Multi-label ultra wide band sensor positioning method based on state space model

Abstract

The invention provides a multi-label ultra-wideband sensor positioning method based on a state space model. The method comprises the following steps: step 1, embedding base station information; step 2, carrying out segmentation, mapping and fusion on the embedded data to obtain fusion features; 3, extracting spatio-temporal characteristics of the measurement value of the ultra-wideband sensor; 4, forming an indoor positioning model by the state space module and the ultra-wideband sensor label position and attitude prediction module; and step 5, training an indoor positioning model to realize synchronous prediction of label positions and postures so as to meet requirements of indoor positioning. Through the method, the measurement noise of the ultra wide band sensor can be efficiently suppressed, the advanced spatial-temporal characteristics of the measurement value of the ultra wide band sensor can be fully extracted, the real-time estimation of the position and attitude of the label can be realized, and higher reliability and robustness are provided for indoor positioning based on the ultra wide band sensor.

Description

Technical Field

[0001] The invention belongs to the technical field of positioning, and in particular relates to a multi-tag ultra-wideband sensor positioning method based on a state space model. Background Art

[0002] At present, ultra-wideband sensors are widely used for indoor positioning due to their high precision and strong real-time performance. In recent years, some ultra-wideband indoor positioning methods based on deep learning have achieved good positioning results in some scenarios, but there are still some shortcomings:

[0003] The current single-tag ultra-wideband positioning method only estimates the tag position, but ignores the robot's posture estimation, and posture information is crucial for moving carriers such as robots; in addition, the existing ultra-wideband indoor positioning process does not fully consider the geometric information that can be used between base stations and tags. These geometric information can provide a more complete and accurate spatial reference frame, so that the ranging value can be more effectively used in the tag positioning calculation. Failure to fully consider the geometric information between base stations and multiple tags is equivalent to losing important correction dimensions, making the positioning results inaccurate. By combining the geometric information between ultra-wideband base stations and between tags, the ranging data can be better corrected and interpreted, thereby achieving improved positioning accuracy and robustness while estimating the posture. In addition, when there are more obstacles or signal interference in complex scenes, the error of the ranging value will increase significantly, making it difficult for the original model to handle complex spatially distributed interference, which in turn leads to a decrease in the accuracy of the positioning results; Summary of the invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to provide a multi-tag ultra-wideband (UWB) sensor positioning method based on the state space model (State Space Model) to address the deficiencies of the prior art, thereby improving the deficiencies of the prior art method, not only providing posture estimation, but also improving the accuracy of tag position estimation. The method of the present invention comprises the following steps: Step 1: embedding base station information based on the pre-calibrated base station position in the indoor environment and the continuous ranging value between the ultra-wideband sensor tag and the base station; Step 2: segment, map and fuse the embedded data to obtain fusion features; Step 3, construct a state space module based on the state space model to extract the spatiotemporal characteristics of the ultra-wideband sensor measurement value; Step 4, establishing an ultra-wideband sensor tag position and posture prediction module to achieve synchronous prediction of tag position and posture; the state space module and the ultra-wideband sensor tag position and posture prediction module form an indoor positioning model; Step 5, design a network training method based on supervised learning, adopt a supervised learning strategy, based on the predicted value of the ultra-wideband sensor tag position and the predicted value of the Euler angle representing the posture obtained by the ultra-wideband sensor tag position and posture prediction module, the pre-collected true quaternion, and the true position value, construct a loss function based on position constraints, posture constraints and joint constraints, which is used to train the indoor positioning model and realize the synchronous prediction of the tag position and posture, thereby meeting the needs of indoor positioning.

[0005] Step 1 includes: Step 1-1, arrange N ultra-wideband sensor base stations indoors, and the coordinates of the N ultra-wideband sensor base stations in the world coordinate system have been calibrated in advance; at the same time, M ultra-wideband sensor tags are fixedly installed on any movable carrier; considering that the measurement frequency of ultra-wideband is usually between 10Hz and 100Hz, ten continuous measurement values ​​between a tag and a base station are collected as samples, which not only avoids excessive calculation, but also can extract enough information in a short time, so as to ensure that the position and attitude information can be updated within seconds. After collecting samples within two or more time steps, the original ranging information is obtained, and the original ranging information is separated along the base station dimension, and the ranging data between each base station and the tag is processed independently. The formula is: , in Indicates equal division along the dimension of the base station, Represents the original ranging information, represents the continuous ranging information between multiple tags and the i-th base station, The dimension is batch×M×10; batch represents the batch size; Step 1-2, convert the single base station position into a feature vector of the same size as the single base station ranging information through a shared fully connected layer, and expand the feature vector at the same time, the formula is: , in represents the coordinates of the i-th base station in the world coordinate system under the batch time step, express Expand the batch times in the first dimension, represents a shared linear transformation from 3D to 10D, represents the transformed and expanded single base station feature vector, where The dimension is 1×3, The dimension is batch×10; Step 1-3, the transformed and expanded single base station feature vector is further expanded along the label dimension, and the formula is: , , in Express Transform the dimension Adjust to shape, express Expand M times in the second dimension, represents the reshaped single base station feature vector, Represents the feature vector of the i-th base station after expansion, with a dimension of batch×M×10; Step 1-4, embed the ranging information between the tag and the single base station into the processed feature vector of the corresponding base station and splice it into new fused data.

[0006] In steps 1-4, the following formula is used to embed the ranging information between the tag and the single base station into the processed feature vector of the corresponding base station and splice it into new fused data: , in Represents a splicing operation, Represents the fused data after embedding, with dimension , where the parameters The value of is N×10.

[0007] Step 2 includes: Step 2-1, segment the continuous ranging information within a fixed time step from the embedded data, specify the appropriate batch size batch through experiments, and select a suitable number from the integer divisor of the batch as the fixed time step size T. The selection of T needs to take into account both computational efficiency and time series information representation; then, separate the ranging information of each ultra-wideband sensor tag, and project the ranging information after fusion of base station information onto the linear transformation. dimensional feature space, the formula is: , in represents the ranging information of the jth tag at the kth time step, Indicated by VI dimensional linear transformation, It means that the dimension after linear transformation is The feature vector of the j-th label at the k-th time step; Step 2-2, concatenate and fuse the multi-label feature vectors extracted at a time. The fusion method is to concatenate and pass them through a fully connected layer together to map the multi-label feature vectors to In the feature space of dimension, the formula is: , in Indicated by VI dimensional linear transformation, Represents the concatenated feature vector of multiple labels at the kth time step, and its dimension is ; The multi-label feature vectors at all times within a fixed time step are concatenated, and the formula is: , in Indicates The multi-label feature vectors of all moments within a fixed time step T are of dimension ; Representative The concatenated feature vector of multiple labels at time steps; Step 2-3, continue to extract the multi-label feature vectors of all moments within the fixed time step T from the data of batch time steps according to steps 2-1 and 2-2, and then perform splicing and fusion. is the parameter S, where Indicates that an integer is rounded down. The formula is: , in, Indicates The multi-label feature vectors of all moments within a fixed time step T are grouped together. It is the fusion feature obtained after the final processing, and the dimension is .

[0008] Step 3 includes: Step 3-1, extract the features of each ultra-wideband sensor tag from the fusion feature F separately: regard the fusion feature F as M dimensions The independent features of the single label are placed in the state space module based on the state space model to deeply extract the long-term spatiotemporal dependency, while effectively suppressing noise interference. The formula is: , in represents the state-space model, represents the time step size of the data passed into the state-space model, The value of is equal to S, represents the sequence length of the time series, The value of is equal to T, represents the feature dimension of the input, The value of equal, represents the feature dimension of the output, The value and input dimension same; The state space module includes linear layers, convolutional layers, activation layers and state space models. For independent features of a label ,in Here, the single ultra-wideband sensor tag feature extracted from the fusion feature F above is the same, and the state space module performs the following transformation: First, the data is normalized, the input features are standardized, and the input data is retained as the residual part, which can effectively avoid the gradient disappearance problem and provide multi-scale information. The formula is: , , in Represents the layer normalization operation, where It represents the result of normalizing the independent feature layer of the label, and the dimension is still ; Indicates independent features The residual of Secondly, the normalized data is retained and processed by the linear layer and the activation layer as the residual to enhance the feature diversity. At the same time, the normalized data is passed through the linear layer and the convolution layer in sequence to perform linear transformation, convolution and activation operations on the data. This operation can capture the local spatiotemporal pattern in the time series. The specific formula is: , , Where ReLU represents the activation function, and Is VI Conv represents a one-dimensional convolution operation. The result of linear transformation and activation operation on the normalized data of the representation layer is used as the residual. Represents the captured local spatiotemporal activation features; Then, the activated data is passed into the state space model and compared with the residual Multiplying the corresponding elements is helpful for modeling long-term dependent information. The formula is: , in represents the state space model, ⊙ represents the multiplication of corresponding elements, Indicates long-term dependence on fusion features, The dimension is ; Finally, the dimension is restored after linear transformation and combined with the previous independent features The residual Connect, fuse the original input and long-term dependent fusion features to prevent information attenuation. The formula is: , in Indicated by VI dimensional linear transformation, is the output of the state space module, with the same dimension as the input Consistency; Step 3-2: In step 3-1, the high-level feature representations of the M single labels obtained after passing through the state space module are , the dimension of each high-level feature representation is , concatenate the M single-label high-level feature representations and reshape them to obtain a dimension of The data is: , in It represents the result of concatenating M high-level features of a single label, and its dimension is ; S The value of T is batch.

[0009] Step 4 includes: Step 4-1, for the estimation of the position of the ultra-wideband sensor tag: the ultra-wideband sensor tag position and posture prediction module includes two fully connected layers, M high-level feature representations of single labels Share two fully connected layers, and sequentially change the dimensions from The dimension is reduced to 32 dimensions, and then from 32 dimensions to 3 dimensions, and the three-dimensional coordinates of M labels in the world coordinate system are obtained in turn. , three-dimensional coordinates The dimensions are all batch×3, and the three-dimensional coordinates For splicing, the formula is: , in Represents the position prediction value of M labels, the dimension is batch×M×3; Step 4-2, estimate the posture.

[0010] Step 4-2 includes: concatenating the M high-level features of a single label obtained in step 3-2 Through two fully connected layers, the dimensions are changed from Dimension reduction Wei, then from The dimension is reduced to 3 dimensions, and the predicted value of the Euler angle representing the posture is obtained .

[0011] Step 5 includes: Step 5-1: Establish a loss function for the prediction of the locations of more than two ultra-wideband sensor tags : , in represents the training batch size, The value of is the same as batch. represents the predicted value of the label position at the kth time step, Represents the true value of the label position at the kth time step; Step 5-2: For posture estimation, the indoor positioning model predicts the Euler angle ,Will Convert to rotation matrix and establish loss function : , in represents the conversion of the rotation matrix to Euler angles, represents the predicted value of the rotation matrix from the carrier coordinate system to the world coordinate system at the kth time step, Predicted values ​​from Euler angles Converted to Represents the true value of the rotation matrix corresponding to the k-th time step, which is transformed from the true value quaternion; Step 5-3, obtain a joint loss function through linear transformation : , , in represents the position of the jth tag in the carrier coordinate system, and its dimension is 1×3×1. represents the position of the jth label in the carrier coordinate system after expansion, and its dimension is batch×1×3. Indicates the position of the jth label at the kth time step; Step 5-4, establish the complete loss function L: , in , , Represent the loss function The weight of The weight and The weight of Step 5-5, using the loss function established in step 5-4, train the indoor positioning model through supervised training.

[0012] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0013] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction is run on a computer, the steps of the method described are executed.

[0014] Compared with the prior art, the present invention has the following beneficial effects: the state-space model can effectively filter noise by parameterizing its input, and its excellent long sequence dependency modeling capability and linear complexity can efficiently extract advanced spatiotemporal features. The design of the joint loss function makes full use of the geometric information provided by multiple labels, so that the model can more accurately realize multi-label position prediction while providing spatial constraints and reference predicted postures. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the method of the present invention.

[0016] Figure 2 It is a schematic diagram of the process of embedding the position of the ultra-wideband sensor base station proposed in the present invention.

[0017] Figure 3 It is a schematic diagram of the state space model mentioned in the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and other advantages of the present invention will become more clear.

[0019] Reference Figure 1 The embodiment of the present invention provides a multi-tag ultra-wideband sensor positioning method based on a state space model, comprising the following steps: Step 1: embedding the UWB sensor base station information to enhance the feature expression capability of the measurement data: embedding the base station information based on the pre-calibrated base station position in the indoor environment and the continuous ranging values ​​between the UWB sensor tag and the base station; specifically, collecting ten continuous measurement values ​​between the tag and the base station as samples in each time step, and converting the corresponding base station position and ranging information into feature vectors of the same size through two independent fully connected layers, and adding them to complete the base station information embedding process; Step 2: Segment, map and fuse the embedded data to enhance the global spatiotemporal relationship modeling capability of the data: segment the embedded data at fixed time steps, project them separately according to the labels, and then fully fuse the single-label projection data globally in space and time, thus laying the foundation for subsequent advanced spatiotemporal feature extraction; Step 3, construct a state space module based on the state space model to extract the spatiotemporal features of the ultra-wideband sensor measurement value: based on the fusion data obtained in step 2, input it into two state space modules in turn to obtain high-level spatiotemporal features; Step 4: Establish an ultra-wideband sensor tag position and posture estimation module to achieve synchronous prediction of tag position and posture: Based on the high-level spatiotemporal features obtained in step 3, first split them according to tags, share two fully connected layers, and predict the position of each tag; at the same time, fuse the high-level features to effectively capture the global relationship between multiple tags and base stations, and then predict the posture of the robot coordinate system through two fully connected layers, and the results are displayed in the form of Euler angles; Step 5, design a network training method based on supervised learning, adopt a supervised learning strategy, based on the predicted value of the ultra-wideband sensor tag position and the predicted value of the Euler angle representing the posture obtained by the indoor positioning model constructed in steps 1 to 4, the pre-collected true quaternion and the true position value, construct a loss function based on position constraints, posture constraints and joint constraints, which is used to train the ultra-wideband indoor positioning model constructed in steps 1 to 4, and realize the synchronous prediction of tag position and posture, thereby meeting the needs of indoor positioning.

[0020] Step 1 specifically includes: Step 1-1, N ultra-wideband sensor base stations are arranged indoors, and their coordinates in the world coordinate system have been calibrated in advance. At the same time, a mobile robot with an EAI wheeled chassis, NVIDIA AGX Xavier, RSLiDAR-16 and ultra-wideband tags is constructed, and M ultra-wideband sensor tags are fixedly installed on it; in the present invention, N is set to 6, M is set to 4, and the 4 tags on the robot have a certain geometric relationship. Ten continuous measurement values ​​between a tag and a base station are collected as samples. After collecting samples in multiple time steps, the original ranging information is obtained, and the original ranging information is separated by base station so as to independently process the ranging data between each base station and the tag. The formula is: , in Indicates equal division along the dimension of the base station, represents the original ranging information, and Represents the continuous ranging information between multiple tags and the i-th base station. Its dimension is batch×4×10. After testing, this experiment determines that the batch size is 1024; Step 1-2, through a shared fully connected layer, the single base station position is converted into a feature vector of the same size as the single base station ranging information, and it is expanded at the same time. The formula is: , in express Expand the batch times in the first dimension, represents a shared linear transformation from 3D to 10D, represents the coordinates of the ith base station in the world coordinate system at the batch time step (which does not change over time), represents the single base station feature vector after transformation and expansion, where The dimension is 1×3, The dimension is batch×10; Step 1-3, the transformed and expanded single base station information is further expanded along the label dimension, the formula is: , , in represents a reshape operation, means that the second dimension is expanded 4 times, and represents the feature vector of the i-th base station after processing, and its dimension is batch×4×10; Step 1-4, embed the distance measurement information between the tag and the single base station into the feature vector of the corresponding base station after processing, and splice it into new fusion data. The formula is: , in Represents a splicing operation, Represents the embedded fusion data, and its dimension can be expressed as ,in The value is 6×10.

[0021] Step 2 specifically includes: like Figure 2 As shown, this figure is a schematic diagram of the process of embedding the position of the ultra-wideband sensor base station proposed by the present invention.

[0022] Step 2-1: Extract the continuous ranging information within a fixed time step from the fused data. In this experiment, the fixed time step size T is set to 16. Then, separate the ranging information of each tag and map it to dimensional feature space, the formula is: , in represents the distance measurement information of the kth label at the kth time step, Indicated by VI dimensional linear transformation, and It means that the dimension after linear transformation is The feature vector of the th label at the kth time step, where The value of is 256, that is, The dimension is batch×4×256.

[0023] Step 2-2, concatenate and fuse the multi-label feature vectors extracted at a certain moment. The fusion method is to concatenate and pass them through a fully connected layer together and map them to In the feature space of dimension, the formula is: , in Represented by VI dimensional linear transformation, is the concatenated feature vector of multiple labels at the kth time step, and its dimension is ,in The value is 512; After that, the multi-label feature vectors at all times within a fixed time step are concatenated, and the formula is: , in For the The multi-label feature vectors of all moments within a fixed time step T are of dimension ; Step 2-3, continue to extract the multi-label feature vectors of all moments within the fixed time step T from the data of batch time steps according to steps 2-1 and 2-2, and concatenate them. is S, where Indicates that the integer is rounded down. It is easy to find that in this experiment, the S value is 64. The formula is: , in It is the projection fusion feature obtained after the final processing, and its dimension is .

[0024] Step 3 specifically includes: like Figure 3 As shown, this figure is a schematic diagram of the state space model mentioned in the present invention.

[0025] Step 3-1, for the obtained projection fusion features , extract the features of each label separately, that is, regard it as 4 dimensions The independent features are put into the state space model to deeply extract the long-term spatiotemporal dependencies, while effectively suppressing noise interference. The formula is: , in represents the state space model, Indicates the time step size of the input network. Here, its value is the same as S, which is 64. Indicates the length of the time series. Here, its value is the same as T, which is 16. Represents the feature dimension of the input, where its value is The same is 512, Represents the feature dimension of the output, which is actually the same as the input dimension same; Specifically, in Set the projection dimensions in The value is 1024, the size of the convolution kernel is set to 3, the step size of the convolution process is set to 1, and the padding of the convolution process is set to 1; Step 3-2: Extract and transform the features of each label layer by layer after passing through the state space model in step 3-1 to generate 4 high-level feature representations of single labels, which are , the dimension of each feature is , and then concatenate them and perform a simple Reshape operation to obtain a dimension of data, that is , which is convenient for subsequent prediction of position and posture. The formula is as follows: , in It represents the result of concatenating the high-level spatiotemporal features of the four tags, and its dimension can be expressed as .

[0026] Step 4 specifically includes: Step 4-1, for the estimation of label position: using the four single-label high-dimensional features obtained in step 3-2, sharing two fully connected layers, the dimensions are reduced from 512 dimensions to 32 dimensions, and then from 32 dimensions to 3 dimensions, and the three-dimensional coordinates of the four labels in the world coordinate system are obtained in turn. , whose dimensions are all batch×3, are concatenated, and the formula is as follows: , in It represents the position estimate of 4 labels, and its dimension is batch×4×3; Step 4-2, for posture estimation: Use the deep feature splicing results obtained in step 3-2 After a simple Reshape fusion, it passes through two fully connected layers in succession, reducing the dimension to 512 and 3 dimensions respectively, and obtaining the Euler angle prediction value representing the posture .

[0027] Step 5 specifically includes: Step 5-1: Establish a loss function for estimating multiple label positions , its mathematical representation is: , in is the training batch size, which is the same as batch, 1024. is the predicted value of the label at the kth time step, is the true value of the label position at the kth time step; Step 5-2: For posture estimation, the predicted value is the Euler angle , convert it into a rotation matrix and establish the loss function , its mathematical representation is: , in It represents the conversion of rotation matrix to Euler angle. is the predicted value of the posture matrix transformed from the robot coordinate system to the world coordinate system at a certain moment, which is composed of the predicted Euler angle Transformed, Represents the true value of the rotation matrix corresponding to the kth time step, which is converted from the true value quaternion; Step 5-3: In order to make full use of the geometric information provided by multiple labels, a joint loss function is obtained through linear transformation , its mathematical representation is: , , in represents the position of the jth tag in the carrier coordinate system, and its dimension is 1×3×1; represents the position of the jth label after expansion in the carrier coordinate system, and its dimension is batch×1×3. Indicates the position of the jth label at the kth time step; Step 5-4, establish the complete loss function L: , in , , Represent the loss function The weight of The weight and The weight of

[0028] In this experiment, we set is 1, is 0.6, is 0.6.

[0029] Step 5-5, using the loss function established in step 5-4, train the indoor positioning model established in steps 1 to 4 through supervised training.

[0030] Using the established cost function, the model established in step 1 to step 4 is trained through supervised training to achieve the purpose of indoor positioning; Through the above steps and methods, the problems of insufficient estimation of carrier attitude and large ranging error in complex environments in ultra-wideband positioning can be effectively solved, the positioning accuracy can be improved, and the attitude estimation can be strengthened, so that the two complement each other.

[0031] In order to prove the superiority of the method of the present invention, some comparative experiments were designed to test the changes in the error of multi-label position prediction in a single living room environment under different loss functions. The experimental results are as follows: First, the loss function The indoor positioning model trained by this method does not have any posture constraints and does not display the posture error. The average error (measured by RMSE) of the label position in multiple test sequences is about 0.153 meters. With loss function After training the model, the average error of label position is about 0.122 meters, and the average error of pose is about 26°; With loss function After training the model, the average error of label position is about 0.109 meters, and the average error of posture is about 21°; Specifically, the position error and attitude error is defined as follows: , , in Indicates the number of test sets under different test sequences;

[0032] Experiments have proved that the construction of the loss function in this method has the function of predicting posture and assisting positioning at the same time.

[0033] The present invention provides a multi-tag ultra-wideband sensor positioning method based on a state space model. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A multi-tag ultra-wideband sensor positioning method based on a state space model, characterized in that: The following steps are involved: Step 1: embedding base station information based on the pre-calibrated base station position in the indoor environment and the continuous ranging value between the ultra-wideband sensor tag and the base station; Step 2: segment, map and fuse the embedded data to obtain fusion features; Step 3, construct a state space module based on the state space model to extract the spatiotemporal characteristics of the ultra-wideband sensor measurement value; Step 4, establishing an ultra-wideband sensor tag position and posture prediction module to achieve synchronous prediction of tag position and posture; the state space module and the ultra-wideband sensor tag position and posture prediction module form an indoor positioning model; Step 5, design a network training method based on supervised learning, adopt a supervised learning strategy, based on the predicted value of the ultra-wideband sensor tag position and the predicted value of the Euler angle representing the posture obtained by the ultra-wideband sensor tag position and posture prediction module, the pre-collected true quaternion, and the true position value, construct a loss function based on position constraints, posture constraints and joint constraints, which is used to train the indoor positioning model and realize the synchronous prediction of the tag position and posture.

2. The method according to claim 1, characterized in that: Step 1 includes: Step 1-1, arrange N ultra-wideband sensor base stations indoors, and the coordinates of the N ultra-wideband sensor base stations in the world coordinate system have been calibrated in advance; at the same time, M ultra-wideband sensor tags are fixedly installed on any movable carrier; collect continuous measurement values ​​between a tag and a base station as samples, collect samples within more than two time steps to obtain the original ranging information, separate the original ranging information along the base station dimension, and independently process the ranging data between each base station and the tag. The formula is: ， in Indicates equal division along the dimension of the base station, Represents the original ranging information, represents the continuous ranging information between multiple tags and the i-th base station, The dimension is batch×M×10; batch represents the batch size; Step 1-2, convert the single base station position into a feature vector of the same size as the single base station ranging information through a shared fully connected layer, and expand the feature vector at the same time, the formula is: ， in represents the coordinates of the i-th base station in the world coordinate system under the batch time step, express Expand the batch times in the first dimension, represents a shared linear transformation from 3D to 10D, represents the transformed and expanded single base station feature vector, where The dimension is 1×3, The dimension is batch×10; Step 1-3, the transformed and expanded single base station feature vector is further expanded along the label dimension, and the formula is: ， ， in Express Perform dimension transformation, Adjust to shape, express Expand M times in the second dimension, represents the reshaped single base station feature vector, Represents the feature vector of the i-th base station after expansion, with a dimension of batch×M×10; Step 1-4, embed the ranging information between the tag and the single base station into the processed feature vector of the corresponding base station and splice it into new fused data.

3. The method according to claim 2, characterized in that In steps 1-4, the following formula is used to embed the ranging information between the tag and the single base station into the processed feature vector of the corresponding base station and splice it into new fused data: ， in Represents a splicing operation, Represents the fused data after embedding, with dimension , where the parameters The value of is N×10.

4. The method according to claim 3, characterized in that Step 2 includes: Step 2-1, segment the continuous ranging information within a fixed time step from the embedded data, specify the batch size batch, select a suitable number from the integer divisor of the batch as the fixed time step size T; separate the ranging information of each ultra-wideband sensor tag separately, and project the ranging information after fusion of base station information to dimensional feature space, the formula is: ， in represents the ranging information of the jth tag at the kth time step, Indicated by VI dimensional linear transformation, It means that the dimension after linear transformation is The feature vector of the j-th label at the k-th time step; Step 2-2, concatenate and fuse the multi-label feature vectors extracted at a time. The fusion method is to concatenate and pass them through a fully connected layer together to map the multi-label feature vectors to In the feature space of dimension, the formula is: ， in Indicated by VI dimensional linear transformation, Represents the concatenated feature vector of multiple labels at the kth time step, and its dimension is ; The multi-label feature vectors at all times within a fixed time step are concatenated, and the formula is: ， in Indicates The multi-label feature vectors of all moments within a fixed time step T are of dimension ; Representative The concatenated feature vector of multiple labels at time steps; Step 2-3, continue to extract the multi-label feature vectors of all moments within the fixed time step T from the data of batch time steps according to steps 2-1 and 2-2, and then perform splicing and fusion. is the parameter S, where Indicates that an integer is rounded down. The formula is: ， in, Indicates The multi-label feature vectors of all moments within a fixed time step T are grouped together. It is the fusion feature obtained after the final processing, and the dimension is .

5. The method according to claim 4, characterized in that Step 3 includes: Step 3-1, extract the features of each ultra-wideband sensor tag from the fusion feature F separately: regard the fusion feature F as M dimensions The independent features of the single label are placed in the state space module based on the state space model to deeply extract the long-term spatiotemporal dependency, while effectively suppressing noise interference. The formula is: ， in represents the state-space model, represents the time step size of the data passed into the state-space model, The value of is equal to S, represents the sequence length of the time series, The value of is equal to T, represents the feature dimension of the input, The value of equal, represents the feature dimension of the output, The value and input dimension same; The state space module includes linear layers, convolutional layers, activation layers and state space models. For independent features of a label , the state-space module performs the following transformations: First, the data is normalized, the formula is: ， ， in Represents the layer normalization operation, where It represents the result of normalizing the independent feature layer of the label, and the dimension is still ; Indicates independent features The residual of Secondly, keep the normalized data and use it as the residual after processing through the linear layer and activation layer. At the same time, let the normalized data pass through the linear layer and convolution layer in turn to perform linear transformation, convolution and activation operations on the data. The specific formula is: ， ， Where ReLU represents the activation function, and Is VI Conv represents a one-dimensional convolution operation. The result after linear transformation and activation operation of the normalized data of the representation layer. Represents the captured local spatiotemporal activation features; Then, the activated data is passed into the state space model and compared with the residual The corresponding elements are multiplied, and the formula is: ， in represents the state space model, ⊙ represents the multiplication of corresponding elements, Indicates long-term dependence on fusion features, The dimension is ; Finally, the dimension is restored after linear transformation and combined with the independent features The residual Connect, fuse the original input and long-term dependent fusion features, the formula is: ， in Indicated by VI dimensional linear transformation, is the output of the state space module; Step 3-2: In step 3-1, the high-level feature representations of the M single labels obtained after passing through the state space module are , the dimension of each high-level feature representation is , concatenate the M single-label high-level feature representations and reshape them to obtain a dimension of The data is: ， in It represents the result of concatenating M high-level features of a single label, and its dimension is ; S The value of T is batch.

6. The method according to claim 5, characterized in that Step 4 includes: Step 4-1, for the estimation of the position of the ultra-wideband sensor tag: the ultra-wideband sensor tag position and posture prediction module includes two fully connected layers, M high-level feature representations of single labels Share two fully connected layers, and sequentially change the dimensions from The dimension is reduced to 32 dimensions, and then from 32 dimensions to 3 dimensions, and the three-dimensional coordinates of M labels in the world coordinate system are obtained in turn. , three-dimensional coordinates The dimensions are all batch×3, and the three-dimensional coordinates For splicing, the formula is: ， in Represents the position prediction value of M labels, the dimension is batch×M×3; Step 4-2, estimate the posture.

7. The method according to claim 6, characterized in that Step 4-2 includes: concatenating the M high-level features of a single label obtained in step 3-2 Through two fully connected layers, the dimensions are changed from Dimension reduction Wei, then from The dimension is reduced to 3 dimensions, and the predicted value of the Euler angle representing the posture is obtained .

8. The method according to claim 7, characterized in that Step 5 includes: Step 5-1: Establish a loss function for the prediction of the locations of more than two ultra-wideband sensor tags : ， in represents the training batch size, The value of is the same as batch. represents the predicted value of the label position at the kth time step, Represents the true value of the label position at the kth time step; Step 5-2: For posture estimation, the indoor positioning model predicts the Euler angle ,Will Convert to rotation matrix and establish loss function : ， in Represents the conversion of rotation matrix to Euler angle, represents the predicted value of the rotation matrix from the carrier coordinate system to the world coordinate system at the kth time step, Predicted values ​​from Euler angles Converted to Represents the true value of the rotation matrix corresponding to the kth time step from the carrier coordinate system to the world coordinate system; Step 5-3, obtain a joint loss function through linear transformation : ， ， in Indicates the position of the jth tag in the carrier coordinate system, with a dimension of 1×3×1. It represents the position of the jth label in the carrier coordinate system after expansion, with the dimension of batch×1×3. Indicates the position of the jth label at the kth time step; Step 5-4, establish the complete loss function L: ， in , , Represent the loss function The weight of The weight and The weight of Step 5-5, using the loss function established in step 5-4, train the indoor positioning model through supervised training.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 8.

10. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 8 are executed.

Images