A high-resolution reconstruction and positioning method and system based on sparse fingerprints

Abstract

The present invention provides a high-resolution reconstruction and positioning method and system based on sparse fingerprints. By setting reference points, received signal strength indication data of a target area is collected and processed to construct an offline fingerprint database. Then, a sparse fingerprint image reconstruction network combining historical fingerprint maps and real-time sparse fingerprint maps is designed and trained using the offline fingerprint database. In the online phase, after collecting the real-time sparse fingerprint map of the positioning target, training is performed using the historical fingerprint map and the real-time sparse fingerprint map, so as to achieve the purpose of high-precision fingerprint map reconstruction and positioning in dynamic and complex environments. The present invention reduces the maintenance cost and saves computing resources, while realizing the reconstruction of the current high-resolution fingerprint map, improving the positioning accuracy within the target area, and greatly enhancing the adaptability of the positioning system in dynamic and complex environments.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technologies, and particularly to a high-resolution reconstruction positioning method and system based on sparse fingerprints. Background Art

[0002] With the development of wireless communication technologies, indoor positioning systems are increasingly widely used in fields such as intelligent buildings, logistics management, and precision medicine. As a common indoor positioning method, Wi-Fi fingerprint positioning constructs a fingerprint database by collecting signal strength data at different positions offline and performs matching in the online stage to achieve precise positioning. However, over time and with environmental changes (such as indoor layout adjustments, personnel flow, equipment differences, etc.), the timeliness of the fingerprint database gradually decreases, resulting in a decline in positioning accuracy and even complete failure in complex environments eventually.

[0003] In the prior art, to solve the timeliness problem of the fingerprint database, increasing the fingerprint collection frequency or introducing environmental modeling is usually adopted. However, these methods often require high computational costs or collection costs and are difficult to maintain high robustness in a dynamically changing environment. Therefore, how to achieve high-precision and high-robustness positioning in complex and changing environments has become a key challenge in current technologies.

[0004] In recent years, deep learning technologies have made significant progress in the fields of image processing and signal processing, especially in image super-resolution and feature reconstruction. However, existing deep learning technologies are mostly applied to traditional image reconstruction tasks, and relevant research on the reconstruction of sparse fingerprint maps and the improvement of positioning accuracy in dynamic environments is not sufficient. Summary of the Invention

[0005] Object of the Invention: The first object of the present invention is to provide a high-resolution reconstruction positioning method based on sparse fingerprints that reduces energy consumption while ensuring accuracy, and the second object is to provide a high-resolution reconstruction positioning system based on sparse fingerprints.

[0006] Technical Solution: A high-resolution reconstruction positioning method based on sparse fingerprints includes the following steps:

[0007] S1. In the offline stage, reference points are evenly set in the target area, and the received signal strength indication (RSSI) values of all anchor nodes in the target area received at each reference point are collected to form the fingerprints of the reference points, and the fingerprints of all reference points form a dense fingerprint database;

[0008] S2. The dense fingerprint databases at several time stamps are collected at fixed time intervals to form an interval dense fingerprint database set;

[0009] S3. Extract the data in the interval-dense fingerprint database according to the anchor nodes to obtain the dense fingerprint maps corresponding to each anchor node at each timestamp, forming an interval-dense fingerprint map set. Set a sparsity coefficient to determine the sparse reference points, and extract the sparse fingerprint maps corresponding to each dense fingerprint map according to the sparse reference points, forming an interval-sparse fingerprint map set;

[0010] S4. Normalize the data in the interval-sparse fingerprint map set and the interval-dense fingerprint map set respectively to obtain the corresponding interval-sparse fingerprint pixel map set and interval-dense fingerprint pixel map set; Extract the dense fingerprint pixel map corresponding to the earliest timestamp from the interval-dense fingerprint pixel map set to form a historical dense fingerprint pixel map set, and the remaining dense fingerprint pixel maps form a real-time dense fingerprint pixel map set; Remove the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval-sparse fingerprint pixel map set, and the remaining sparse fingerprint pixel maps form a real-time sparse fingerprint pixel map set;

[0011] S5. Construct a sparse fingerprint image reconstruction network, use the historical dense fingerprint pixel map set and the real-time sparse fingerprint pixel map set as the input of the sparse fingerprint image reconstruction network. For each anchor node, train the sparse fingerprint image reconstruction network with the real-time dense fingerprint pixel map to obtain the optimal sparse fingerprint image reconstruction network corresponding to each anchor node;

[0012] S6. In the online stage, the target to be located collects the RSSI values on all anchor nodes in the target area to form the real-time fingerprint of the target to be located. Normalize the real-time fingerprint to obtain the real-time pixel fingerprint. At the same time, receive the RSSI values on all anchor nodes at the sparse reference points to obtain an online sparse fingerprint map set, which is converted into an online sparse fingerprint pixel map set after normalization. Match the online sparse fingerprint pixel map set with the historical dense fingerprint pixel map set, and use the matched historical dense fingerprint pixel map set and online sparse fingerprint pixel map set as the input of the optimal sparse fingerprint image reconstruction network to output a real-time dense pixel fingerprint set;

[0013] S7. Calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, sort the calculated Euclidean distances, select several real-time dense pixel fingerprints with the smallest Euclidean distances, and perform weighted calculation on the reference point coordinates corresponding to these several real-time dense pixel fingerprints to obtain the coordinates of the target to be located.

[0014] Specifically, in step S1, reference points are uniformly set in the target area at a fixed interval distance, and the number of reference points is:

[0015] ,

[0016] where: is the number of reference points, is the length of the target area, is the width of the target area, is the set fixed interval distance.

[0017] Specifically, in step S3, setting the sparsity coefficient to determine the sparse reference points includes: setting the sparsity coefficient and determining the number of sparse reference points selected for the long side and the wide side of the target area according to the sparsity coefficient:

[0018] ,

[0019] In the formula: is the number of sparse reference points selected for the long side of the target area, is the number of sparse reference points selected for the wide side of the target area; is the number of reference points selected for the long side of the target area; is the number of reference points selected for the wide side of the target area; is the sparsity coefficient.

[0020] Specifically, in step S4, the normalization process includes:

[0021] Set the range of RSSI values to [-100dBm, 0dBm], set the RSSI values less than -100dBm to -100dBm. If the RSSI values of the corresponding anchor nodes are not collected at the reference points or sparse reference points, set the missing RSSI values to -100dBm as well, and then convert all RSSI values to the range of [0, 1] through the normalization operation.

[0022] Specifically, in step S5: The sparse fingerprint image reconstruction network includes an upsampling layer, a channel merging layer, a convolutional encoding layer, a feature fusion module, and a convolutional decoding layer connected in sequence; the convolutional encoding layer includes a first convolutional layer and a second convolutional layer connected in sequence; the feature fusion module includes four residual blocks connected in sequence, and each residual block contains two convolutional layers; the convolutional decoding layer includes a third convolutional layer and a fourth convolutional layer connected in sequence.

[0023] Specifically, in the sparse fingerprint image reconstruction network, the upsampling layer performs upsampling processing on the real-time sparse fingerprint pixel map to align the resolution of the real-time sparse fingerprint pixel map with that of the historical dense fingerprint pixel map; the channel merging layer merges the real-time sparse fingerprint pixel map and the historical dense fingerprint pixel map in the channel direction; the first convolutional layer extracts shallow features from the merged pixel map; the second convolutional layer extracts deep features in the pixel map; the feature fusion module fuses the deep features and the shallow features; the third convolutional layer reduces the dimensionality of the high-dimensional features; the fourth convolutional layer generates the predicted real-time dense fingerprint pixel map.

[0024] Specifically, the convolution kernel of the first convolutional layer has a width of 3, a height of 3, 2 channels, 32 convolution kernels, a padding of 1, and a stride of 1; the convolution kernel of the second convolutional layer has a width of 3, a height of 3, 32 channels, 64 convolution kernels, a padding of 1, and a stride of 1; in the residual block, the convolution kernel of the convolutional layer has a width of 3, a height of 3, 64 channels, 64 convolution kernels, a padding of 1, and a stride of 1; the convolution kernel of the third convolutional layer has a width of 3, a height of 3, 64 channels, 32 convolution kernels, a padding of 1, and a stride of 1; the convolution kernel of the fourth convolutional layer has a width of 3, a height of 3, 32 channels, 1 convolution kernel, a padding of 1, and a stride of 1.

[0025] Specifically, in step S5, the training of the sparse fingerprint image reconstruction network includes:

[0026] Taking the historical dense fingerprint pixel map and the real-time sparse fingerprint pixel map under the anchor node as input data and inputting them into the sparse fingerprint image reconstruction network to calculate the predicted real-time dense fingerprint pixel map; calculating the loss value using the mean square error based on the predicted real-time dense fingerprint pixel map and the real-time dense fingerprint pixel map under the anchor node; performing gradient calculation on the weights and biases of the sparse fingerprint image reconstruction network through the loss value, and propagating the gradient using the chain rule; adjusting the parameters of the sparse fingerprint image reconstruction network based on the update rule of the optimizer, adopting cosine annealing learning rate scheduling, and the learning rate gradually decays during the training process; taking the real-time sparse fingerprint pixel map and the real-time dense fingerprint pixel map under the anchor node as eigenvalues and labels respectively, dividing them into a training set and a validation set according to a set ratio, training on the training set and evaluating the loss on the validation set in each round of iteration; setting the batch size and the number of iterations; saving the model parameters with the best performance according to the change of the loss value to obtain the optimal sparse fingerprint image reconstruction network corresponding to the anchor node, and repeating the above process to obtain the optimal sparse fingerprint image reconstruction network corresponding to all anchor nodes.

[0027] Specifically, in step S7, the weighted calculation of the reference point coordinates corresponding to 3 real-time dense pixel fingerprints includes: calculating the weights of the 3 reference point coordinates respectively using the Euclidean distance:

[0028] ,

[0029] In the formula: is the weight corresponding to the i-th reference point, is the Euclidean distance between the real-time dense pixel fingerprint and the real-time pixel fingerprint of the i-th reference point, is a very small positive number greater than 0 to avoid the Euclidean distance being 0;

[0030] Calculating the coordinates of the target to be located:

[0031] ,

[0032] In the formula: is the abscissa of the target to be located, is the ordinate of the target to be located, is the abscissa of the i-th reference point, is the ordinate of the i-th reference point.

[0033] The present invention also provides a high-resolution reconstruction positioning system based on sparse fingerprints, including:

[0034] Offline information acquisition module: used for evenly setting reference points in the target area during the offline stage, collecting the received signal strength indication (RSSI) values of all anchor nodes in the target area received at each reference point to form the fingerprints of the reference points, and the fingerprints of all reference points form a dense fingerprint library;

[0035] Offline fingerprint library module: used for collecting the dense fingerprint library at a number of timestamps at fixed time intervals to form an interval dense fingerprint library set;

[0036] Fingerprint map generation module: used for extracting the data in the interval dense fingerprint library set according to the anchor nodes, obtaining the dense fingerprint map corresponding to each anchor node at each timestamp, forming an interval dense fingerprint map set, setting a sparse coefficient to determine sparse reference points, and extracting the sparse fingerprint map corresponding to each dense fingerprint map according to the sparse reference points to form an interval sparse fingerprint map set;

[0037] Normalization processing module: used for respectively performing normalization processing on the data in the interval sparse fingerprint map set and the interval dense fingerprint map set to obtain the corresponding interval sparse fingerprint pixel map set and interval dense fingerprint pixel map set; then extracting the dense fingerprint pixel map corresponding to the earliest timestamp from the interval dense fingerprint pixel map set to form a historical dense fingerprint pixel map set, and the remaining dense fingerprint pixel maps form a real-time dense fingerprint pixel map set; removing the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval sparse fingerprint pixel map set, and the remaining sparse fingerprint pixel maps form a real-time sparse fingerprint pixel map set;

[0038] Network construction and training module: used for constructing a sparse fingerprint image reconstruction network, using the historical dense fingerprint pixel map set and the real-time sparse fingerprint pixel map set as the input of the sparse fingerprint image reconstruction network, and for each anchor node, training the sparse fingerprint image reconstruction network with the real-time dense fingerprint pixel map to obtain the optimal sparse fingerprint image reconstruction network corresponding to each anchor node;

[0039] Online data acquisition module: It is used to collect the RSSI values on all anchor nodes in the target area by the target to be located during the online phase, form the real-time fingerprint of the target to be located, perform normalization processing on the real-time fingerprint to obtain the real-time pixel fingerprint. At the same time, receive the RSSI values on all anchor nodes at the sparse reference points to obtain the online sparse fingerprint atlas, which is converted into an online sparse fingerprint pixel atlas after normalization processing. Match the online sparse fingerprint pixel atlas with the historical dense fingerprint pixel atlas, and use the matched historical dense fingerprint pixel atlas and online sparse fingerprint pixel atlas as the input of the optimal sparse fingerprint image reconstruction network to output the real-time dense pixel fingerprint set;

[0040] Positioning output module: It is used to calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, sort the calculated Euclidean distances, select several real-time dense pixel fingerprints with the smallest Euclidean distances, and perform weighted calculation on the reference point coordinates corresponding to these several real-time dense pixel fingerprints to obtain the coordinates of the target to be located.

[0041] Beneficial effects: Compared with the prior art, the remarkable effects of the present invention are:

[0042] 1. The present invention designs a sparse fingerprint image reconstruction network, combines the historical dense fingerprint pixel map and the sparse fingerprint map obtained by current acquisition and processing for feature learning, realizes the reconstruction of the current high-resolution fingerprint map, improves the positioning accuracy in the target area, and greatly enhances the adaptability of the positioning system in a dynamic and complex environment.

[0043] 2. After adopting the sparse fingerprint image reconstruction network, accurate positioning can be achieved by using the historical dense fingerprint pixel atlas, reducing the update frequency of the fingerprint database, and thus reducing the costs of data acquisition and system maintenance.

[0044] 3. Combining sparse acquisition with deep learning optimizes the computing resources, reduces the resource consumption during the computing process, and thus reduces the dependence on computing resources, being suitable for lightweight deployment. Description of the Drawings

[0045] Figure 1 is the flowchart of the method of the present invention.

[0046] Figure 2 is the schematic diagram of the sparse fingerprint pixel map of the present invention.

[0047] Figure 3 is the flowchart of the sparse fingerprint image reconstruction network of the present invention.

[0048] Figure 4 is the high-resolution reconstruction effect diagram of the sparse fingerprint image reconstruction network of the present invention.

[0049] Figure 5 This is a schematic diagram for predicting the location of the user to be located in the present invention. Specific Embodiment

[0050] The following further describes a preferred solution of the present invention in conjunction with the accompanying drawings.

[0051] Embodiment 1

[0052] Please refer to Figure 1 As shown, this embodiment provides a high-resolution reconstruction positioning method based on sparse fingerprints, including the following steps:

[0053] S1. In the offline stage, reference points are evenly set in the target area, and the received signal strength indication (RSSI) values of all anchor nodes in the target area received at each reference point are collected to form the fingerprint Fp (Fingerprint) of the reference point. The fingerprints of all reference points constitute the dense fingerprint database FD (Fingerprint Database).

[0054] In this embodiment, reference points are evenly set in the target area at a fixed interval distance, and the number of reference points is:

[0055] ,

[0056] In the formula: is the number of reference points, is the length of the target area, is the width of the target area, is the set fixed interval distance.

[0057] Taking any reference point i as an example, at time stamp the fingerprint composed of the RSSI values of p anchor nodes received in the positioning area is denoted as :

[0058] ,

[0059] In the formula: represents the received signal strength indication (RSSI) value of anchor node j received by reference point i at time stamp .

[0060] S2. Collect the dense fingerprint databases at several time stamps at a fixed time interval to form an interval dense fingerprint database set.

[0061] In this embodiment, the dense fingerprint database composed of the fingerprints received from p anchor nodes by all m reference points in the positioning area at time stamp is :

[0062] ,

[0063] wherein: represents the fingerprint of the reference point k at the time stamp .

[0064] The fingerprint database information is collected at fixed time intervals, and the time interval is , and the dense fingerprint databases at x time stamps collected form an interval dense fingerprint database set :

[0065] ,

[0066] wherein: represents the dense fingerprint database at the time stamp u, and the relationship between time stamps is as follows:

[0067] .

[0068] S3. Extract the data in the interval dense fingerprint database set according to the anchor nodes to obtain the dense fingerprint maps corresponding to each anchor node at each time stamp, form an interval dense fingerprint map set, set a sparsity coefficient to determine the sparse reference points, and extract the sparse fingerprint maps corresponding to each dense fingerprint map according to the sparse reference points to form an interval sparse fingerprint map set.

[0069] In this embodiment, there are p anchor nodes in the target area, and the data in the interval dense fingerprint database set is extracted according to the anchor nodes to form corresponding p dense fingerprint maps. Taking as an example, the corresponding p dense fingerprint maps can be extracted from this fingerprint database, the dense fingerprint map on the anchor node j is as follows:

[0070] ,

[0071] wherein: is the number of reference points set for the long side of the target area, is the number of reference points set for the wide side of the target area.

[0072] The fingerprint database The p dense fingerprint maps formed on p anchor nodes are merged as the dense fingerprint map set at the time stamp :

[0073] .

[0074] Since there are p anchor nodes in the positioning area, there are p dense fingerprint maps formed at each timestamp; the dense fingerprint map sets at all timestamps are combined to form an interval dense fingerprint map set composed of dense fingerprint map sets at x timestamps :

[0075] ,

[0076] To save positioning energy consumption, in the online stage, an online sparse fingerprint map is collected at sparse reference points and reconstructed into a high-resolution dense fingerprint map through a deep learning network, which specifically includes:

[0077] Let the sparse coefficient be , then the number of reference points selected for the corresponding long side and the number of reference points selected for the wide side are:

[0078] ,

[0079] In the formula: is the number of sparse reference points selected for the long side of the target area, is the number of sparse reference points selected for the wide side of the target area; is the number of reference points selected for the long side of the target area; is the number of reference points selected for the wide side of the target area.

[0080] Each dense fingerprint map corresponds to a sparse fingerprint map. Taking the dense fingerprint map on anchor node j in the fingerprint database as an example, the corresponding sparse fingerprint map of anchor node j at timestamp is as follows:

[0081] ,

[0082] In the formula: represents the total number of sparse reference points.

[0083] The p sparse fingerprint maps formed on p anchor nodes in the fingerprint database are combined as the sparse fingerprint map set at timestamp :

[0084] ,

[0085] The interval sparse fingerprint map set corresponding to the interval dense fingerprint map set is composed of x sparse fingerprint map sets at x timestamps

[0086] .

[0087] S4. Normalize the data in the sparse interval fingerprint map set and the dense interval fingerprint map set respectively to obtain the corresponding sparse interval fingerprint pixel map set and dense interval fingerprint pixel map set.

[0088] In this embodiment, to obtain data more suitable for deep learning network optimization, it is necessary to further process the data in the sparse interval fingerprint map set and the dense interval fingerprint map set , specifically:

[0089] Through normalization processing, the RSSI value is normalized and mapped to the range of [0, 1]. Set the RSSI value range to be within [-100dBm, 0dBm]. Set the RSSI value less than -100dBm to -100dBm. For the reference point where the RSSI value of the corresponding anchor node is not collected, also set the RSSI value corresponding to this reference point to the lowest RSSI value of -100dBm; through the normalization operation, all RSSI values are converted to the range of [0, 1]. The specific conversion formula is:

[0090] ,

[0091] ,

[0092] ,

[0093] .

[0094] In the formula: represents the minimum element value in the dense fingerprint map on anchor node j at timestamp u, that is, the minimum RSSI value; represents the maximum element value in the dense fingerprint map on anchor node j at timestamp u, that is, the maximum RSSI value; represents the normalized RSSI value corresponding to anchor node j on dense reference point i at timestamp u; represents the normalized RSSI value corresponding to anchor node j on sparse reference point l at timestamp u; represents the maximum element value in the sparse fingerprint map on anchor node j at timestamp u, that is, the maximum RSSI value; represents the minimum element value in the sparse fingerprint map on anchor node j at timestamp u, that is, the minimum RSSI value; i represents the dense reference point sequence number; j represents the anchor node sequence number; l represents the sparse reference point sequence number.

[0095] Through the above formula and method, process the data in the dense fingerprint map and the sparse fingerprint map to obtain the corresponding dense fingerprint pixel map and sparse fingerprint pixel map as:

[0096] ,

[0097] ,

[0098] wherein: represents the dense fingerprint pixel map under the anchor node j at the time stamp u; represents the sparse fingerprint pixel map under the anchor node j at the time stamp u.

[0099] Correspondingly, integrating the sparse fingerprint pixel maps and the dense fingerprint pixel maps under all anchor nodes forms a sparse fingerprint pixel map set and a dense fingerprint pixel map set :

[0100] ,

[0101] wherein: represents the dense fingerprint pixel map set at the time stamp u; represents the sparse fingerprint pixel map set at the time stamp u.

[0102] Correspondingly, the set of interval dense fingerprint pixel maps formed by the dense fingerprint pixel map sets at x time stamps and the set of interval sparse fingerprint pixel maps formed by the sparse fingerprint pixel map sets at x time stamps can be represented by the following sets:

[0103] ,

[0104] Split the sparse fingerprint pixel maps at x time stamps according to different anchor nodes:

[0105] ,

[0106] wherein: represents the sparse fingerprint pixel map set under the anchor node j.

[0107] Split the dense fingerprint pixel maps at x time stamps according to different anchor nodes:

[0108] ,

[0109] wherein: represents the dense fingerprint pixel map set under the anchor node j.

[0110] Please refer to Figure 2 shown Figure 2 is the sparse fingerprint pixel map of the anchor node 2 at the time stamp t 23 with a size of 16×16 and a sparsity coefficient of , , the anchor node 2 at timestamp t 23 The sparse fingerprint map and the anchor node 2 at timestamp t 23 The sparse fingerprint pixel map The matrix representation is as follows:

[0111] ,

[0112] .

[0113] In the formula: Indicates missing data.

[0114] It can be seen that through the normalization operation, the missing values of the RSSI values are filled, and all the RSSI values are normalized to the range of [0, 1].[[]END]]

[0115] S5. Extract the dense fingerprint pixel map corresponding to the earliest timestamp from the interval dense fingerprint pixel map set to form the historical dense fingerprint pixel map set, and the remaining dense fingerprint pixel maps form the real-time dense fingerprint pixel map set; remove the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval sparse fingerprint pixel map set, and the remaining sparse fingerprint pixel maps form the real-time sparse fingerprint pixel map set; construct a sparse fingerprint image reconstruction network (SFIR, Sparse Fingerprint Image Reproduction), use the historical dense fingerprint pixel map set and the real-time sparse fingerprint pixel map set as the input of the sparse fingerprint image reconstruction network, and for each anchor node, use the real-time dense fingerprint pixel map to train the sparse fingerprint image reconstruction network to obtain the optimal sparse fingerprint image reconstruction network corresponding to each anchor node.

[0116] Taking the anchor node j as an example, extract the dense fingerprint pixel map set under the anchor node j and the corresponding sparse fingerprint pixel map set , select the dense fingerprint pixel map corresponding to the earliest timestamp among the x timestamps in the dense fingerprint pixel map set under the anchor node j as the historical dense fingerprint pixel map under the anchor node j, and the data size is: ; the remaining x - 1 dense fingerprint pixel maps are used as the real-time dense fingerprint pixel maps under the anchor node j, and the data size is: ; remove the earliest timestamp from the x timestamps in the sparse fingerprint pixel map set under the anchor node j, and the remaining x - 1 sparse fingerprint pixel maps are used as the real-time sparse fingerprint pixel maps under the anchor node j, and the data size is: .

[0117] Next, taking the anchor node j as an example, the specific structure of the SFIR network is described:

[0118] Please refer toFigure 3 As shown in Figure 3 , the SFIR network includes an upsampling layer, a channel merging layer, a convolutional encoding layer, a feature fusion module, and a convolutional decoding layer connected in sequence; the convolutional encoding layer includes a first convolutional layer and a second convolutional layer connected in sequence; the feature fusion module includes four residual blocks connected in sequence, and each residual block contains two convolutional layers; the convolutional decoding layer includes a third convolutional layer and a fourth convolutional layer connected in sequence.

[0119] The input of the upsampling layer is: the historical dense fingerprint pixel map under anchor node j and the real-time sparse fingerprint pixel map under anchor node j; the output is: the real-time upsampled fingerprint pixel map under anchor node j after upsampling and the historical dense fingerprint pixel map under anchor node j; the input data size is: , and the output data size is: , and the function of this layer is to upsample the real-time sparse fingerprint pixel map under anchor node j, from low resolution to high resolution for alignment with the historical dense fingerprint pixel map.

[0120] The input of the channel merging layer is: the real-time upsampled fingerprint pixel map under anchor node j after upsampling and the historical dense fingerprint pixel map under anchor node j; the input data sizes are: and ; the output data size is: ; the function of this layer is to merge the two pixel maps in the channel direction to form a two-channel input tensor.

[0121] The first convolutional layer: Input: , Output: , Convolution kernel size: , indicating that the width of the convolution kernel is 3, the height is 3, the number of channels is 2, and the number of convolution kernels is 32. The following uses this representation, padding is 1, stride is 1, and the activation function is set to the ReLU function. The function of this layer is to extract shallow features, and the convolution kernels extract features from the merged image.

[0122] The second convolutional layer: Input: , Output: , Convolution kernel size: , Padding is 1, stride is 1, and the activation function is set to the ReLU function. The function of this layer is to extract deeper features, and the convolution kernels deepen the network feature representation.

[0123] The feature fusion module: contains four residual blocks with the same structure, Input: , Output: , and each residual block contains two convolutional layers, and the convolution kernel size is: , filled with 1, with a step size of 1. The activation function of the previous convolutional layer is the ReLU function, and the activation function of the subsequent convolutional layer is not set, which is used to perform skip connections for the residual blocks; repeat 4 residual blocks, and the input and output sizes always remain , and the function of this layer is to retain the input features, while learning the residual information and fusing the deep and shallow features.

[0124] The third convolutional layer: Input: , Output: , Convolution kernel size: , filled with 1, with a step size of 1, and the activation function is the ReLU function. The function of this layer is to reduce the high-dimensional features to low-dimensional features in preparation for the output stage.

[0125] The fourth convolutional layer: Input: , Output: , Convolution kernel size: , filled with 1, with a step size of 1, and the activation function is a linear function. The function of this layer is to generate the predicted real-time dense fingerprint pixel map under the anchor node j.

[0126] Output layer: Output size: , and the function of this layer is to output the predicted real-time dense fingerprint pixel map under the anchor node j.

[0127] SFIR network training stage:

[0128] For p anchor nodes, p SFIR networks need to be trained specifically. Taking the anchor node j as an example, the training process is described as follows:

[0129] The mean squared error loss function is selected as the loss function; the Adam optimizer is used, and the initial learning rate is set to 0.001; the historical dense fingerprint pixel map under the anchor node j and the real-time sparse fingerprint pixel map under the anchor node j are used as input data and input into the corresponding SFIR network of the anchor node j to calculate the predicted output: the predicted real-time dense fingerprint pixel map under the anchor node j; according to the predicted output and the ground truth: the real-time dense fingerprint pixel map under the anchor node j, the loss value is calculated using the mean squared error (MSE); the gradients of the weights and biases of the network are calculated through the loss value, and the gradients are propagated using the chain rule; based on the update rule of the optimizer (Adam), the network parameters are adjusted, and the learning rate gradually decays during the training process, and the cosine annealing learning rate schedule is adopted.

[0130] Take the real-time sparse fingerprint pixel map under the anchor node j and the real-time dense fingerprint pixel map under the anchor node j as eigenvalues and labels respectively, and divide them into a training set and a validation set according to a ratio of 8:2. Training will be performed on the training set and the loss will be evaluated on the validation set in each iteration (Epoch). The batch size is set to 32 and the number of iterations (Epoch) is set to 30; according to the change of the validation loss, save the model parameters with the best performance and use the early stopping strategy to avoid overfitting.

[0131] Using the same method as above, the SFIR network for determining the optimal model parameters corresponding to all anchor nodes can be obtained, denoted as :

[0132] .

[0133] In the formula: represents the SFIR network for determining the optimal model parameters corresponding to the j-th anchor node.

[0134] Please refer to Figure 4 as shown, for the anchor node 2 corresponding to Figure 2 at timestamp t 23 The predicted real-time dense fingerprint pixel map obtained by high-resolution reconstruction of the sparse fingerprint pixel map through the SFIR network, and the size is restored to 64×64.

[0135] S6. In the online stage, the target to be located collects the RSSI values on all anchor nodes in the target area to form the real-time fingerprint of the target to be located, normalizes the real-time fingerprint to obtain the real-time pixel fingerprint. At the same time, the RSSI values on all anchor nodes are received at the sparse reference point to obtain the online sparse fingerprint atlas, which is converted into an online sparse fingerprint pixel atlas after normalization. The online sparse fingerprint pixel atlas is matched with the historical dense fingerprint pixel atlas, and the matched historical dense fingerprint pixel atlas and the online sparse fingerprint pixel atlas are used as the input of the optimal sparse fingerprint image reconstruction network to output the real-time dense pixel fingerprint set.

[0136] In this embodiment, the RSSI values corresponding to p anchor nodes in the positioning area are collected by the user to be located as the real-time fingerprint :

[0137] ,

[0138] In the formula: represents the RSSI value of the anchor node j received by the user to be located.

[0139] Perform a normalization operation on the real-time fingerprint to obtain the corresponding normalized real-time fingerprint :

[0140] ,

[0141] wherein: represents the normalized RSSI value of the anchor node j received by the user to be located.

[0142] Meanwhile, a positioning request is sent to the positioning system, and the sparse reference points receive the RSSI values on all anchor nodes to form p online sparse fingerprint maps, which constitute an online sparse fingerprint map set :

[0143] ,

[0144] wherein: represents the online sparse fingerprint map under the anchor node p.

[0145] Through the conversion method of the sparse fingerprint pixel map recorded in step S4, the online sparse fingerprint map is converted into an online sparse fingerprint pixel map. Taking the anchor node j as an example, the online sparse fingerprint pixel map under the anchor node j is expressed as: ; then the online sparse fingerprint pixel maps under the corresponding p anchor nodes constitute an online sparse fingerprint pixel map set :

[0146] .

[0147] Match the online sparse fingerprint pixel map set with the historical dense fingerprint pixel map set saved in the network. The corresponding historical dense fingerprint pixel map set is :

[0148] ,

[0149] wherein: represents the historical dense fingerprint pixel map under the anchor node p at the earliest timestamp.

[0150] The matched data is used as the online network input :

[0151] ,

[0152] wherein: represents the input data of the SFIR network corresponding to the jth anchor node.

[0153] Bring into to obtain the real-time dense pixel fingerprint set :

[0154] ,

[0155] In the formula: represents the real-time pixel fingerprint at the reference point r.

[0156] S7. Calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, sort the calculated Euclidean distances, select several real-time dense pixel fingerprints with the smallest Euclidean distances, and perform weighted calculation on the reference point coordinates corresponding to these several real-time dense pixel fingerprints to obtain the coordinates of the target to be located.

[0157] In this embodiment: Calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time pixel fingerprint in the real-time dense pixel fingerprint set. The formula is as follows:

[0158] ,

[0159] In the formula: represents the Euclidean distance value between the real-time pixel fingerprint under the reference point r and the normalized real-time pixel fingerprint.

[0160] Sort the calculated Euclidean distances. In this embodiment, select the reference points corresponding to the three real-time pixel fingerprints with the smallest Euclidean distances:

[0161] ,

[0162] The corresponding reference point coordinates are , , , where is the index of the reference point in the real-time dense pixel fingerprint set; calculate the weight of each distance :

[0163] ,

[0164] In the formula: represents a very small positive number to avoid the division-by-zero problem when the distance is zero.

[0165] The calculated corresponding coordinates of the user to be located are :

[0166] ,

[0167] In the formula: is the abscissa of the target to be located, is the ordinate of the target to be located, is the abscissa of the i-th reference point, is the ordinate of the i-th reference point.

[0168] Please refer to Figure 5As shown in the figure, it is a predicted location map of the user to be located obtained by applying the above solution in a specific scenario of the present invention. The normalized real-time pixel fingerprint of the user to be located is: [0.2, 0.3, 0.4, 0.1]; Some of the normalized real-time pixel fingerprint sets of the reference points are shown in Table 1 below:

[0169] Table 1: Real-time pixel fingerprint set (normalized)

[0170] Reference point Reference point A Reference point B Reference point C Reference point D Reference point E Real-time pixel fingerprint [0.1,0.3,0.5,0.0] [0.2,0.4,0.3,0.2] [0.3,0.2,0.4,0.1] [0.0,0.3,0.6, 0.1] [0.2,0.3,0.5,0.0] …… Coordinates (2,3) (5,8) (8,6) (1,1) (3,7) ……

[0171] Calculate the Euclidean distance point by point using the above formula :

[0172] 。

[0173] Sort the Euclidean distances from small to large to obtain the three reference points with the smallest Euclidean distances as reference point C, reference point E, and reference point A; Calculate the weights according to the above formula. Let , and the corresponding weights are obtained as follows:

[0174] 。

[0175] According to the above formula for solving the coordinates of the user to be located, the coordinates of the user to be located are: (x, y) = (4.49, 5.55), and the specific positional relationship is as Figure 5 shown. Through the method provided by the present invention, high-precision positioning tasks can be completed with low energy consumption and high robustness, which has far-reaching research significance and broad application prospects.

[0176] Example 2

[0177] This embodiment provides a high-resolution reconstruction positioning system based on sparse fingerprints corresponding to the high-resolution reconstruction positioning method described in Embodiment 1, including:

[0178] Offline information collection module: Used in the offline stage, reference points are uniformly set in the target area, and the received signal strength indication (RSSI) values of all anchor nodes in the target area received at each reference point are collected to form the fingerprint of the reference point, and the fingerprints of all reference points form a dense fingerprint library;

[0179] Offline fingerprint library module: Used to collect the dense fingerprint library at several time stamps at fixed time intervals to form an interval dense fingerprint library set;

[0180] Fingerprint map generation module: Used to extract the data in the interval dense fingerprint library set according to the anchor nodes, obtain the dense fingerprint map corresponding to each anchor node at each time stamp, form an interval dense fingerprint map set, set a sparse coefficient to determine the sparse reference points, and extract the sparse fingerprint map corresponding to each dense fingerprint map according to the sparse reference points to form an interval sparse fingerprint map set;

[0181] Normalization processing module: used to perform normalization processing on the data in the interval sparse fingerprint map set and the interval dense fingerprint map set respectively, to obtain the corresponding interval sparse fingerprint pixel map set and interval dense fingerprint pixel map set;

[0182] Network construction and training module: used to extract the dense fingerprint pixel map corresponding to the earliest timestamp from the interval dense fingerprint pixel map set to form a historical dense fingerprint pixel map set, and the remaining dense fingerprint pixel maps form a real-time dense fingerprint pixel map set; remove the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval sparse fingerprint pixel map set, and the remaining sparse fingerprint pixel maps form a real-time sparse fingerprint pixel map set; construct a sparse fingerprint image reconstruction network, use the historical dense fingerprint pixel map set and the real-time sparse fingerprint pixel map set as the input of the sparse fingerprint image reconstruction network, and for each anchor node, use the real-time dense fingerprint pixel map to train the sparse fingerprint image reconstruction network to obtain the optimal sparse fingerprint image reconstruction network corresponding to each anchor node;

[0183] Online data acquisition module: used in the online stage, the target to be located collects the RSSI values on all anchor nodes in the target area to form the real-time fingerprint of the target to be located, performs normalization processing on the real-time fingerprint to obtain the real-time pixel fingerprint, and at the same time receives the RSSI values on all anchor nodes at the sparse reference point to obtain an online sparse fingerprint map set, which is converted into an online sparse fingerprint pixel map set after normalization processing. Match the online sparse fingerprint pixel map set with the historical dense fingerprint pixel map set, and use the matched historical dense fingerprint pixel map set and online sparse fingerprint pixel map set as the input of the optimal sparse fingerprint image reconstruction network to output a real-time dense pixel fingerprint set;

[0184] Positioning output module: used to calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, sort the calculated Euclidean distances, select several real-time dense pixel fingerprints with the smallest Euclidean distance, and perform weighted calculation on the reference point coordinates corresponding to these several real-time dense pixel fingerprints to obtain the coordinates of the target to be located.

Claims

1. A high-resolution reconstruction positioning method based on sparse fingerprints, characterized in that: The following steps are involved: S1, in the offline stage, reference points are evenly set in the target area and the RSSI values ​​of all anchor nodes are received to obtain the fingerprints of the reference points. The fingerprints of all reference points constitute a dense fingerprint library; S2, collecting dense fingerprint libraries at several timestamps to form an interval dense fingerprint library set; S3, extracting data from the interval dense fingerprint library, obtaining a dense fingerprint map of each anchor node, forming an interval dense fingerprint map set, setting a sparse reference point, and extracting a sparse fingerprint map corresponding to the dense fingerprint map, forming an interval sparse fingerprint map set; S4, normalizing the interval sparse fingerprint atlas and the interval dense fingerprint atlas to obtain the interval sparse fingerprint pixel atlas and the interval dense fingerprint pixel atlas, respectively, and extracting the historical dense fingerprint pixel atlas, the real-time dense fingerprint pixel atlas and the real-time sparse fingerprint pixel atlas; S5, constructing a sparse fingerprint image reconstruction SFIR network, taking the historical dense fingerprint pixel atlas and the real-time sparse fingerprint pixel atlas as input, using the real-time dense fingerprint pixel atlas to train the SFIR network, and obtaining the optimal SFIR network corresponding to each anchor node; S6, in the online stage, the real-time pixel fingerprint of the target to be located is obtained, and the online sparse fingerprint pixel atlas corresponding to the sparse reference point is obtained, the online sparse fingerprint pixel atlas is matched with the historical dense fingerprint pixel atlas, and the atlas is used as the input of the corresponding optimal SFIR network, and the real-time dense pixel fingerprint set is output; S7. Calculate the Euclidean distance between the real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, select several real-time dense pixel fingerprints with the smallest Euclidean distance, perform weighted calculation on the corresponding reference point coordinates, and obtain the coordinates of the target to be located.

2. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1 is characterized in that: In step S1, reference points are evenly set in the target area, and the number of reference points is: ， Where: is the number of reference points, is the length of the target area, is the width of the target area, The fixed interval distance is set.

3. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In step S3, setting the sparse reference points includes: setting a sparse coefficient, and determining the number of sparse reference points selected for the long side and the wide side of the target area according to the sparse coefficient: ， Where: The number of sparse reference points selected for the long side of the target area, The number of sparse reference points selected for the broad side of the target area; The number of reference points selected for the long side of the target area; The number of reference points selected for the broadside of the target area; is the sparse coefficient.

4. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In step S4, the normalization process includes: The RSSI value range is set to [-100dBm, 0dBm], and the RSSI value less than -100dBm is set to -100dBm. If the RSSI value of the corresponding anchor node is not collected at the reference point or sparse reference point, the missing RSSI value is also set to -100dBm, and then all RSSI values ​​are converted to the range of [0, 1] through normalization.

5. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In step S5: the sparse fingerprint image reconstruction network includes an upsampling layer, a channel merging layer, a convolutional coding layer, a feature fusion module and a convolutional decoding layer connected in sequence; the convolutional coding layer includes a first convolutional layer and a second convolutional layer connected in sequence; the feature fusion module includes four residual blocks connected in sequence, each residual block includes two convolutional layers; the convolutional decoding layer includes a third convolutional layer and a fourth convolutional layer connected in sequence.

6. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 5, characterized in that: In the sparse fingerprint image reconstruction network, the upsampling layer performs upsampling processing on the real-time sparse fingerprint pixel map, and aligns the resolution of the real-time sparse fingerprint pixel map with the historical dense fingerprint pixel map; the channel merging layer merges the real-time sparse fingerprint pixel map with the historical dense fingerprint pixel map in the channel direction; the first convolution layer extracts shallow features from the merged pixel map; the second convolution layer extracts deep features in the pixel map; the feature fusion module fuses deep features with shallow features; the third convolution layer reduces the dimension of high-dimensional features; and the fourth convolution layer generates a predicted real-time dense fingerprint pixel map.

7. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In the step S4, respectively extracting the historical dense fingerprint pixel atlas, the real-time dense fingerprint pixel atlas and the real-time sparse fingerprint pixel atlas includes: extracting the dense fingerprint pixel map corresponding to the earliest timestamp from the interval dense fingerprint pixel atlas to form the historical dense fingerprint pixel atlas, and the remaining dense fingerprint pixel maps constitute the real-time dense fingerprint pixel atlas; removing the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval sparse fingerprint pixel atlas, and the remaining sparse fingerprint pixel maps constitute the real-time sparse fingerprint pixel atlas.

8. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In step S5, the training of the sparse fingerprint image reconstruction network includes: The historical dense fingerprint pixel map and the real-time sparse fingerprint pixel map under the anchor node are input as input data into the sparse fingerprint image reconstruction network, and the real-time dense fingerprint pixel map is calculated and predicted; the loss value is calculated using the mean square error according to the predicted real-time dense fingerprint pixel map and the real-time dense fingerprint pixel map under the anchor node; the weights and biases of the sparse fingerprint image reconstruction network are gradient calculated by the loss value, and the gradient is propagated using the chain rule; based on the update rule of the optimizer, the parameters of the sparse fingerprint image reconstruction network are adjusted, and the cosine annealing learning rate scheduling is adopted, and the learning rate gradually decays during the training process; the real-time sparse fingerprint pixel map and the real-time dense fingerprint pixel map under the anchor node are used as feature values ​​and labels respectively, and are divided into training sets and validation sets according to the set ratio. Each round of iteration will be trained on the training set and the loss will be evaluated on the validation set; the batch size and the number of iterations are set; according to the change of the loss value, the model parameters with the best performance are saved to obtain the optimal sparse fingerprint image reconstruction network corresponding to the anchor node, and the above process is repeated to obtain the optimal sparse fingerprint image reconstruction network corresponding to all anchor nodes.

9. The high-resolution reconstruction positioning method based on sparse fingerprint according to claim 1, characterized in that: In step S7, weighted calculation of the reference point coordinates corresponding to the three real-time dense pixel fingerprints includes: using Euclidean distance to calculate the weights of the three reference point coordinates respectively: ， Where: is the weight corresponding to the i-th reference point, is the Euclidean distance between the real-time dense pixel fingerprint of the i-th reference point and the real-time pixel fingerprint, It is a very small positive number greater than 0 to avoid the Euclidean distance being 0; Calculate the coordinates of the target to be located: ， Where: is the horizontal coordinate of the target to be located, is the ordinate of the target to be located, is the horizontal coordinate of the i-th reference point, is the ordinate of the i-th reference point.

10. A high-resolution reconstruction positioning system based on sparse fingerprints, characterized in that: include: Offline information collection module: used to evenly set reference points in the target area during the offline phase, collect the RSSI values ​​of all anchor nodes in the target area received at each reference point to form the fingerprint of the reference point, and the fingerprints of all reference points constitute a dense fingerprint library; Offline fingerprint library module: used to collect dense fingerprint libraries at several timestamps at fixed time intervals to form an interval dense fingerprint library set; Fingerprint map generation module: used to extract the data in the interval dense fingerprint library according to the anchor node, obtain the dense fingerprint map corresponding to each anchor node at each time stamp, form an interval dense fingerprint map set, set the sparse coefficient to determine the sparse reference point, and extract the sparse fingerprint map corresponding to each dense fingerprint map according to the sparse reference point to form an interval sparse fingerprint map set; Normalization processing module: used to normalize the data in the interval sparse fingerprint atlas and the interval dense fingerprint atlas respectively, to obtain the corresponding interval sparse fingerprint pixel atlas and the interval dense fingerprint pixel atlas; then extract the dense fingerprint pixel map corresponding to the earliest timestamp from the interval dense fingerprint pixel atlas to form the historical dense fingerprint pixel atlas, and the remaining dense fingerprint pixel maps form the real-time dense fingerprint pixel atlas; remove the sparse fingerprint pixel map corresponding to the earliest timestamp from the interval sparse fingerprint pixel atlas, and the remaining sparse fingerprint pixel maps form the real-time sparse fingerprint pixel atlas; Network construction and training module: used to construct a sparse fingerprint image reconstruction network. The historical dense fingerprint pixel atlas and the real-time sparse fingerprint pixel atlas are used as the input of the sparse fingerprint image reconstruction network. For each anchor node, the sparse fingerprint image reconstruction network is trained using the real-time dense fingerprint pixel map to obtain the optimal sparse fingerprint image reconstruction network corresponding to each anchor node. Online data acquisition module: used for collecting RSSI values ​​of all anchor nodes in the target area of ​​the target to be located in the online stage to form the real-time fingerprint of the target to be located, normalizing the real-time fingerprint to obtain the real-time pixel fingerprint, and receiving the RSSI values ​​of all anchor nodes at the sparse reference point to obtain the online sparse fingerprint atlas, which is converted into the online sparse fingerprint pixel atlas after normalization, and the online sparse fingerprint pixel atlas is matched with the historical dense fingerprint pixel atlas, and the matched historical dense fingerprint pixel atlas and the online sparse fingerprint pixel atlas are used as the input of the optimal sparse fingerprint image reconstruction network, and the real-time dense pixel fingerprint set is output; Positioning output module: used to calculate the Euclidean distance between the normalized real-time pixel fingerprint and each real-time dense pixel fingerprint in the real-time dense pixel fingerprint set, sort the calculated Euclidean distances, select several real-time dense pixel fingerprints with the smallest Euclidean distances, and perform weighted calculation on the reference point coordinates corresponding to these several real-time dense pixel fingerprints to obtain the coordinates of the target to be located.

Images