Low-altitude three-dimensional radio map reconstruction method based on scenario guidance and sparse measurement
By combining deep neural networks with sparse measurement and environmental scene understanding, the problem of high accuracy in predicting radio signal coverage in 3D urban environments using traditional methods has been solved, and efficient 3D radio map reconstruction has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional two-dimensional coverage prediction methods are difficult to meet the needs of high-precision radio signal coverage prediction in complex three-dimensional urban environments. Existing technologies have high computational overhead or cannot capture fine spatial changes.
By employing a deep neural network architecture, combined with sparse measurement and environmental scene understanding, and utilizing a sparse data converter, a radio map data feature extraction module, a multi-scale convolutional pyramid, and a layer-by-layer reconstruction module, high-precision 3D radio map reconstruction is achieved.
It significantly reduces measurement overhead, improves reconstruction quality, generates 3D radio maps that better conform to the laws of physical propagation, and enhances the robustness and accuracy of reconstruction.
Smart Images
Figure CN122368360A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of wireless communication, specifically a method for reconstructing low-altitude three-dimensional radio maps based on scene guidance and sparse measurement. Background Technology
[0002] Radio maps are a key tool for planning and optimizing wireless communication networks, used to characterize spatial signal coverage distribution. As fifth-generation and future wireless networks expand into low-altitude applications such as UAV communication and urban air traffic, traditional two-dimensional coverage prediction methods struggle to meet the high-precision prediction requirements of complex three-dimensional urban environments. Existing technologies primarily rely on computationally intensive ray tracing simulations or simplified empirical propagation models. The former incurs enormous computational overhead, making real-time application difficult, while the latter fails to capture subtle spatial variations and complex multipath effects. Existing image super-resolution reconstruction techniques derive prior information from the signal space constituted by the image pixel values themselves, neglecting environmental information. Furthermore, existing wireless channel map construction techniques rely on two-dimensional diffuse images, failing to demonstrate continuity in the three-dimensional vertical direction. Summary of the Invention
[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement. Through a deep neural network architecture, it integrates environmental scene understanding with sparse measurement data to achieve high-precision, physically consistent 3D reference signal received power (RSRP) field reconstruction, significantly reducing measurement overhead and improving reconstruction quality. By fusing environmental building distribution with sparse RSRP measurements, utilizing feature pyramids to capture environmental patterns, and combining a hierarchical reconstruction strategy to maintain vertical dependence, high-precision 3D radio map reconstruction is achieved.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement. The method involves constructing and training a reconstruction network that includes a sparse data converter module, a radio map data feature extraction module, a multi-scale convolutional pyramid module, and a radio map layer-by-layer reconstruction module in the offline stage, and then performing real-time map reconstruction in the online stage using the trained reconstruction network.
[0006] The sparse data converter module transforms irregular sparse raw measurement data points into regular dense data so that the neural network can extract features and obtain a three-dimensional confidence map; the radio map data feature extraction module extracts RSRP semantic features based on the three-dimensional projected RSRP map; the multi-scale convolutional pyramid module extracts features from the environmental map to obtain a multi-scale environmental feature tensor; the radio map layer-by-layer reconstruction module reconstructs the two-dimensional RSRP map layer by layer and accumulates them to obtain a three-dimensional RSRP map.
[0007] The method specifically includes:
[0008] Step 1: Obtain a two-dimensional building distribution map of the target area through low-altitude field surveys using drones or an open map platform;
[0009] Step 2: Obtain a 3D RSRP map of the target area through simulation based on the ray tracing algorithm, and randomly sample from it to obtain sparse RSRP data points;
[0010] Step 3: Map the sparse RSRP data points obtained in Step 2 to three dimensions using the sparse data converter module, convert them into a structured mesh representation, and obtain a three-dimensional projection map and a three-dimensional confidence map of the target area.
[0011] Step 4: Extract features from the 3D projected map obtained in Step 3 using the radio map data feature extraction module to obtain RSRP semantic features;
[0012] Step 5: Extract features from the building distribution map obtained in Step 1 using the multi-scale convolutional pyramid module to obtain environmental features;
[0013] Step 6: Using the radio map layer-by-layer reconstruction module, the two-dimensional RSRP map is reconstructed layer by layer based on the three-dimensional sparse map and three-dimensional confidence map obtained in Step 3, the semantic features obtained in Step 4, and the environmental features obtained in Step 5. The two-dimensional RSRP map is then accumulated to obtain the three-dimensional RSRP map.
[0014] Step 7: Calculate the target loss based on the obtained simulated 3D RSRP map and the reconstructed 3D RSRP map obtained in Step 6;
[0015] Step 8: Update the sparse data converter module in Step 3, the radio map data feature extraction module in Step 4, the multi-scale convolutional pyramid module in Step 5, and the radio map layer-by-layer reconstruction module in Step 6 based on the target loss obtained in Step 7, to obtain a complete three-dimensional radio map reconstruction model.
[0016] Step 9, Online Phase: Based on the trained model, reconstruct a 3D RSRP map of the target area using the real-time input building distribution map of the target area and the sparse RSRP data points obtained from actual UAV measurements.
[0017] Technical effect
[0018] This invention introduces a detail-enhanced residual block containing three branch paths at different scales into the feature pyramid model, enabling the extraction and fusion of multi-scale environmental features in building distribution maps. Simultaneously, by introducing an inter-layer communication mechanism into the layer-by-layer reconstruction model, information from adjacent layers is fused during the reconstruction of each height layer, ensuring the continuity and smooth transition of the signal field in the vertical dimension. Compared to existing technologies, this invention, through the multi-scale perception strategy in the multi-scale convolutional pyramid module, allows the reconstruction network to analyze and fuse multi-scale features from details to the global scale in environmental scene images. This structure enables the network to analyze the geometric topology in the building distribution map, generating a signal field that better conforms to the laws of physical propagation, thereby achieving a more physically consistent 3D radio map reconstruction. Furthermore, through the inter-layer communication mechanism in the layer-by-layer reconstruction model, in terms of vertical continuity, the reconstructed RSRP field achieves a smooth transition in the vertical dimension, ensuring the vertical coherence of the signal field and achieving effective modeling of the vertical signal propagation structure. In terms of robustness, when measurement data is lacking at a specific height layer, this mechanism allows the network to use features from adjacent layers to fill in the gaps, effectively improving the robustness and accuracy of the reconstruction. Attached Figure Description
[0019] Figure 1 This is a flowchart of the present invention;
[0020] Figure 2 This is a schematic diagram of the invention;
[0021] Figure 3 This is a schematic diagram of the sparse data converter module;
[0022] Figure 4 This is a schematic diagram of the structure of the radio map data feature extraction module;
[0023] Figure 5 This is a schematic diagram of the structure of a multi-scale convolutional pyramid module;
[0024] Figure 6 A schematic diagram of the structure of the radio map layer-by-layer reconstruction module;
[0025] Figure 7 This is a comparison of reconstruction quality at different heights in a 3D building occlusion scenario based on simulation data. Detailed Implementation
[0026] like Figure 1 and Figure 2 As shown in this embodiment, a low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement is proposed. The method includes: a sparse data converter module, a radio map data feature extraction module, a multi-scale convolutional pyramid module, and a radio map layer-by-layer reconstruction module. Specifically, the sparse data converter module converts irregular sparse original measurement data points into regularized dense data so that a neural network can extract features and obtain a 3D confidence map; the radio map data feature extraction module reconstructs the map layer by layer based on the dimension... 3D projection RSRP map The extracted dimension is RSRP semantic features The multi-scale convolutional pyramid module extracts features from the environment map E to obtain a multi-scale environment feature tensor. The radio map layer-by-layer reconstruction module reconstructs the two-dimensional RSRP map layer by layer, and accumulates the results to obtain the three-dimensional RSRP map. Where: C is the number of feature channels, environment map L represents the number of horizontal grid points in the target area, W represents the number of vertical grid points in the target area, and B represents the batch size.
[0027] like Figure 3 As shown, the sparse data converter module includes: a coordinate mapping and quantization unit, a trilinear weighted interpolation calculation unit, and a sparse feature accumulation unit. Specifically: the coordinate mapping and quantization unit maps normalized three-dimensional floating-point coordinates to a discrete grid index space, providing basic calculation parameters for trilinear interpolation; the trilinear weighted interpolation calculation unit calculates the contribution weight of each measurement point to its eight neighboring grid points based on the trilinear interpolation principle; and the sparse feature accumulation unit distributes the RSRP value of the measurement point to the eight neighboring grid vertices according to the calculated weights, and uses parallel scattering operations to achieve optimized accumulation.
[0028] The sparse RSRP data points are specifically... This represents the value of the 3D coordinates of the i-th sparse data point after normalization to [0, 1]. Indicates the location The RSRP value at point N represents the total number of data points obtained through random sampling.
[0029] The coordinate mapping and quantization unit receives the input normalized three-dimensional coordinate tensor, which has the following shape: The three dimensions correspond to the x, y, and h coordinates, respectively. The normalized three-dimensional coordinates are... The indices mapped to the target mesh are multiplied by, for x, y, and h, respectively. , , Rounding down, we get the integer and decimal parts. The integer part is used as the bottom-left index of the grid cell where the measurement point is located, and the decimal part is used as the decimal part. This refers to the relative position of the measurement point within the grid cell.
[0030] The contribution weights mentioned are specifically as follows: ,in: for , indicating that for any grid vertex surrounding a measurement point Its weight equal The product of the weighting factors in the three directions, where the weighting factor is [value] in each direction. According to the index The value is determined, and a shape is ultimately obtained. The weight tensor.
[0031] The aforementioned allocation preferably avoids the use of inefficient loops, flattens the batch index, coordinate index, weights, and values all to one dimension, and then flattens the four-dimensional coordinates... After encoding the data into a unique one-dimensional memory address index, the scattering accumulation operator in PyTorch is used to process all sampled points. RSRP value Multiply by its corresponding weight Accumulated into a 3D projection RSRP map The corresponding positions, and all sampling points weight Accumulated into the 3D confidence map The corresponding positions are used to obtain the three-dimensional projected RSRP map. and 3D confidence map Specifically: Where: H is the number of height grids in the target area, and the confidence map A higher pixel value indicates a denser concentration of real sampling points and more reliable measurement data in the vicinity of that location. This value is subsequently used for weighting in the loss function, guiding the network to focus on optimizing regions with real data.
[0032] like Figure 4As shown, the radio map data feature extraction module includes: an input convolutional unit, two detail enhancement residual block units, a fine-grained attention unit, and an output convolutional unit. Specifically: the input convolutional unit expands the projected feature maps of multiple height layers from the channel dimension, providing sufficient feature representation capabilities for subsequent deep processing; the detail enhancement residual block units extract deep patterns from preliminary semantic features while preserving spatial details; the fine-grained attention unit dynamically adjusts the importance of different spatial locations and channels in the feature maps, enhancing key features and suppressing irrelevant features; and the output convolutional unit refines and adjusts the deepened semantic features processed by the fine-grained attention unit, providing a suitable input format for the subsequent hierarchical reconstruction module.
[0033] The main body of the input convolutional unit consists of a two-dimensional convolutional layer, a normalization layer, and an activation layer, wherein: the kernel size of the two-dimensional convolutional layer is [missing information]. The step size is 1, the padding is 1, the number of input channels is H (corresponding to H height layers), and the number of output channels is C. The normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce non-linear features. The input dimension of this unit is... 3D projection RSRP map The output dimension is Preliminary semantic features.
[0034] The detail enhancement residual block unit employs three parallel paths for feature extraction: the local perception path consists of a convolutional kernel with a size of [missing information]. It consists of two-dimensional convolutional layers padded with 1s, responsible for extracting regular local features; the wide-area perception path consists of convolutional kernels with a size of... It consists of two-dimensional convolutional layers with an inflation rate of 2 and a padding of 2, utilizing dilated convolution to expand the receptive field and capture a wide range of signal fading trends; the channel fusion path consists of convolutional kernels with a size of... It consists of two-dimensional convolutional layers responsible for cross-channel information exchange. Each path is followed by a normalization layer and an activation layer. After the input features pass through the above three paths simultaneously, the output results are added element-wise to achieve multi-scale feature fusion.
[0035] The fine-grained attention unit includes a spatial attention path and a channel attention path, wherein the spatial attention path is used to identify pixel locations in an image that contain valid measurement information or key geometric features, and its structure is as follows: Dimension The input features are first processed by a convolution kernel with a size of The two-dimensional convolutional layers and ReLU activation layers reduce the number of channels by 1 / 4, that is... Then pass through a convolution kernel with a size of Two-dimensional convolutional layers and ReLU activation layers capture spatial information, and finally, a convolutional kernel of size [size missing] is used to [process the process]. Two-dimensional convolutional layers reduce the number of channels to The output is compressed to between 0 and 1 using a Sigmoid activation layer. The channel attention path identifies the feature channels most important to the current task, with the following structure: First, the dimension... The input features are subjected to global average pooling, compressing the spatial dimensions (W and L) of each channel to 1x1, i.e. Secondly, through a convolution kernel with a size of The two-dimensional convolutional layers and ReLU activation layers reduce the number of channels by 1 / 4, that is... Then pass through a convolution kernel with a size of The two-dimensional convolutional layer increases the number of channels by four times, that is, it returns to... The output is compressed to between 0 and 1 using a Sigmoid activation layer. After passing through two parallel paths, the input features are multiplied element-wise with the output features of the two parallel paths to obtain the dimension. This deepens semantic features, thereby enhancing regional features with high confidence measurements and suppressing invalid responses in noisy or blank areas.
[0036] The output convolutional unit comprises a two-dimensional convolutional layer, a normalization layer, and an activation layer, wherein the kernel size of the two-dimensional convolutional layer is [missing information]. The number of channels remains at C; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce non-linear features. The output dimension of this unit is... RSRP semantic features .
[0037] like Figure 5 As shown, the multi-scale convolutional pyramid module includes: three encoder units, two decoder units, and an output convolutional unit. The encoder units extract low-level visual features from the original environment map, increasing the channel dimension while maintaining spatial resolution. The decoder units upsample the deep features back to a large scale and fuse them with the corresponding layer features of the encoder. The output convolutional unit fuses the stitched features and adjusts them to the target number of channels.
[0038] The encoder unit includes: a two-dimensional convolutional layer, a normalization layer, an activation layer, and detail enhancement residual block units, wherein: the kernel size of the two-dimensional convolutional layer is [missing information]. The stride of the two-dimensional convolutional layer in the first-level encoder unit is 1, and the stride of the two-dimensional convolutional layer in the second and third-level encoder units is 2; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce nonlinear features; the structure of the detail enhancement residual block unit is as described in
[0042] . The dimension is... The environment map E is taken as input, and after passing through the first-stage encoder unit, the output dimension is... After passing through the second-stage encoder unit, the output dimension is After passing through the third-level encoder unit, the output dimension is .
[0039] The decoder unit includes: a 2D transposed convolutional layer, a normalization layer, an activation layer, and skip connections, wherein the kernel size of the 2D transposed convolutional layer is [missing information]. The step size is 2; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce non-linear features; skip connections fuse the encoder features to compensate for information loss caused by downsampling. The dimension is... The encoder output is used as input, and in the first-level decoder unit, the output dimension of the two-dimensional transposed convolutional layer is... After a skip connection with the second-level encoder, the dimension is In the second-level decoder unit, the output dimension of the two-dimensional transposed convolutional layer is... After a skip connection with the first-level encoder, the dimension is .
[0040] The output convolutional unit is a convolutional kernel with a size of A two-dimensional convolutional layer is used, where this unit receives features from the output of the second-stage decoder. Due to skip connections doubling the number of channels, this unit integrates information between channels, generating a dimension of... Final scene features .
[0041] like Figure 6 As shown, the radio map layer-by-layer reconstruction module includes: a fusion feature extraction unit, a local enhancement branch unit, a global enhancement branch unit, a feature integration and output unit, and an inter-layer information transfer unit. Specifically: the fusion feature extraction unit fuses features from multiple input sources and extracts abstract features; the local enhancement branch unit extracts local fine features, preserving spatial details and high-frequency information; the global enhancement branch unit extracts global semantic features, capturing large-scale patterns and environmental context, and modeling the macroscopic trends of radio propagation; the feature integration and output unit integrates the information output from the two branches and outputs the final single-layer RSRP map through progressive convolution; and the inter-layer information transfer unit, after generating the RSRP reconstruction map at the current layer height, passes the features to the reconstructor of the next layer, maintaining inter-layer continuity and consistency.
[0042] The fusion feature extraction unit includes: a data concatenation layer, a two-dimensional convolutional layer, a normalization layer, an activation layer, and a detail enhancement residual block unit, wherein: the data concatenation layer concatenates the six parts of the input data: environmental features semantic features Projected map of current height h Confidence map of current height h The height of the next floor Projected map and the height of the next floor Projected map The concatenated data dimensions are Then, through a convolution kernel size of It consists of a 2D convolutional layer, a normalization layer, an activation layer, and two cascaded detail-enhancing residual block units, with an output dimension of... The fusion characteristics.
[0043] The local enhancement branch unit includes: three two-dimensional convolutional layers and three activation layers, wherein: the first two-dimensional convolutional layer uses a convolutional kernel size of... Two-dimensional convolution and ReLU activation function are used for channel compression, with an output dimension of Features; the second two-dimensional convolutional layer uses a kernel size of The first two-dimensional convolutional layer uses ReLU activation for detail extraction, while maintaining the same output dimension; the third two-dimensional convolutional layer uses a kernel size of [missing value]. Two-dimensional convolution and ReLU activation function are used for feature refinement, with an output dimension of . Its characteristics.
[0044] The global enhancement branch unit includes: a pooling layer, a semantic encoder, and an upsampling layer, wherein: the pooling layer compresses the input features and the output dimension is... The semantic encoder extracts features by discarding high-frequency details, forcing its convolutional kernels to only see the overall global picture of the entire map. The semantic encoder contains two 2D convolutional layers and two activation layers, compressing the input features in terms of channel count and extracting features, with an output dimension of... The features; upsampling stretches the resolution of the input features back to the original size, and the output dimension is Its characteristics.
[0045] The feature integration and output unit concatenates the outputs of the two branches, with an output dimension of... After defining the features, three convolutional layers with kernel sizes of [size missing] are then applied. The two-dimensional convolution is used to fuse features and finally map them into a single-channel RSRP reconstructed map. The output dimension of the first two-dimensional convolutional layer and the ReLU activation function is... The characteristics of the second two-dimensional convolutional layer and the ReLU activation function are that the output dimension is The characteristics of the third two-dimensional convolutional layer and the ReLU activation function are that the output dimension is Single-layer RSRP reconstruction map .
[0046] The inter-layer information transmission unit comprises a total of There are 10 sub-units (because the highest-height layer doesn't need to be passed to the next layer), and each sub-unit consists of a convolutional kernel of size 1. The two-dimensional convolutional layer, normalization layer, activation layer, and one convolutional kernel are of size . In a two-dimensional convolutional layer, after processing the h-th layer, the sub-units input the projection features of the current layer. , dimension Update features with unchanged output dimensions And add it to the environmental features of the next layer. Input in progress.
[0047] like Figure 1 As shown, this embodiment relates to a low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement of the above system, including the following steps:
[0048] Step 1: Obtain a two-dimensional environmental map of the target area through low-altitude drone surveys or an open map platform. Specifically, this can be represented as... If point If the point is in free space, the value is 0; if the point is in free space, the value is 0. If the location is a building or obstacle, the value is the ratio of the building's height at that point to the highest point of the scene. In this embodiment, a building distribution map E of 30,000 urban environment scenes is obtained from the OpenStreetMap open-source map database. Each scene is 512 meters × 512 meters in size, 120 meters in height, and each pixel has a resolution of 2 meters × 2 meters. Therefore, the input dimension of the environment map is... In specific applications, the scene size can be adjusted according to the actual environment size.
[0049] Step 2: Obtain a 3D RSRP map of the target area through simulation using a ray tracing algorithm, and randomly sample from it to obtain a sparse RSRP data point set. The ray tracing algorithm is specifically implemented using, but is not limited to, the Sionna ray tracing open-source simulation platform developed by NVIDIA. The 3D RSRP map obtained from the simulation is as follows: Obtained by random sampling In this embodiment, NVIDIA's Sionna ray tracing open-source simulation platform was used to generate a simulated 3D RSRP map R with 30,000 scenes and a resolution of 12×256×256. Among them, 26,000 urban environment scenes were used as the training set, 2,800 as the test set, and 1,200 as the validation set. Sparse RSRP data points were obtained by sampling from it using three sampling numbers: 50,000 data points, 40,000 data points, and 30,000 data points.
[0050] Step 3: The sparse RSRP data points obtained in Step 2 are mapped to a 3D model using the sparse data converter module, converting them into a structured mesh representation to obtain a 3D projection map and a 3D confidence map of the target area. In this embodiment, both the 3D projection map and the 3D confidence map have a dimension of 1. .
[0051] Step 4: Extract features from the 3D projected map obtained in Step 3 using the radio map data feature extraction module to obtain RSRP semantic features. In this embodiment, the RSRP semantic feature dimension is [dimension not specified]. C is set to 64, which represents the number of channels.
[0052] Step 5: Extract features from the building distribution map obtained in Step 1 using a multi-scale convolutional pyramid module to obtain environmental features; in this embodiment, the output dimension of the environmental features is... .
[0053] Step 6: Using the radio map layer-by-layer reconstruction module, a two-dimensional RSRP map is reconstructed layer by layer based on the three-dimensional sparse map and three-dimensional confidence map obtained in Step 3, the semantic features obtained in Step 4, and the environmental features obtained in Step 5. The reconstructed three-dimensional RSRP map is then obtained by summing these features. In this embodiment, the output dimension of the three-dimensional RSRP reconstructed map is [missing information]. .
[0054] Step 7: Based on the simulated 3D RSRP map obtained in Step 2 and the 3D reconstructed RSRP map obtained in Step 6, apply the loss function... To minimize the simulation of 3D RSRP maps Reconstructing a 3D RSRP Map with a Network The differences between the modules are used to train the sparse data converter module in step 3, the radio map data feature extraction module in step 4, the multi-scale convolutional pyramid module in step 5, and the radio map layer-by-layer reconstruction module in step 6, specifically as follows: Among them: reconstruction losses Marginal loss ,in: The smoothing factor is used to avoid the zero-weight problem in the unmeasured region, and s represents two different scales. Time represents the original resolution. The time represents the resolution after downsampling; and Sobel operators at different scales are used to preserve edge information; and Represents a simulated 3D RSRP map and reconstructing 3D RSRP maps Slices at the same height level.
[0055] Step 8: In the online phase, based on the trained model, a 3D RSRP map of the target area is reconstructed from the real-time input building distribution map of the target area and the sparse RSRP data points obtained from the actual measurement by the UAV.
[0056] Network reconstruction effect as follows Figure 7 As shown in the figure, the RSRP reconstruction quality is compared at four representative height levels. The yellow areas represent buildings, and the blue areas surrounding the buildings represent weak signal regions obscured by the buildings. The figure demonstrates that the neural network reconstruction results can effectively preserve the complexity of the spatial signal structure and accurately capture the four-quadrant propagation patterns radiating outward from the central transmitter. The results indicate that the neural network method can well meet the accuracy requirements of practical wireless network planning in modeling signal propagation in complex urban environments.
[0057] As shown in Table 1, the evaluation index calculations demonstrate that this invention significantly improves performance across all evaluation metrics. With 50,000 sampling points, this invention achieves a mean square error of 0.012, a root mean square error of 0.111, a mean absolute error of 0.030, and a correlation coefficient of 0.954, achieving a coverage quality of 97%. Sampling density analysis shows that 40,000 sampling points achieve the best balance, reducing the number of measurements required by 20% while achieving 95% of the best performance. However, below this threshold, performance degrades significantly: a configuration with 30,000 sampling points shows a 277% increase in MSE and a decrease in correlation from 0.953 to 0.840, establishing a critical sampling threshold for practical deployment.
[0058] Table 1 Comparison of Evaluation Indicators for 3D Reconstructed Maps
[0059] Compared with existing technologies, this invention combines environmental information from wireless communication systems with deep learning technology to reconstruct a more accurate three-dimensional radio map, significantly improving the accuracy and efficiency of low-altitude network planning.
[0060] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for reconstructing low-altitude 3D radio maps based on scene guidance and sparse measurement, characterized in that, By constructing and training a reconstruction network that includes a sparse data converter module, a radio map data feature extraction module, a multi-scale convolutional pyramid module, and a radio map layer-by-layer reconstruction module in the offline stage, real-time map reconstruction is performed in the online stage using the trained reconstruction network. The sparse data converter module transforms irregular sparse raw measurement data points into regular dense data so that the neural network can extract features and obtain a three-dimensional confidence map; the radio map data feature extraction module extracts RSRP semantic features based on the three-dimensional projected RSRP map; the multi-scale convolutional pyramid module extracts features from the environmental map to obtain a multi-scale environmental feature tensor; the radio map layer-by-layer reconstruction module reconstructs the two-dimensional RSRP map layer by layer and accumulates them to obtain a three-dimensional RSRP map.
2. The low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement according to claim 1, characterized in that, The sparse data converter module includes: a coordinate mapping and quantization unit, a trilinear weighted interpolation calculation unit, and a sparse feature accumulation unit. Specifically: the coordinate mapping and quantization unit maps normalized 3D floating-point coordinates to a discrete grid index space, providing basic calculation parameters for trilinear interpolation; the trilinear weighted interpolation calculation unit calculates the contribution weight of each measurement point to its eight neighboring grid points based on the trilinear interpolation principle; and the sparse feature accumulation unit distributes the RSRP value of the measurement point to the eight neighboring grid vertices according to the calculated weights, and uses parallel scattering operations to achieve optimized accumulation. The sparse RSRP data points are specifically... This represents the value of the 3D coordinates of the i-th sparse data point after normalization to [0, 1]. Indicates the location The RSRP value at point N represents the total number of data points obtained through random sampling.
3. The low-altitude 3D radio map reconstruction method based on scene guidance and sparse measurement according to claim 2, characterized in that, The coordinate mapping and quantization unit receives the input normalized three-dimensional coordinate tensor, which has the following shape: The three dimensions correspond to the x, y, and h coordinates, respectively. The normalized three-dimensional coordinates are... The indices mapped to the target mesh are multiplied by, for x, y, and h, respectively. , , Rounding down, we get the integer and decimal parts. The integer part is used as the bottom-left index of the grid cell where the measurement point is located, and the decimal part is used as the decimal part. As the relative position of the measurement point within the grid cell; The contribution weights mentioned are specifically: ,in: for , indicating that for any grid vertex surrounding a measurement point Its weight equal The product of the weighting factors in the three directions, where the weighting factor is [value] in each direction. According to the index The value is determined, and a shape is ultimately obtained. The weight tensor; To avoid using inefficient loops, the allocation process flattens the batch index, coordinate index, weights, and values to one dimension, and then flattens the four-dimensional coordinates. After encoding the data into a unique one-dimensional memory address index, the scattering accumulation operator in PyTorch is used to process all sampled points. RSRP value Multiply by its corresponding weight Accumulated into a 3D projection RSRP map The corresponding positions, and all sampling points weight Accumulated into the 3D confidence map The corresponding positions are used to obtain the three-dimensional projected RSRP map. and 3D confidence map Specifically: Where: H is the number of height grids in the target area, and the confidence map A higher pixel value indicates a denser concentration of real sampling points and more reliable measurement data in the vicinity of that location. This value is subsequently used for weighting in the loss function, guiding the network to focus on optimizing regions with real data.
4. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 1, characterized in that, The radio map data feature extraction module includes: an input convolutional unit, two detail enhancement residual block units, a fine-grained attention unit, and an output convolutional unit. Specifically: the input convolutional unit expands the projected feature maps of multiple height layers from the channel dimension, providing sufficient feature representation capabilities for subsequent deep processing; the detail enhancement residual block units extract deep patterns from preliminary semantic features while preserving spatial details; the fine-grained attention unit dynamically adjusts the importance of different spatial locations and channels in the feature maps, enhancing key features and suppressing irrelevant features; and the output convolutional unit refines and adjusts the deepened semantic features processed by the fine-grained attention unit, providing a suitable input format for the subsequent hierarchical reconstruction module.
5. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 4, characterized in that, The main body of the input convolutional unit consists of a two-dimensional convolutional layer, a normalization layer, and an activation layer, wherein: the kernel size of the two-dimensional convolutional layer is [missing information]. The step size is 1, the padding is 1, the number of input channels is H (corresponding to H height layers), and the number of output channels is C; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce non-linear features, and the input dimension of this unit is... 3D projection RSRP map The output dimension is Preliminary semantic features; The detail enhancement residual block unit employs three parallel paths for feature extraction: the local perception path consists of a convolutional kernel with a size of [missing information]. It consists of two-dimensional convolutional layers padded with 1s, responsible for extracting regular local features; the wide-area perception path consists of convolutional kernels with a size of... It consists of two-dimensional convolutional layers with an inflation rate of 2 and a padding of 2, utilizing dilated convolution to expand the receptive field and capture a wide range of signal fading trends; the channel fusion path consists of convolutional kernels with a size of... It consists of two-dimensional convolutional layers, which are responsible for cross-channel information interaction. Each path is followed by a normalization layer and an activation layer. After the input features pass through the above three paths at the same time, the output results are added element by element to achieve the fusion of multi-scale features. The fine-grained attention unit includes a spatial attention path and a channel attention path, wherein the spatial attention path is used to identify pixel locations in an image that contain valid measurement information or key geometric features, and its structure is as follows: Dimension The input features are first processed by a convolution kernel with a size of The two-dimensional convolutional layers and ReLU activation layers reduce the number of channels by 1 / 4, that is... Then pass through a convolution kernel with a size of Two-dimensional convolutional layers and ReLU activation layers capture spatial information, and finally, a convolutional kernel of size [size missing] is used to [process the process]. Two-dimensional convolutional layers reduce the number of channels to The output is compressed to between 0 and 1 using a Sigmoid activation layer. Channel attention paths are used to identify the feature channels most important to the current task, with the following structure: First, the dimension... The input features are subjected to global average pooling, compressing the spatial dimensions (W and L) of each channel to 1x1, i.e. Secondly, through a convolution kernel with a size of The two-dimensional convolutional layers and ReLU activation layers reduce the number of channels by 1 / 4, that is... Then pass through a convolution kernel with a size of The two-dimensional convolutional layer increases the number of channels by four times, that is, returns to... The output is compressed to between 0 and 1 using a Sigmoid activation layer. After passing through two parallel paths, the input features are multiplied element-wise with the output features of the two parallel paths to obtain the dimension. This deepens semantic features, thereby enhancing regional features with high confidence measurements and suppressing invalid responses in noisy or blank regions; The output convolutional unit comprises a two-dimensional convolutional layer, a normalization layer, and an activation layer, wherein the kernel size of the two-dimensional convolutional layer is [missing information]. The number of channels remains at C; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer, introducing non-linear features, and the output dimension of this unit is... RSRP semantic features .
6. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 1, characterized in that, The multi-scale convolutional pyramid module includes three encoder units, two decoder units, and an output convolutional unit. The encoder units extract low-level visual features from the original environment map, increasing the channel dimension while maintaining spatial resolution. The decoder units upsample the deep features back to a large scale and fuse them with the corresponding layer features from the encoder. The output convolutional unit fuses the stitched features and adjusts them to the target number of channels.
7. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 6, characterized in that, The encoder unit includes: a two-dimensional convolutional layer, a normalization layer, an activation layer, and detail enhancement residual block units, wherein: the kernel size of the two-dimensional convolutional layer is [missing information]. The stride of the two-dimensional convolutional layer in the first-level encoder unit is 1, and the stride of the two-dimensional convolutional layer in the second and third-level encoder units is 2; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce nonlinear features; the detail enhancement residual block unit structure is as described in [0042], with a dimension of The environment map E is taken as input, and after passing through the first-stage encoder unit, the output dimension is... After passing through the second-stage encoder unit, the output dimension is After passing through the third-level encoder unit, the output dimension is ; The decoder unit includes: a 2D transposed convolutional layer, a normalization layer, an activation layer, and skip connections, wherein the kernel size of the 2D transposed convolutional layer is [missing information]. The step size is 2; the normalization layer performs two-dimensional batch normalization to stabilize the feature distribution; the activation layer is a ReLU layer to introduce non-linear features; skip connections fuse the encoder features to compensate for information loss caused by downsampling, with a dimension of [missing information]. The encoder output is used as input, and in the first-level decoder unit, the output dimension of the two-dimensional transposed convolutional layer is... After a skip connection with the second-level encoder, the dimension is In the second-level decoder unit, the output dimension of the two-dimensional transposed convolutional layer is... After a skip connection with the first-level encoder, the dimension is ; The output convolutional unit is a convolutional kernel with a size of The two-dimensional convolutional layer receives features from the output of the second-stage decoder. Due to skip connections doubling the number of channels, this unit integrates information between channels, generating a dimension of... Final scene features .
8. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 1, characterized in that, The radio map layer-by-layer reconstruction module includes: a fusion feature extraction unit, a local enhancement branch unit, a global enhancement branch unit, a feature integration and output unit, and an inter-layer information transfer unit. Specifically: the fusion feature extraction unit fuses features from multiple input sources and extracts abstract features; the local enhancement branch unit extracts fine local features, preserving spatial details and high-frequency information; the global enhancement branch unit extracts global semantic features, capturing large-scale patterns and environmental context, and modeling the macroscopic trends of radio propagation; the feature integration and output unit integrates the information output from the two branches and outputs the final single-layer RSRP map through progressive convolution; and the inter-layer information transfer unit, after generating the RSRP reconstruction map at the current layer height, passes the features to the reconstructor of the next layer, maintaining inter-layer continuity and consistency.
9. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to claim 8, characterized in that, The fusion feature extraction unit includes: a data concatenation layer, a two-dimensional convolutional layer, a normalization layer, an activation layer, and a detail enhancement residual block unit, wherein: the data concatenation layer concatenates the six parts of the input data: environmental features semantic features Projected map of current height h Confidence map of current height h The height of the next floor Projected map and the height of the next floor Projected map The concatenated data dimensions are Then, through a convolution kernel size of It consists of a 2D convolutional layer, a normalization layer, an activation layer, and two cascaded detail-enhancing residual block units, with an output dimension of... The fusion characteristics; The local enhancement branch unit includes: three two-dimensional convolutional layers and three activation layers, wherein: the first two-dimensional convolutional layer uses a convolutional kernel size of... Two-dimensional convolution and ReLU activation function are used for channel compression, with an output dimension of Features; the second two-dimensional convolutional layer uses a kernel size of The first two-dimensional convolutional layer uses ReLU activation for detail extraction, while maintaining the same output dimension; the third two-dimensional convolutional layer uses a kernel size of [missing value]. Two-dimensional convolution and ReLU activation function are used for feature refinement, with an output dimension of . Features; The global enhancement branch unit includes: a pooling layer, a semantic encoder, and an upsampling layer, wherein: the pooling layer compresses the input features and the output dimension is... The semantic encoder extracts features by discarding high-frequency details, forcing its convolutional kernels to only see the overall global picture of the entire map. The semantic encoder contains two 2D convolutional layers and two activation layers, compressing the input features in terms of channel count and extracting features, with an output dimension of... The features; upsampling stretches the resolution of the input features back to the original size, and the output dimension is Features; The feature integration and output unit concatenates the outputs of the two branches, with an output dimension of... After defining the features, three convolutional layers with kernel sizes of [size missing] are then applied. The two-dimensional convolution is used to fuse features and finally map them into a single-channel RSRP reconstructed map. The output dimension of the first two-dimensional convolutional layer and the ReLU activation function is... The characteristics of the second two-dimensional convolutional layer and the ReLU activation function are that the output dimension is The characteristics of the third two-dimensional convolutional layer and the ReLU activation function are that the output dimension is Single-layer RSRP reconstruction map ; The inter-layer information transmission unit comprises a total of There are 10 sub-units (because the highest-height layer doesn't need to be passed to the next layer), and each sub-unit consists of a convolutional kernel of size 1. The two-dimensional convolutional layer, normalization layer, activation layer, and one convolutional kernel are of size . In a two-dimensional convolutional layer, after processing the h-th layer, the sub-units input the projection features of the current layer. , dimension Update features with unchanged output dimensions And add it to the environmental features of the next layer. Input in progress.
10. The low-altitude three-dimensional radio map reconstruction method based on scene guidance and sparse measurement according to any one of claims 1-9, characterized in that, specifically include: Step 1: Obtain a two-dimensional building distribution map of the target area through low-altitude field surveys using drones or an open map platform; Step 2: Obtain a 3D RSRP map of the target area through simulation based on the ray tracing algorithm, and randomly sample from it to obtain sparse RSRP data points; Step 3: Map the sparse RSRP data points obtained in Step 2 to three dimensions using the sparse data converter module, convert them into a structured mesh representation, and obtain a three-dimensional projection map and a three-dimensional confidence map of the target area. Step 4: Extract features from the 3D projected map obtained in Step 3 using the radio map data feature extraction module to obtain RSRP semantic features; Step 5: Extract features from the building distribution map obtained in Step 1 using the multi-scale convolutional pyramid module to obtain environmental features; Step 6: Using the radio map layer-by-layer reconstruction module, the two-dimensional RSRP map is reconstructed layer by layer based on the three-dimensional sparse map and three-dimensional confidence map obtained in Step 3, the semantic features obtained in Step 4, and the environmental features obtained in Step 5. The two-dimensional RSRP map is then accumulated to obtain the three-dimensional RSRP map. Step 7: Calculate the target loss based on the obtained simulated 3D RSRP map and the reconstructed 3D RSRP map obtained in Step 6; Step 8: Update the sparse data converter module in Step 3, the radio map data feature extraction module in Step 4, the multi-scale convolutional pyramid module in Step 5, and the radio map layer-by-layer reconstruction module in Step 6 based on the target loss obtained in Step 7, to obtain a complete three-dimensional radio map reconstruction model. Step 9, Online Phase: Based on the trained model, reconstruct a 3D RSRP map of the target area using the real-time input building distribution map of the target area and the sparse RSRP data points obtained from actual UAV measurements.