the original GPS positioning point, positioning time and satellite status data sent by the terminal, and converts the environmental data, positioning time and satellite status data into a matrix form;

[0152] The encoder extracts features from the matrix data output by the multi-source data fusion module and outputs a feature map encoding the environment, time and satellite status information;

[0153] The distance encoder predicts the positioning error of the original GPS positioning point based on the feature map and the error threshold. If the positioning error prediction value is greater than the error threshold, the position decoder is called, otherwise the original GPS positioning point is not corrected;

[0154] The position decoder outputs the corrected positioning position of the original GPS positioning point based on the feature map and sends it to the terminal, and sends the original GPS positioning point and the corrected positioning position to the database for storage.

[0155] The system is implemented as follows:

[0156] (1) Figure 5 As shown, when this system is needed for position correction, the client first obtains the GPS estimated position through the Android interface and other supporting data, and then convert these data (including Timestamp t, GPS satellite status) and error threshold δ are passed to the server. (The threshold δ is adjusted according to user needs)

[0157] (2) After receiving the request, the server retrieves the location from the database The nearby road network and building data are used, and the environment matrix M is constructed using the multi-source data fusion module. e 、Sky map matrix M s and the timestamp matrix M tOnce these input matrices are ready, the system will feed them into the encoder to get the feature map, and then feed the feature map into the range decoder to predict the GPS estimated position If the positioning error is less than the threshold δ, the system will is considered as the correct position p′. If the positioning error is greater than the threshold δ, that is, because It is not accurate enough for the user, so the position decoder needs to be called to predict the correct position, that is, the feature map is input into the position decoder to predict the positioning position. If the threshold δ is not provided, the distance decoder is skipped and the position decoder is directly executed to infer p′.

[0158] (3) Once the position correction request is processed, the system will record The record is stored in the database, indicating that the requester visited location p' at time t. These records help the system dynamically update the user's mobility model.

[0159] (4) Finally, the server sends the position p′ back to the client.

Claims

1. A method for GPS enhanced positioning in urban canyons based on deep learning, characterized in that: include: Obtain the influencing factor data of GPS accuracy in the urban canyon, including environmental data, satellite status data and positioning time data, and convert them into a matrix of the same size; the environmental data includes the building height and road distribution data around the initial GPS position; wherein the influencing factor data is converted into a matrix form as follows: a. For environmental data, a square area is constructed with the original estimated GPS position as the center, and the area is divided into several square grid cells; the square grid cells are assigned values ​​according to the environment they belong to, and the environmental matrix is ​​constructed; b. For satellite status data, draw a sky map based on the altitude angle, azimuth angle and signal-to-noise ratio of the GPS satellite; construct a sky map matrix of the same size as the environment matrix, map the satellites in the sky map to the sky map matrix, and assign values ​​to the matrix grid elements where there are satellites using the signal-to-noise ratio of the corresponding satellite; c. For the positioning time data, the timestamp at the time of positioning is converted into an n-dimensional vector, and the n-dimensional vector is converted into a matrix of the same size as the environment matrix and the sky map matrix, i.e., the timestamp matrix, using a multilayer perceptron; Inputting the converted multiple matrices of the same size into a trained encoder to obtain a feature map encoding environment, time and satellite status information; the encoder is an encoder based on a residual network block; Input the feature map into a trained distance decoder to obtain a predicted value of the positioning error of the original GPS positioning point; Compare the positioning error prediction value with the preset error threshold value. If the positioning error prediction value is greater than the acceptable error value, input the feature map into the trained position decoder to obtain the corrected positioning position of the original GPS positioning point. The encoder, distance decoder and position encoder are trained in the following manner: marked GPS positioning point data, corresponding road network, building measurement data and satellite status data are used as training samples to perform model training on the encoder, distance encoder and position encoder; The training target of the distance decoder is a one-hot encoding vector, and the element value corresponding to the correct error interval is 1; the cross entropy loss function is used to measure the distance between the output of the distance decoder and the training target; The training target of the position decoder is the Gaussian representation matrix of the true position, and the element corresponding to the true position in the Gaussian representation matrix is ​​the maximum value 1; the mean square error loss function is used to measure the distance between the output of the position decoder and the training target.

2. The method according to claim 1, characterized in that The way to map satellites in the sky map to the sky map matrix is: The central element of the sky map matrix corresponds to the center of the sky map. The distance between the satellite element and the central element on the sky map matrix is ​​determined according to the satellite's altitude angle. The clockwise angle between the line connecting the satellite element and the central element and the "directly above" direction is determined according to the satellite's azimuth angle, so as to map the satellite in the sky map to the sky map matrix.

3. The method according to claim 1, characterized in that Convert the timestamp at the time of positioning into a 7-dimensional vector, the 7 elements are: Element 1: an element that characterizes the progression of location time in a natural day; Element 2: An element that characterizes the progress of the satellite in a revisit cycle; Elements 3 to 6: Apply sine and cosine calculations to elements 1 and 2, respectively, to assist in finding periodic elements; Element 7: an element representing that the positioning time is before or after 12:00am on the same day in the time zone of the positioning location; The 7-dimensional vector is obtained by combining the 7 elements in any order.

4. The method according to claim 1, characterized in that: The output of the position decoder is a probability matrix, and the final output of the position decoder is obtained after adopting the environment matrix constraint; Set the elements in the environment matrix that have buildings or other inaccessible elements to 0, otherwise to 1, that is, valid candidate positions. The assigned environment matrix is ​​used as a constraint masker. The final output of the position decoder is the constraint masker dot product probability matrix. The element with the largest value in the final output matrix is ​​the element corresponding to the corrected GPS positioning position, and the longitude and latitude corresponding to the center of the element is the output corrected position.

5. The method according to claim 1, characterized in that During training, the encoder, distance encoder and position encoder are connected according to their input-output relationship and trained together.

6. The method according to claim 5, characterized in that When the encoder, distance encoder, and position encoder are trained together, the overall loss function of the training model is set to L overall =L dis ·λ+L pos Where L dis is the loss function of the distance encoder, L pos is the loss function of the position encoder, and λ is the regularization parameter used to balance the impact of the distance decoder and the position decoder on the encoder.

7. The method according to claim 1, characterized in that Also includes: When it is necessary to continuously locate a mobile object, estimate the moving speed of the mobile object and limit the reachable area of ​​the mobile object to assist in determining the correct GPS positioning position, including: Calculate the average moving speed of the mobile positioning object based on the latest corrected GPS positioning positions and corresponding positioning times; Construct a square area with the original estimated GPS position as the center, and divide the area into several square grid units; The corrected GPS positioning position of the mobile positioning object corresponding to the previous positioning time is used as the center of the circle, and the average moving speed × the time difference between adjacent positioning times is used as the radius to draw a circle to construct a circular area; According to the overlap between the square area and the circular area, the grid cells are assigned values. If the grid cells overlap with the circular area, a value of 1 is assigned. Otherwise, a value of ρ is assigned, where 0≤ρ≤1. The matrix formed after the assignment is called a moving masker. The output of the motion masker is multiplied by the output of the position decoder to obtain the final probability matrix, and the longitude and latitude corresponding to the center of the grid cell with the largest value in the probability matrix is ​​output as the corrected position for continuous positioning.

8. An urban canyon GPS enhanced positioning system based on deep learning, characterized in that: include: The terminal obtains the original GPS positioning point, positioning time and satellite status data of the positioning object, presets the required error threshold, and sends it to the server; Server, including: database and deep learning model; The database stores environmental data and historical calculation data; the environmental data includes building height and road distribution data; The deep learning model performs positioning correction on the original GPS positioning point, including a multi-source data fusion module, an encoder, a distance encoder and a position encoder; The multi-source data fusion module provides input parameters for the deep learning model. The multi-source data fusion module searches for environmental data corresponding to the original GPS positioning point and positioning time from the database according to the original GPS positioning point, positioning time and satellite status data sent by the terminal, and converts the environmental data, positioning time and satellite status data into matrices of the same size; the method of converting the influencing factor data into a matrix form is: a. For environmental data, a square area is constructed with the original estimated GPS position as the center, and the area is divided into several square grid cells; the square grid cells are assigned values ​​according to the environment they belong to, and the environmental matrix is ​​constructed; b. For satellite status data, draw a sky map based on the altitude angle, azimuth angle and signal-to-noise ratio of the GPS satellite; construct a sky map matrix of the same size as the environment matrix, map the satellites in the sky map to the sky map matrix, and assign values ​​to the matrix grid elements where there are satellites using the signal-to-noise ratio of the corresponding satellite; c. For the positioning time data, the timestamp at the time of positioning is converted into an n-dimensional vector, and the n-dimensional vector is converted into a matrix of the same size as the environment matrix and the sky map matrix, i.e., the timestamp matrix, using a multilayer perceptron; The encoder extracts features from the matrix data output by the multi-source data fusion module and outputs a feature map encoding the environment, time and satellite status information; The distance encoder predicts the positioning error of the original GPS positioning point based on the feature map and the error threshold. If the positioning error prediction value is greater than the error threshold, the position decoder is called, otherwise the original GPS positioning point is not corrected; The position decoder outputs the corrected positioning position of the original GPS positioning point based on the feature map and sends it to the terminal, and sends the original GPS positioning point and the corrected positioning position to the database for storage; The encoder, distance decoder and position encoder are trained in the following manner: marked GPS positioning point data, corresponding road network, building measurement data and satellite status data are used as training samples to perform model training on the encoder, distance encoder and position encoder; The training target of the distance decoder is a one-hot encoding vector, and the element value corresponding to the correct error interval is 1; the cross entropy loss function is used to measure the distance between the output of the distance decoder and the training target; The training target of the position decoder is the Gaussian representation matrix of the true position, and the element corresponding to the true position in the Gaussian representation matrix is ​​the maximum value 1; the mean square error loss function is used to measure the distance between the output of the position decoder and the training target.