3D indoor radio map estimation method and system based on 2D deep learning

Abstract

The invention discloses a 3D indoor radio map estimation method (R2Net) based on 2D deep learning, a 3D indoor radio map is estimated according to a building layout map, a furniture layout map and a transmitter position, and the influence of the height of an indoor transmitter, the height of a receiver and the height of an obstacle on the radio map is considered. Since the input and output of the 2D deep learning algorithm are 2D pictures, the invention proposes that the heights of buildings, furniture and transmitters are represented by pixel values, and 3D radio maps with different heights are generated by using different output channels. And 3D radio map estimation is converted into a 2D deep learning task, so that the calculation overhead and the memory demand are reduced. The 2D deep learning method R2Net provided by the invention has strong generalization ability, enhances feature extraction according to indoor path loss features, and improves estimation precision and efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a 3D indoor radio map estimation method and system based on 2D deep learning. Background Art

[0002] Radio maps describe the spatial distribution of received signal strength in the form of images. The position of each pixel in the image corresponds to the physical position of the receiver, and its pixel value represents the path loss value. Radio maps have broad application prospects, such as fingerprint positioning, robot path planning, aerial base station deployment, interference management, etc. For example, using radio maps, computer vision technology can be used to improve the accuracy and efficiency of fingerprint positioning.

[0003] The primary problem in using radio maps to empower related applications is the acquisition of radio maps. Traditional radio maps are mostly obtained through field surveys, but due to physical limitations, there are some inaccessible areas, resulting in missing radio map data. In order to complete the radio map, radio map reconstruction that estimates the unmeasured path loss based on measured values ​​has received widespread attention. Field surveys require that base stations have been deployed. However, base stations are not deployed during the network planning phase, and radio maps are also important for optimizing the base station deployment locations. Physical simulation methods can search for possible propagation paths between transmitters and receivers based on the provided environmental data without measured values, and calculate path losses based on antenna characteristics, obstacle materials, scattering, and diffraction parameters. Among them, searching for radio propagation paths requires a huge computational overhead, which makes physical simulation time-consuming. Therefore, efficient radio map estimation methods need to be studied urgently.

[0004] Due to the characteristics of fast inference speed and high accuracy of deep learning, radio map estimation based on deep learning has attracted widespread attention. Existing indoor radio map estimation based on deep learning mainly focuses on 2D radio maps, ignoring the influence of the height of transmitters, receivers and obstacles, while height must be considered in indoor scenes. Although 3D indoor radio map estimation can be achieved by changing the 2D operation of existing 2D radio map estimation methods to 3D, the computational overhead of 3D deep learning algorithms is large. Therefore, further research is needed on computationally efficient 3D indoor radio map estimation. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a 3D indoor radio map estimation method and system based on 2D deep learning, which solves the problem of low efficiency and accuracy of 3D indoor radio map estimation in the prior art.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] A 3D indoor radio map estimation method based on 2D deep learning includes the following steps:

[0008] Step 1: construct the input and output of several 3D radio map estimates obtained by pre-simulation into 2D pictures containing height information, use pixel values ​​to represent the heights of buildings, furniture and transmitters, and use different output channels to generate 3D radio maps of different heights, wherein the 2D pictures include a building layout map, a furniture layout map and a transmitter location map;

[0009] Step 2: construct a 2D deep learning method estimation model based on the characteristics of the 3D radio map, and divide the 2D images obtained in step 1 into a training set, a cross-validation set, and a test set;

[0010] Step 3: Set the training parameters and loss function of the 2D deep learning method estimation model;

[0011] Step 4: Use the training set to train the 2D deep learning method estimation model, use the cross-validation set for verification, and update the parameters of the 2D deep learning method estimation model in real time until the training is completed, and use the test set to evaluate the estimation accuracy of the model;

[0012] Step 5: Input the environmental parameters into the optimal 2D deep learning method estimation model obtained in step 4 to estimate the 3D indoor radio map.

[0013] The specific architecture of the 2D deep learning method estimation model includes an encoder and a decoder, wherein the encoder includes several cascaded residual modules; the decoder includes several nearest neighbor upsampling modules; and a path loss feature enhancement module is connected between the encoder and the decoder.

[0014] The specific process of building a 2D deep learning method estimation model is as follows:

[0015] Step 2.1, based on the U-Net encoder and decoder structure model, according to the characteristics of path loss, random dropout layer and cascade residual module are used to enhance feature extraction;

[0016] Step 2.2: Using the nearest neighbor interpolation algorithm, the decoder restores the features extracted by the encoder to the size of the original image, thereby achieving pixel-level radio map estimation;

[0017] Step 2.3: Use skip connection to pass the features proposed by the encoder to the decoder.

[0018] In step 2.1, for feature maps with relatively low resolution, a random dropout layer is used to enhance generalization; nonlinear feature extraction is enhanced through a cascaded residual module.

[0019] The training parameters in step 3 include the network initial learning rate, the learning rate decay method, and the number of network iterations.

[0020] The loss function selects the mean square error, which is specifically expressed as follows:

[0021]

[0022] in, is the ith estimated radio map, V(i) is the true value; N train is the number of training samples.

[0023] In step 1, the specific method of using pixel values ​​to represent the height of buildings, furniture, and transmitters is as follows:

[0024] Use pixel value 0 to represent no object, and use pixel value 1 to represent the highest object h max , height h env The value of the pixel where the object is located is calculated according to the following formula:

[0025] v env =(h env +ε) / (h max +ε), where ε is a constant set to distinguish objects located on the ground.

[0026] A computer-readable storage medium, characterized in that: the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method are called.

[0027] A 3D indoor radio map estimation system based on 2D deep learning includes a processor module, a 3D radio map simulation module, and a memory module; wherein the processor module calls the method to estimate the 3D indoor radio map; the 3D radio map simulation module is used to simulate and obtain the data set required for model training and perform preprocessing; the memory module is used to classify and store the preprocessed data.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. With the proposed method, all deep learning-based 2D radio map estimation methods can estimate 3D radio maps, reducing computational overhead and memory requirements. The proposed method can also be used for 2D radio map estimation to input environmental height information.

[0030] 2. The present invention also proposes a 2D deep learning method for 3D indoor radio map estimation 2 Net (radioresidual network), the proposed method R 2 Net enhances feature extraction according to the path loss characteristics to improve estimation accuracy and achieve faster estimation speed.

[0031] 3. Proposed method R 2 Net also achieves strong generalization ability by randomly dropping layers. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of input and output of 3D indoor radio map estimation of the present invention.

[0033] Figure 2 The 3D indoor radio map estimation method R based on 2D deep learning of the present invention 2 Net structure diagram.

[0034] Figure 3 This is a schematic diagram of the cascaded residual module used in the present invention.

[0035] Figure 4 This is a schematic diagram of the path loss feature enhancement module proposed in the present invention.

[0036] Figure 5 This is a schematic diagram of the nearest neighbor upsampling module used in the present invention.

[0037] Figure 6 0.5m high indoor radio map estimated by the present invention and the comparison algorithm, where a is the true value obtained by simulation, b to e are the radio maps estimated by the comparison algorithm, b to e are all 2D deep learning methods originally used to estimate 2D radio maps, and the method proposed by the present invention is used to estimate 3D radio maps, and f is the 2D deep learning method R proposed by the present invention. 2 Net estimated radio map. DETAILED DESCRIPTION

[0038] The structure and working process of the present invention will be further described below in conjunction with the accompanying drawings.

[0039] The present invention proposes a 3D indoor radio map estimation method based on 2D deep learning, which estimates the 3D indoor radio map according to the building layout, furniture layout and transmitter position, and considers the influence of the height of indoor transmitters, receivers and obstacles on the radio map. Existing indoor radio map estimation based on deep learning mainly focuses on 2D radio maps, ignoring the influence of the height of transmitters, receivers and obstacles, while the height must be considered in indoor scenes. Although 3D indoor radio map estimation can be achieved by changing the 2D operation of the existing 2D radio map estimation method to 3D, the 3D deep learning algorithm has a large computational overhead. In order to reduce the computational overhead, the present invention proposes a 2D deep learning algorithm, a radio residual network (radioresidual network, i.e., R 2Net) to estimate the 3D radio map. Since the input and output of the 2D deep learning algorithm are both 2D images, the present invention proposes to use pixel values ​​to represent the heights of buildings, furniture and transmitters, and use different output channels to generate 3D radio maps of different heights. The present invention converts the 3D radio map estimation into a 2D deep learning task, reducing the computational overhead and memory requirements. 2 Net has strong generalization ability and enhances feature extraction according to indoor path loss characteristics, thus improving estimation accuracy and efficiency.

[0040] A 3D indoor radio map estimation method based on 2D deep learning includes the following steps:

[0041] Step 1: construct the input and output of several 3D radio map estimates obtained by pre-simulation into 2D pictures containing height information, use pixel values ​​to represent the heights of buildings, furniture and transmitters, and use different output channels to generate 3D radio maps of different heights, wherein the 2D pictures include a building layout map, a furniture layout map and a transmitter location map;

[0042] Step 2: construct a 2D deep learning method estimation model based on the characteristics of the 3D radio map, and divide the 2D images obtained in step 1 into a training set, a cross-validation set, and a test set;

[0043] Step 3: Set the training parameters and loss function of the 2D deep learning method estimation model;

[0044] Step 4: Use the training set to train the 2D deep learning method estimation model, use the cross-validation set for verification, and update the parameters of the 2D deep learning method estimation model in real time until the training is completed, and use the test set to evaluate the estimation accuracy of the model;

[0045] Step 5: Input the environmental parameters into the optimal 2D deep learning method estimation model obtained in step 4 to estimate the 3D indoor radio map.

[0046] Specific embodiments, such as Figures 1 to 6 As shown,

[0047] This embodiment uses pixel values ​​to represent the height of buildings, furniture, and transmitters, uses different output channels to generate 3D radio maps at different heights, and constructs the input and output of the 3D radio map estimation into 2D images, see the attached Figure 1 , converting 3D radio map estimation into a 2D deep learning task to reduce computational overhead and memory requirements. The green and white circles emphasize the transmitter locations.

[0048] See attached Figure 2 -Attached Figure 5 , the R2 Based on the encoder and decoder structure of 2D U-Net and according to the characteristics of indoor path loss, a random dropout layer is used to enhance the feature extraction of penetration loss, and a cascaded residual module is used to enhance the feature extraction of diffraction loss. At the same time, the nearest neighbor interpolation algorithm is used to reduce the information loss caused by random dropout layer and downsampling, so as to improve the estimation accuracy.

[0049] A 3D indoor radio map estimation method based on 2D deep learning proposed in this embodiment includes the following steps:

[0050] Step 1: construct the input and output of several 3D radio map estimates obtained by pre-simulation into 2D pictures containing height information, use pixel values ​​to represent the heights of buildings, furniture and transmitters, use different output channels to generate 3D radio maps of different heights, and reduce the output dimension from 3D to 2D, wherein the 2D pictures include a building layout map, a furniture layout map and a transmitter location map;

[0051] In this embodiment, the specific method of using pixel values ​​to represent the heights of buildings, furniture, and transmitters is as follows:

[0052] Use pixel value 0 to represent no object, and use pixel value 1 to represent the highest object h max , height h env The value of the pixel where the object is located is calculated according to the following formula:

[0053] v env =(h env +ε) / (h max +ε), where ε is a constant set to distinguish objects located on the ground, usually a small value. In this embodiment, ε takes the value of 0.1.

[0054] Step 2: Construct a 2D deep learning estimation model (i.e., R 2 Net neural network model), the 2D images obtained in step 1 are divided into a training set, a cross-validation set, and a test set;

[0055] The R 2 Net neural network model is as follows:

[0056] Overall network architecture: see attached Figure 2 ,This network uses 2D U-Net as the basic network architecture, and adds a cascaded residual module to the U-Net encoder to enhance the diffraction loss feature extraction. The structure of the cascaded residual module is shown in the attached Figure 3 ;

[0057] The proposed path loss feature enhancement module: R 2Net adopts the proposed path loss feature enhancement module including cascaded residual modules and random dropout layers between the encoder and decoder of U-Net to enhance the extraction of penetration loss and diffraction loss features. The structure of the proposed path loss feature enhancement module is shown in the attached Figure 4 , where the circled area shows that the proposed path loss feature enhancement module effectively extracts the path loss features at locations far away from the transmitter;

[0058] Nearest Neighbor Upsampling Module: R 2 Net uses the nearest neighbor upsampling module to reduce the information loss caused by random dropout layers and downsampling. The structure of the nearest neighbor upsampling module is shown in the attached Figure 5 .

[0059] The R 2 The specific construction method of Net neural network model is as follows:

[0060] Step 2.1, based on the U-Net encoder and decoder structure model, according to the characteristics of penetration loss, that is, the same building and the same furniture layout, once the transmitter position changes, the penetration loss will cause a large change in the 3D indoor radio map, requiring the model to have strong generalization ability. The proposed method R 2 Net enhances generalization by using random dropout layers in feature maps with lower resolution. Due to the sudden signal power attenuation and propagation direction change caused by diffraction loss, that is, indoor obstacles (such as furniture) have a nonlinear effect on path loss, the model needs to extract nonlinear features at the same time. The proposed method R 2 Net enhances nonlinear feature extraction through cascaded residual modules.

[0061] Step 2.2: Using the nearest neighbor interpolation algorithm, the decoder restores the features extracted by the encoder to the size of the original image, thereby realizing pixel-level radio map estimation; enhancing feature extraction through random dropout layers and downsampling layers will lead to information loss. In order to reduce the impact of information loss, the interpolation method commonly used by U-Net, namely deconvolution, cannot be used. Therefore, the proposed method R 2 Net uses the nearest neighbor interpolation algorithm.

[0062] Step 2.3: To improve the estimation accuracy, a skip connection is used to pass the features proposed by the encoder to the decoder.

[0063] Step 3, setting the training parameters and loss function of the 2D deep learning method estimation model; the training parameters include setting the network initial learning rate, learning rate decay method, and number of network iterations; the initial learning rate used here is 0.0001, the learning rate decay method is 0.00001 after 30 iterations, the number of network iterations is 50 times, and the loss function selects the mean square error;

[0064] The loss function is specifically expressed as follows:

[0065]

[0066] in, is the ith estimated radio map, V(i) is the true value; N train is the number of training samples.

[0067] Step 4: Use the training set to train the 2D deep learning method estimation model, and use the cross-validation set for verification to avoid overfitting, that is, select the model parameters with the smallest mean square error of the cross-validation set in 50 iterations, and update the parameters of the 2D deep learning method estimation model in real time until the training is completed, and use the test set to evaluate the estimation accuracy of the model;

[0068] Step 5: Input the environmental parameters into the optimal 2D deep learning method estimation model obtained in step 4 to estimate the 3D indoor radio map.

[0069] This embodiment uses simulation software to generate a 3D indoor radio map dataset 3DiRM3200, which includes 3200 3D radio maps, 200 building layouts, 200 furniture layouts, and 3200 randomly generated transmitter location maps. The experiment of 3D indoor radio map estimation uses the generated 3DiRM3200 dataset, and the 3200 samples are randomly divided into a training set of 2560 samples, a cross-validation set of 320 samples, and a test set of 320 samples.

[0070] The mentioned R 2 Net and the comparison methods RadioUNet, FadeNet, RadioTrans and PPNet are all trained on Nvidia Geforce RTX 3090 graphics cards. Among them, the comparison methods RadioUNet, FadeNet, RadioTrans and PPNet are deep learning methods for 2D radio map estimation. After adopting the method proposed in the present invention, they can be used to estimate 3D radio maps. Each model is trained 50 times. To avoid overfitting, the model parameters with the smallest MSE on the cross-validation set are selected. Since the simulation software for generating the data set 3DiRM3200 runs on the CPU, in order to facilitate the comparison of the inference speed of the simulation software and the deep learning method, all methods are tested on the Intel Core i5-11400F.

[0071] Experimental verification has shown that the 3D indoor radio map estimated by the invented method is more accurate than the existing radio map estimation method based on deep learning.

[0072] The samples randomly selected in the test set are on the radio map at a height of 0.5m. See the attached Figure 6 , the radio map estimated by the proposed R2Net is closest to the true value, especially for the locations far away from the transmitter in the radio map. This shows that the proposed R2Net enhances the indoor path loss feature extraction, thereby improving the estimation accuracy.

[0073] On the test set, the normalized mean square errors of the comparison methods RadioUNet, FadeNet, RadioTrans, and PPNet are 0.0292, 0.0583, 0.0714, and 0.1433, respectively. 2 Net can achieve a normalized mean square error of 0.0268, which is 8.22% smaller than RadioUNet and 81.30% smaller than PPNet. 2 Net can significantly improve the estimation accuracy of 3D indoor radio maps.

[0074] On an Intel Core i5-11400F, the simulation software estimates 0.04 3D indoor radio maps per second. The comparison methods RadioUNet, FadeNet, RadioTrans, and PPNet estimate 2, 1, 2, and 1 3D indoor radio maps per second, respectively. 2 Net can estimate three 3D indoor radio maps per second. 2 Net is faster than other comparison algorithms, and R 2 Net is two orders of magnitude faster than the main path model of the physical simulation method. In addition, the comparison methods RadioUNet, FadeNet, RadioTrans and PPNet contain 13M, 65M, 55M and 15M parameters respectively. 2 The number of parameters of RNet is 8M, which is much smaller than the comparison algorithm, 38.46% smaller than RadioUNet and 87.69% smaller than FadeNet. This shows that R 2 Net occupies less memory than the comparison algorithms. The comparison methods RadioUNet, FadeNet, RadioTrans and PPNet require 25.8G, 51.7G, 18.7G and 34.7G respectively. 2 The number of multiplication-addition-accumulation operations required by Net is 6.5G, which is less than the comparison algorithm, that is, the computational efficiency is higher.

[0075] A computer-readable storage medium, characterized in that: the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method are called.

[0076] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0077] A 3D indoor radio map estimation system based on 2D deep learning includes a processor module, a 3D radio map simulation module, and a memory module; wherein the processor module calls the method to estimate the 3D indoor radio map; the 3D radio map simulation module is used to simulate and obtain the data set required for model training and perform preprocessing; the memory module is used to classify and store the preprocessed data.

[0078] It should be understood that the present solution is not limited to the above-mentioned specific implementation methods, and the devices and structures not described in detail should be understood to be implemented in a common manner in the art; any technician familiar with the art can use the above-disclosed methods and technical contents to make many possible changes and modifications to the technical solution of the present solution without departing from the scope of the technical solution of the present solution, or modify it into an equivalent embodiment with equivalent changes, which does not affect the essential content of the present solution. Therefore, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present solution without departing from the content of the technical solution of the present solution still falls within the scope of protection of the technical solution of the present solution.

Claims

1. A 3D indoor radio map estimation method based on 2D deep learning, characterized by: The steps include: Step 1: construct the input and output of several 3D radio map estimates obtained by pre-simulation into 2D pictures containing height information, use pixel values ​​to represent the heights of buildings, furniture and transmitters, and use different output channels to generate 3D radio maps of different heights, wherein the 2D pictures include a building layout map, a furniture layout map and a transmitter location map; Step 2: construct a 2D deep learning method estimation model based on the characteristics of the 3D radio map, and divide the 2D images obtained in step 1 into a training set, a cross-validation set, and a test set; Step 3: Set the training parameters and loss function of the 2D deep learning method estimation model; Step 4: Use the training set to train the 2D deep learning method estimation model, use the cross-validation set for verification, and update the parameters of the 2D deep learning method estimation model in real time until the training is completed, and use the test set to evaluate the estimation accuracy of the model; Step 5: Input the environmental parameters into the optimal 2D deep learning method estimation model obtained in step 4 to estimate the 3D indoor radio map.

2. The 3D indoor radio map estimation method based on 2D deep learning according to claim 1, characterized in that: The specific architecture of the 2D deep learning method estimation model includes an encoder and a decoder, wherein the encoder includes several cascaded residual modules; the decoder includes several nearest neighbor upsampling modules; and a path loss feature enhancement module is connected between the encoder and the decoder.

3. The 3D indoor radio map estimation method based on 2D deep learning according to claim 2, characterized in that: The specific process of building a 2D deep learning method estimation model is as follows: Step 2.1, based on the U-Net encoder and decoder structure model, according to the characteristics of path loss, random dropout layer and cascade residual module are used to enhance feature extraction; Step 2.2: Using the nearest neighbor interpolation algorithm, the decoder restores the features extracted by the encoder to the size of the original image, thereby achieving pixel-level radio map estimation; Step 2.3: Use skip connection to pass the features proposed by the encoder to the decoder.

4. The 3D indoor radio map estimation method based on 2D deep learning according to claim 3, characterized in that: In step 2.1, for feature maps with relatively low resolution, a random dropout layer is used to enhance generalization; nonlinear feature extraction is enhanced through a cascaded residual module.

5. The 3D indoor radio map estimation method based on 2D deep learning according to claim 1, characterized in that: The training parameters in step 3 include the network initial learning rate, the learning rate decay method, and the number of network iterations.

6. The 3D indoor radio map estimation method based on 2D deep learning according to claim 5, characterized in that: The loss function selects the mean square error, which is specifically expressed as follows: in, is the ith estimated radio map, V(i) is the true value; N train is the number of training samples.

7. The 3D indoor radio map estimation method based on 2D deep learning according to claim 1, characterized in that: In step 1, the specific method of using pixel values ​​to represent the height of buildings, furniture, and transmitters is as follows: Use pixel value 0 to represent no object, and use pixel value 1 to represent the highest object h max , height h env The value of the pixel where the object is located is calculated according to the following formula: v env =(h env +ε) / (h max +ε), where ε is a constant set to distinguish objects located on the ground.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method described in any one of claims 1 to 7 are called.

9. A 3D indoor radio map estimation system based on 2D deep learning, characterized by: It includes a processor module, a 3D radio map simulation module, and a memory module; wherein the processor module calls the method described in any one of claims 1 to 7 to estimate a 3D indoor radio map; the 3D radio map simulation module is used to simulate and obtain the data set required for model training, and perform preprocessing; the memory module is used to classify and store the preprocessed data.

Images