A UWB Indoor Pedestrian Localization Method Based on Graph Convolutional Neural Network

Abstract

The present invention discloses a UWB indoor pedestrian positioning method based on a graph convolutional neural network. During the positioning process, a pedestrian holds a UWB tag, which emits pulse signals to UWB base stations. After the base stations receive the positioning packets, they are sent to the upper computer through the network for positioning data storage. At the same time, the positioning data is used to construct and train a graph convolutional neural network model for indoor positioning. The input of the graph convolutional neural network model is the base station coordinates and the positioning data in the time series during the movement of the pedestrian, and the output is the coordinates of the tag. When the pedestrian is to be positioned, the real-time acquired positioning data and the base station coordinates are input into the trained graph convolutional neural network model, and the coordinates of the tag to be positioned are predicted. Since the position information of adjacent moments is introduced on the basis of the graph convolutional neural network, compared with the existing UWB positioning technology, the present invention reduces the influence of non-line-of-sight propagation and multipath effects on the positioning accuracy and improves the indoor positioning accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of information engineering, relates to the technical field of high-precision indoor positioning, and particularly relates to a UWB indoor pedestrian positioning method based on a graph convolutional neural network. Background Art

[0002] Spatial location is an essential and important element for human beings in social activities. With the continuous development of human society, the continuous expansion of the activity range, and the increasing complexity of urban space, the spatial information of human activities has attracted more and more attention. Accurate indoor location information is required in fields such as special population guardianship, large venue management, Internet of Things, and personal location services. Therefore, high-precision indoor positioning technology is of great significance. In the outdoor environment, positioning and navigation technologies are becoming increasingly mature and have been widely applied in people's lives. However, in the indoor environment, the intensity and quality of satellite signals drop rapidly. Limited by conditions such as positioning time, positioning accuracy, and complex indoor environments, there are many difficulties in indoor positioning. Currently, existing indoor positioning methods mainly include WiFi, Bluetooth, infrared, ultra-wideband, RFID, ZigBee, etc. Among them, ultra-wideband indoor positioning technology has advantages such as fast data transmission rate, low transmission power, low device power consumption, and high security. However, due to the complex indoor environment and the susceptibility of wireless signals to multipath effects, it is still difficult to achieve precise positioning using the current UWB indoor positioning methods. In addition, traditional UWB-based TDOA positioning methods generally use direct calculation methods, such as the classical Chan algorithm and Taylor algorithm. The direct solution method not only generates an invalid solution that affects the accuracy, but also the Taylor expansion method requires an initial estimated position. When the initial estimated position is inaccurate, its calculation convergence is slow, and it may even diverge so that a good approximate solution cannot be obtained.

[0003] As an efficient machine learning method, neural networks themselves have good robustness, self-adaptive self-organization, high fault tolerance and other characteristics, and have been widely applied in real life, such as speech recognition, text recognition, etc. Although some commonly used neural network models, such as convolutional neural networks, have been widely applied in the above scenarios, it is difficult for them to be applied to graph data structures containing different attribute features. The proposed graph convolutional neural network can well solve the above problems. GCN can be directly applied to graph data structures, sharing local parameters of graph data through the graph structure, not only expanding the range of feature extraction, but also making the feature extraction process no longer limited to data with Euclidean structure. Summary of the Invention

[0004] In view of the above problems, in order to overcome the defects of the existing technology and improve the accuracy of indoor positioning, the purpose of the present invention is to provide a UWB indoor pedestrian positioning method based on a graph convolutional neural network, which applies the graph neural network to the calculation of position information for indoor positioning, and to a certain extent reduces the indoor positioning error caused by the non-line-of-sight propagation and multipath propagation of positioning signals. This method uses UWB base stations, UWB tags, a background server and a data processing terminal to construct a positioning system, obtains preliminary TDOA positioning information, uses the TDOA positioning information and the three-dimensional coordinate information of the UWB base stations as the feature representation of the positioning points, constructs a graph data structure between the positioning points, and uses the constructed graph data structure to build a graph convolutional neural network to extract the feature correlation between adjacent positioning nodes. Finally, the accurate position of the positioning tag is calculated using this graph convolutional neural network. Since the present invention combines the advantages of the UWB positioning method and the graph neural network, it does not require additional laying of a large number of hardware devices, nor does it require modification of the existing hardware devices, and has a more accurate indoor space positioning ability, providing theoretical and technical support for indoor scene positioning and indoor intelligent management, and having great advantages and commercial prospects in the application scenarios of high-precision indoor positioning.

[0005] The specific technical solution for achieving the purpose of the present invention is as follows:

[0006] A UWB indoor pedestrian positioning method based on a graph convolutional neural network, which uses tags and at least four UWB base stations to upload TDOA data packets to the host computer, and uses the graph convolutional neural network built and trained in the host computer to complete the indoor position positioning of pedestrians; the method includes the following steps:

[0007] Step 1: Sample collection

[0008] Arrange several UWB base stations indoors, and use a millimeter-level laser rangefinder to measure the coordinate information of the base stations; during the positioning process, pedestrians hold UWB tags that can emit signals. After the UWB base stations receive the pulse signals sent by the tags, they transmit the original TDOA positioning data to the local server as the prior data information for training the graph convolutional neural network.

[0009] Step 2: Train the model

[0010] Preprocess the base station coordinates measured by the laser rangefinder and the original TDOA positioning data received by the local server to obtain a data format that conforms to the input of the graph convolutional neural network, construct and train a graph convolutional neural network for indoor positioning, adjust the network parameters, and train to obtain an optimal network model;

[0011] Step 3: Apply the model

[0012] In the prediction stage of the pedestrian position coordinates, the TDOA data between the to-be-located tag and the base stations during the pedestrian movement and the base station position coordinates are input into the trained graph convolutional neural network model;

[0013] Step 4: Coordinate calculation

[0014] Based on the input of the obtained positioning data, the graph convolutional neural network model outputs the accurate coordinates of the UWB tag held by the pedestrian.

[0015] In Step 1, at least four UWB base stations are set, including one main base station and several slave base stations.

[0016] In Step 2, the original TDOA positioning data is the Time Difference of Arrival, which is important information required to obtain the indoor pedestrian coordinates.

[0017] In Step 2, the preprocessing of the base station coordinates measured by the laser rangefinder and the original TDOA positioning data received by the local server specifically includes:

[0018] (1) Convert the base station coordinates in the UWB positioning system coordinate system to the Cartesian coordinate system established for the pedestrian positioning area;

[0019] (2) The 32Hz UWB tag continuously transmits signals for 10s at a fixed coordinate point to obtain 320 TDOA data at this position point. Filter out the data where all base stations participate in positioning simultaneously and perform an averaging process to obtain the final TDOA feature data at this point;

[0020] (3) Convert the positioning data into the feature matrix format allowed by the graph neural network.

[0021] In Step 2, the construction and training of the graph convolutional neural network for indoor positioning, its construction specifically includes:

[0022] (1) The graph data form (Graph) is a type of data structure composed of nodes and edges; the graph G is represented by the node set V (Vertex) and the edges E (Edge) connecting the nodes:

[0023] G=(V,E)

[0024] (2) Since the movement trajectory of pedestrians indoors naturally has a temporal and spatial connection between the front and rear positions, the indoor pedestrian movement trajectory is constructed into the graph data form described in the graph neural network, where each positioning point is represented as a node on the graph, and the data feature of each positioning point is the feature of the node in the graph; the positioning points generated at two adjacent moments are connected by edges;

[0025] (3) The features of each node include the three-dimensional coordinate information of n base stations and (n - 1) TDOA data information generated by n base stations, totaling 3*n + (n - 1) = 4*n - 1 features; therefore, the input data of the graph convolutional neural network is a feature matrix of m*(4*n - 1) dimensions, that is, m consecutive position nodes in the pedestrian movement trajectory are selected, and each position node has (4*n - 1) features;

[0026] (4) Considering that the position coordinates of a pedestrian at a certain moment during the movement are closest to those at the previous and next moments, that is, the relationship is the closest, the construction method of the adjacency matrix is as follows: the positioning point at a certain moment in the pedestrian movement trajectory is only connected by an edge to the positioning points at the previous and next moments;

[0027] (5) The graph convolutional neural network for indoor positioning includes two graph convolutional layers and a fully connected layer connected in sequence; the graph convolutional neural network for indoor positioning is constructed.

[0028] For the graph convolutional neural network for indoor positioning, the updated state of the node is expressed as:

[0029] h v = f(x v , x co[v] , h ne[v] , x ne[v] )

[0030] where x and h represent the input feature and the hidden state respectively, co[v] and ne[v] represent the set of edges and the set of nodes connected to node v respectively; x v , x co[v] , h ne[v] , x ne[v] represent the node feature, the feature of the edge of this node, the hidden state of the adjacent nodes of this node, and the feature of the adjacent nodes of this node respectively; f is the local transfer function;

[0031] The node output is expressed as:

[0032] o v = g(h v , x v )

[0033] where g is the local output function;

[0034] Therefore, all the states, outputs, features, and node features in the graph can be represented in matrix form as H, O, X, X N :

[0035] H = F(H, X)

[0036] O = G′(H, X N)

[0037] Among them, F is the global transfer function, and G' is the global output function;

[0038] According to the Banach fixed-point theory, the graph neural network uses the following iterative method to solve the node state:

[0039] H t+1 = F(H t , X)

[0040] where H t represents the t-th iteration of H;

[0041] For the graph convolutional neural network for indoor positioning, after the hierarchical propagation of its graph convolutional layer, the output after graph convolution is expressed as:

[0042]

[0043] where X is the initial attribute matrix of all nodes; A is the adjacency matrix, I N is the n-order identity matrix, is the adjacency matrix with self-loops added; W is the learnable parameter in the training of the graph neural network; Z is the output after graph convolution; D is the degree matrix; is the matrix I N + D -1 / 2 AD -1 / 2 is the normalization result;

[0044] The iterative process of each layer of graph convolution is formally defined as:

[0045]

[0046] where, W t represents the parameter of the corresponding layer; σ represents the activation function.

[0047] In step 2, for the construction and training of the graph convolutional neural network for indoor positioning, its training specifically includes:

[0048] (1) First, input the node attribute features X and the adjacency matrix A of the entire graph, and through two graph convolutional layers, obtain the node embedding matrix Z':

[0049]

[0050] where the activation function used when training the graph convolutional neural network is the ELU activation function, and its specific form is:

[0051]

[0052] α represents a positive constant that determines the slope of the corresponding exponential function when the input x is negative;

[0053] (2) Use the fully connected layer to output the final output of the graph convolutional neural network:

[0054]

[0055] where ω represents the weight parameter of the connection between layers in the neural network, and b represents a bias term;

[0056] (3) Compare the predicted result train with the true label Y on the nodes V of the training set, and calculate the deviation between them using the MSELoss loss function; and the true label Y, and calculate the deviation between them using the MSELoss loss function;

[0057] (4) Calculate the gradient of the weight according to the loss function, and perform training by the stochastic gradient descent method to update the weight parameters of the network;

[0058] (5) Stop training until the loss is lower than the expected threshold, and save the model parameters at this time, that is, obtain the optimal graph convolutional neural network model.

[0059] In step 3, in the prediction stage of the pedestrian position coordinates, it specifically includes:

[0060] (1) Preprocess the TDOA value between the to-be-located tag and the base station obtained in real time and the base station position coordinates;

[0061] (2) Input the preprocessed data into the graph convolutional neural network model.

[0062] The beneficial effects of the present invention include:

[0063] 1) The present invention establishes a model based on the graph convolutional neural network for UWB indoor pedestrian positioning in a complex indoor environment, which can overcome the phenomenon of poor positioning accuracy caused by partial signal attenuation and signal instability;

[0064] 2) The present invention utilizes the feature extraction ability of the graph convolutional neural network and the mapping ability of the neural network to establish a more accurate mapping relationship between the positioning data features and the real data in the indoor environment where the signal is affected by non-line-of-sight propagation and multipath propagation, improving the positioning accuracy;

[0065] 3) The present invention combines the graph convolutional neural network with the UWB indoor positioning method, effectively reducing the influence of non-line-of-sight and multipath effects on the positioning accuracy, without the need to modify the existing hardware devices, and has great advantages and commercial prospects in the application scenarios of high-precision indoor positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is the flowchart of the UWB indoor pedestrian positioning method based on the graph convolutional neural network in the embodiment of the present invention;

[0067] Figure 2 It is the schematic diagram of the acquisition process of the positioning signal in the embodiment of the present invention;

[0068] Figure 3 It is the schematic diagram of the construction method of the graph data in the embodiment of the present invention;

[0069] Figure 4 It is the model structure diagram of the graph convolutional neural network in the embodiment of the present invention. Detailed implementation manners

[0070] Combined with the following specific embodiments and the accompanying drawings, the present invention will be further described in detail. The processes, conditions, experimental methods, etc. for implementing the present invention are all common knowledge and well-known common sense in the art except for the specifically mentioned content below, and the present invention has no special restrictive content.

[0071] The method of the present invention requires at least 4 UWB base stations, one of which is the main base station. In the construction and training stage of the graph convolutional neural network, first, a pedestrian holds a UWB tag, and the tag emits a pulse signal to the UWB base station. When the UWB background server receives the collected positioning data, it performs data cleaning at the data processing end and jointly constructs a graph data structure with the UWB base station coordinate values and sends it into the graph convolutional neural network to extract the feature associations between adjacent positioning nodes. According to the loss between the network output result and the tag value, the weight gradient is calculated, and the weight parameters are updated to obtain the final network model. During the indoor positioning process, the pedestrian holds a UWB tag, and the collected TDOA data and the coordinate values of the UWB base station are jointly sent into the trained graph convolutional neural network as node features, and the final positioning result coordinate values are output.

[0072] Embodiment

[0073] The present invention establishes UWB indoor pedestrian positioning based on a graph convolutional neural network model for complex indoor environments.

[0074] As Figure 1 shown, this embodiment provides a UWB indoor pedestrian positioning method based on a graph convolutional neural network, and the method includes:

[0075] Step 1: Sample acquisition

[0076] Arrange several UWB base stations indoors. During the positioning process, a pedestrian holds a UWB tag that can emit signals. After the UWB base station receives the pulse signal sent by the tag, it transmits the original data to the local server as the prior data information for training the graph convolutional neural network;

[0077] In step 1, four base stations are set, including one main base station and three slave base stations.

[0078] TDOA data is the time difference of arrival. The position information is calculated by the time difference of arrival of the ultra-wideband signals transmitted by the UWB positioning tag to two base stations.

[0079] Step 2: Train the model

[0080] Preprocess the collected base station coordinate information and TDOA data to obtain a data format that conforms to the input of the graph convolutional neural network. Construct and train a graph convolutional neural network for UWB indoor positioning, adjust the network parameters, and train to obtain an optimal model;

[0081] In step 2, the process of collecting positioning data is as Figure 2 shown. The preprocessing of the base station coordinate information includes: converting the base station coordinates in the UWB positioning system coordinate system to the Cartesian coordinate system established in a specific positioning area; the 32Hz UWB tag continuously transmits signals for 10s at fixed coordinate points to obtain 320 TDOA data of this position point, and screen out the data where 4 base stations participate in positioning at the same time for averaging to obtain the final TDOA feature data of this point; send the positioning data into the network model in the format of a feature matrix that conforms to the input allowed by the graph neural network. The graph convolutional neural network model includes two graph convolutional layers and a fully connected layer connected in sequence.

[0082] Among them, the construction method of the graph data is as Figure 3 shown. In the graph data structure, the features of each node include the three-dimensional coordinate information of four base stations and three TDOA data information generated by the four base stations, totaling 15 features. That is, the feature of node n is represented as (X1, Y1, Z1, X2, Y2, Z2, X3, Y3, Z3, X4, Y4, Z4, TDOA1, TDOA2, TDOA3). Represent the time-series positioning data generated by pedestrians walking indoors as a graph data structure. The nodes in the graph are the coordinate points at each moment, and the coordinate points generated at two adjacent moments are connected by edges. Among them, the feature of each node is the above 15. The construction method of the adjacency matrix is to connect the position points at a moment with the position points at the previous and next moments by edges.

[0083] Select 14 coordinate points in the pedestrian trajectory for prediction at a time. Therefore, the input data of the graph convolutional neural network is a 14*15 feature matrix.

[0084] The node update state of the graph convolutional neural network is expressed as:

[0085] h v =f(x v ,x co[v] ,h ne[v],x ne[v] )

[0086] Among them, x and h represent the input feature and the hidden state respectively, and co[v] and ne[v] represent the set of edges and the set of nodes connected to node v respectively. x v , x co[v] , h ne[v] , x ne[v] represent the node feature, the feature of the edge of this node, the hidden state of the adjacent nodes of this node, and the feature of the adjacent nodes of this node respectively. f is the local transfer function.

[0087] The node output of the graph convolutional neural network is expressed as:

[0088] o v = g(h v , x v )

[0089] Among them, g is the local output function.

[0090] Therefore, all the states, outputs, features, and node features in the graph can be represented in matrix form as:

[0091] H = F(H, X)

[0092] O = G(H, X N )

[0093] According to the Banach fixed-point theory, the update process of the node can be expressed as:

[0094] H t+1 = F(H t , X)

[0095] Therefore, the hierarchical propagation process of the described graph convolutional neural network is expressed as:

[0096]

[0097] Among them, X is the initial attribute matrix of all nodes; A is the adjacency matrix, I N is the n-order identity matrix, is the adjacency matrix with self-loops added; W is the learnable parameter in the network model training. Z is the output after graph convolution. D is the degree matrix. is the normalization result of the matrix I N + D -1 / 2 AD -1 / 2 .

[0098] In the described graph convolutional neural network, each layer of graph convolution is defined as:

[0099]

[0100] Among them,

[0101] When training the graph convolutional neural network, the activation function used is the ELU activation function, and its specific form is:

[0102]

[0103] where α represents a positive constant, which determines the slope of the corresponding exponential function when the input x is negative.

[0104] Among them, the specific steps in the training stage include:

[0105] (1) First, input the node attribute features X and the adjacency matrix A of the entire graph, and through two graph convolutional layers, obtain the node embedding matrix Z':

[0106]

[0107] (2) Then use a fully connected layer to output the final output of the graph convolutional neural network:

[0108]

[0109] (3) On the nodes V train in the training set, compare the predicted results with the true label Y, and use the MSELoss loss function to calculate the deviation between them.

[0110] (4) Calculate the gradient of the weights according to the loss function, and perform training through the stochastic gradient descent method to update the weight parameters of the network.

[0111] (5) Stop training until the loss is lower than the expected threshold, and save the model parameters at this time.

[0112] Among them, the graph convolutional neural network model includes two graph convolutional layers and a fully connected layer connected in sequence. The model diagram of the graph convolutional neural network is as Figure 4 shown.

[0113] Step 3: Model application

[0114] In actual application, input the TDOA value between the to-be-located label and the base station and the base station location coordinates during the pedestrian movement process into the trained graph convolutional neural network;

[0115] In step 3, the specific steps in the prediction stage of the pedestrian position coordinates include:

[0116] (1) Preprocess the TDOA value between the to-be-located label and the base station and the base station location coordinates obtained in real time;

[0117] (2) Load the trained graph convolutional neural network model;

[0118] (3) Input the preprocessed data into the graph convolutional neural network model, and output the position coordinates of the to-be-localized tag.

[0119] Step 4: Coordinate calculation

[0120] Based on the acquired positioning data input, the graph convolutional neural network outputs the accurate coordinates of the UWB tag held by the pedestrian.

[0121] In Step 4, a comparison of some positioning results obtained by the UWB indoor pedestrian positioning method based on the graph convolutional neural network and the positioning accuracy of this positioning method and that of the prior art is shown in Table 1. The mean square error is used as an evaluation index to compare the performance of each algorithm. It can be seen from the error comparison of each algorithm in Table 1 that the average error of some coordinates obtained by the UWB indoor pedestrian positioning method based on the graph convolutional neural network proposed in the present invention is 0.2657 m, which is better than the traditional Chan-Taylor positioning method and the ordinary MLP fully connected neural network method. In summary, the indoor pedestrian positioning method proposed in this paper has higher accuracy than other algorithms.

[0122] Table 1

[0123]

[0124] The present invention establishes a model based on the graph convolutional neural network for UWB indoor pedestrian positioning in a complex indoor environment, which can overcome the phenomenon of poor positioning accuracy caused in part by signal attenuation and signal instability. By utilizing the feature extraction ability of the graph convolutional neural network and the mapping ability of the neural network, in an indoor environment where the signal is affected by non-line-of-sight propagation and multipath propagation, a more accurate mapping relationship between the positioning data features and the real data is established, improving the positioning accuracy. The combination of the graph convolutional neural network and the UWB indoor positioning method effectively reduces the influence of non-line-of-sight and multipath effects on the positioning accuracy, without the need to modify the existing hardware devices, and has great advantages and commercial prospects in the application scenarios of high-precision indoor positioning.

[0125] The above embodiments are only examples and do not represent the limitation of the scope of the present invention. These embodiments can also be implemented in various other ways and can be subject to various omissions, substitutions, and changes without departing from the technical idea of the present invention.

Claims

1. A UWB indoor pedestrian positioning method based on graph convolutional neural network, characterized in that, This method uses UWB tags and at least four UWB base stations to upload TDOA data packets to the host computer, and uses the constructed and trained graph convolutional neural network model in the host computer to complete indoor pedestrian position localization; the method includes the following steps: Step 1: Sample collection Arrange several UWB base stations indoors, and use a millimeter-level laser rangefinder to measure the coordinate information of the base stations; during the positioning process, the pedestrian holds a UWB tag that emits signals. After the UWB base stations receive the pulse signals emitted by the tag, they transmit the original TDOA positioning data to the local server as the prior data information for training the graph convolutional neural network. Step 2: Train the model Preprocess the base station coordinates measured by the laser rangefinder and the original TDOA positioning data received by the local server to obtain a data format that conforms to the input of the graph convolutional neural network, construct and train a graph convolutional neural network for indoor positioning, adjust the network parameters, and train to obtain the optimal network model. Step 3: Model application In the prediction stage of the pedestrian position coordinates, input the TDOA data between the tag to be located and the base stations during the pedestrian's movement and the base station position coordinates into the trained graph convolutional neural network model. Step 4: Coordinate calculation The graph convolutional neural network model outputs the accurate coordinates of the UWB tag held by the pedestrian according to the obtained positioning data input.

2. The UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, wherein In step 1, at least four UWB base stations are set, including one main base station and several slave base stations.

3. The UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, characterized in that, In step 2, the original TDOA positioning data is the time difference of arrival, which is important information required to obtain the indoor pedestrian coordinates.

4. A UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, characterized in that In step 2, the preprocessing of the base station coordinates measured by the laser rangefinder and the original TDOA positioning data received by the local server specifically includes: (1) Convert the base station coordinates in the UWB positioning system coordinate system to the Cartesian coordinate system established for the pedestrian positioning area. (2) The UWB tag at 32 Hz continuously emits signals for 10 s at a fixed coordinate point, obtains 320 TDOA data at this position point, screens out the data where all base stations participate in positioning at the same time for averaging processing, and obtains the final TDOA feature data at this point. (3) Convert the positioning data into a feature matrix format that conforms to the input allowed by the graph neural network.

5. The UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, characterized in that In step 2, the construction of the graph convolutional neural network for indoor positioning specifically includes: (1) The graph data form is a type of data structure composed of nodes and edges; the graph G is represented by the node set V and the edges E connecting the nodes: G=(V,E) (2) Since the movement trajectory of pedestrians indoors naturally has the time and space connection of the front and rear positions, the indoor pedestrian movement trajectory is constructed into the graph data form described in the graph neural network, where each positioning point is represented as a node on the graph, and the data feature of each positioning point is the feature of the node in the graph; the positioning points generated at two adjacent moments are connected by edges. (3) The features of each node include the three-dimensional coordinate information of n base stations and (n - 1) TDOA data information generated by n base stations, totaling 3*n + (n - 1) = 4*n - 1 features; therefore, the input data of the graph convolutional neural network is a feature matrix of m*(4*n - 1) dimensions, that is, m consecutive position nodes in the time series of the pedestrian movement trajectory are selected, and each position node has (4*n - 1) features; (4) Considering that the position coordinates of a pedestrian at a certain moment during movement are closest to, and thus most closely related to, the position coordinates at the previous and next moments, the construction method of the adjacency matrix is as follows: The positioning point at a certain moment in the pedestrian movement trajectory is only connected by an edge to the positioning points at the previous and next moments; (5) The graph convolutional neural network for indoor positioning described above includes two graph convolutional layers and a fully connected layer connected in sequence; thus, the graph convolutional neural network for indoor positioning is constructed.

6. The UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 5, wherein For the graph convolutional neural network for indoor positioning described above, the updated state of the node is expressed as: h v = f(x v , x co[v] , h ne[v] , x ne[v] ) Among them, x and h represent the input feature and the hidden state respectively, and co[v] and ne[v] represent the set of edges and the set of nodes connected to node v respectively; x v , x co[v] , h ne[v] , x ne[v] represent the node feature, the feature of the edge of this node, the hidden state of the adjacent nodes of this node, and the feature of the adjacent nodes of this node respectively; f is the local transfer function; The node output is expressed as: o v = g(h v , x v ) where g is the local output function; Therefore, all the states, outputs, features, and node features in the figure are respectively represented in matrix form as H, O, X, X N : H = F(H, X) O = G′(H, X N ) where F is the global transfer function and G' is the global output function; According to the Banach fixed-point theorem, the graph neural network uses the following iterative method to solve the node state: H t+1 = F(H t , X) where H t represents the t-th iteration of H.

7. A UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 5, characterized in that For the graph convolutional neural network for indoor positioning described above, after the hierarchical propagation of its graph convolutional layer, the output after graph convolution is expressed as: where X is the initial attribute matrix of all nodes; A is the adjacency matrix, and I N is the identity matrix of order n, is the adjacency matrix with self-loops added; W is a learnable parameter in the training of the graph neural network; Z is the output after graph convolution; D is the degree matrix; is the matrix I N +D -1 / 2 AD -1 / 2 is the normalization result of; The iterative process of each layer of graph convolution is formally defined as: Among them, W t represents the parameters of the corresponding layer; σ represents the activation function.

8. A UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, characterized in that, In step 2, for the construction and training of the graph convolutional neural network for indoor positioning, its training specifically includes: (1) First, input the node attribute features X and the adjacency matrix A of the entire graph, and through two graph convolutional layers, obtain the node embedding matrix Z': where the activation function used when training the graph convolutional neural network is the ELU activation function, and its specific form is: α represents a positive constant that determines the slope of the corresponding exponential function when the input x is negative; (2) Use the fully connected layer to output the final output of the graph convolutional neural network: where ω represents the weight parameter of the connection between layers in the neural network, and b represents a bias term; (3) Compare the prediction results train on the node V of the training set with the true label Y, and calculate the deviation between them using the MSELoss function; (4) Calculate the gradient of the weight according to the loss function, and perform training through the stochastic gradient descent method to update the weight parameters of the network; (5) Stop training until the loss is lower than the expected threshold, and save the model parameters at this time, that is, obtain the optimal graph convolutional neural network model.

9. A UWB indoor pedestrian positioning method based on a graph convolutional neural network according to claim 1, characterized in that, In step 3, for the prediction stage of the pedestrian position coordinates, it specifically includes: (1) Preprocess the TDOA values between the real-time acquired positioning tag and the base stations and the base station position coordinates; (2) Input the preprocessed data into the graph convolutional neural network model.

Images