A method for three-dimensional reconstruction of large-scale outdoor scenes based on voxelized point clouds
By using a voxelized point cloud optimization method, 3D UNet and SE modules are used to recalibrate the features of the point cloud, and the GIoU loss function is used to constrain the global geometry to generate a high-quality initial point cloud. This solves the problem of poor initial point cloud quality in large-scale outdoor scene 3D reconstruction and improves reconstruction accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF SCI & TECH
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-19
AI Technical Summary
In large-scale outdoor scene 3D reconstruction, existing technologies suffer from poor initial point cloud quality, resulting in low reconstruction efficiency and accuracy. Furthermore, existing methods struggle to effectively process large amounts of point cloud information, exhibiting issues such as over-smoothing, poor noise robustness, and high computational costs.
A voxelization point cloud optimization method is adopted. The point cloud is recalibrated using 3D UNet and SE modules, and the global geometry is constrained by the GIoU loss function. Combined with the inverse voxelization process, a high-quality initial point cloud is generated as the input for 3D Gaussian splashing.
It significantly improves the training effect and convergence speed of 3D Gaussian splashing, generates point clouds with low noise and complete detail preservation, adapts to the initial input requirements of 3D Gaussian splashing, and improves reconstruction accuracy and efficiency.
Smart Images
Figure CN122244324A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer graphics technology, and in particular to a method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds. Background Technology
[0002] With the rapid development of 3D reconstruction technology, 3D Gaussian splashing plays an indispensable role in digital cities. This technology can recreate the spatial form and geometric structure of urban infrastructure such as buildings, roads, and green spaces. Planners can use these models to scientifically conduct urban layout planning, traffic flow analysis, building sunlight simulation, and other tasks, improving the efficiency and quality of urban construction and contributing to the development of smart cities.
[0003] The reconstruction effect and convergence speed of 3D Gaussian splashing are highly dependent on the quality of the initial input point cloud. As the basic data source for 3D reconstruction, the point cloud provides scene geometry and texture information, and is also a key intermediate layer from the raw data to the final 3D model. The higher the point cloud accuracy, the closer it is to the geometry of the real scene, which improves the training speed of 3D Gaussian splashing. The lower the accuracy, the more likely local rendering blur will occur.
[0004] However, large-scale outdoor scene 3D reconstruction requires processing massive amounts of point clouds, resulting in a cubic increase in computation. Generating low-quality initial point clouds leads to low reconstruction efficiency and post-processing overload. Traditional methods, such as Poisson reconstruction, suffer from over-smoothing, loss of fine geometric details, and high computational costs. Statistical filtering can only remove local noise, with thresholds dependent on manual settings and poor generalization. Mainstream modern optimization methods, such as point cloud-based learning methods, struggle to model global spatial geometric consistency and exhibit poor robustness to noise. Methods based on neural implicit fields suffer from low inference efficiency and slow speed. Existing 3D Gaussian splash optimization methods mostly optimize only the rendering process, not the initial point cloud, thus failing to repair shapes at the geometric level. Therefore, achieving fast, high-quality outdoor scene 3D reconstruction with abundant point cloud information is crucial for large-scale outdoor scene 3D reconstruction based on 3D Gaussian splash optimization. Summary of the Invention
[0005] The main objective of this invention is to provide a method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds, which can generate high-quality initial point clouds and significantly improve the training effect and convergence speed of 3D Gaussian splashing.
[0006] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows: A method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds is characterized by optimizing the point cloud quality after voxelization, and then using the devoxed point cloud as the initial input for 3D Gaussian splashing. The method includes the following steps: Step 1: Input the outdoor scene image to be reconstructed to obtain a large amount of initial point cloud information, and then voxelize the point cloud; Step 2: Perform 3D feature recalibration. Input the voxelized point cloud into the network model composed of 3D UNet and SE, and perform voxel feature weighting. Step 3, use Guide geometric and spatial constraints; Step 4: Reverse map the voxel optimization results into a point cloud; Step 5: Use the optimized point cloud as the initial input for 3D Gaussian splashing and perform 3D reconstruction.
[0007] Furthermore, in step 1, a statistical outlier removal method is applied to the collected raw point cloud to calculate the Euclidean distance between each point and all points in its k-neighborhood. : Calculate the mean of the Euclidean distance within the k-neighborhood. and standard deviation v ,when When a neighboring point is identified as an outlier, it is removed; after preprocessing, clean point cloud information is obtained, providing high-quality input for subsequent voxelization.
[0008] The preprocessed point cloud is converted into voxels, and the point cloud is divided using uniform voxels with a voxel resolution of 0.01. A 3D voxel mesh is then constructed, with the mesh boundary determined by the maximum and minimum coordinates of the preprocessed point cloud. Each voxel... For a given 3D mesh cell, its center coordinates Determined by the mesh generation rules, the eigenvalues of a voxel are obtained by fusing features such as the mean coordinates and density of all points within that voxel. The fusion formula is as follows: As shown below: in, voxels The number of points contained within. For point The density value is calculated by the number of neighboring points. If there are no point clouds within a voxel, the feature value is set to 0, resulting in a final dimension of [dimensional value missing]. The voxel feature maps are used as inputs to the subsequent network structure.
[0009] Furthermore, the specific structure of the network model used in step 2 is as follows: 3D UNet employs an encoder-decoder structure, achieving feature extraction and reconstruction through progressive downsampling and upsampling. The final output is an optimized voxel feature map with the same dimension as the input. The mathematical expression for 3D convolution is shown below: in, As input, voxel feature map, The output is a voxel feature map, where W is a 3D convolution kernel with a size of [size missing]. b is the bias term; the 3D max pooling layer achieves downsampling by taking the maximum value within the convolution window, and the 3D deconvolution layer achieves upsampling by padding with zeros.
[0010] An SE module is inserted after each convolutional block in the encoding and decoding stages of the 3D UNet to perform weighting and labeling of the voxel feature channels; the SE module process is as follows: Global average pooling is performed on the 3D voxel feature map to compress the three-dimensional feature of each channel into a one-dimensional feature value: in, This is the voxel feature map of the c-th channel. The squeezing feature value of the c-th channel; through two fully connected layers and The activation function performs a non-linear mapping on the squeezed feature values to obtain the weight coefficients for each channel: in, From feature dimensions Compress to , The compression factor is 1. Restore the feature dimensions to , for Activation function, ensuring weight coefficients The weight coefficients obtained from the excitation are multiplied by the original channel feature map to achieve feature recalibration. Furthermore, in step 3, to ensure that the optimized voxel set can accurately reconstruct the global geometry of the original point cloud, a method is adopted. As the loss function of the network, let the 3D bounding box corresponding to the real voxel set be... The 3D bounding box corresponding to the optimized voxel set output by the network is , The calculation formula is as follows: in, This represents the intersection-union ratio (IoU) between the ground truth bounding box and the predicted bounding box. To be able to contain simultaneously and The minimum 3D bounding box, This represents the volume of a 3D bounding box. The loss function is as follows: The total loss function of the network is The weighted sum of the loss function and the L2 loss function is used to simultaneously constrain the geometry of the voxels and the accuracy of the eigenvalues. The L2 loss function is as follows: in, To optimize the voxels eigenvalues, For real voxels eigenvalues, total loss function as follows: in , These are the weighting coefficients.
[0011] Furthermore, in step 4, the optimized voxel feature map is... Reverse mapping to discrete point cloud This ensures that the output point cloud can be directly used as the initial input for 3DGS without additional format conversion, reducing the introduction of errors. The inverse mapping employs a voxel in-point sampling strategy, the specific process of which is as follows: The optimized voxels are filtered out, and empty voxels with an eigenvalue of 0 are removed, while voxels containing effective geometric features are retained. For each effective voxel Based on its density characteristics Determine the number of sampling points The higher the density, the more sampling points are required to ensure that the point cloud density matches the geometric details and that at least one point is sampled for each region. The formula is as follows: in, This is the sampling coefficient, used to control the overall sampling amount. In each effective voxel... Within a three-dimensional spatial range, uniform random sampling is used to generate There are discrete points, and the coordinates of each point are determined by the coordinates of the voxel center and a random offset: in, s is the voxel resolution, ensuring that the sampling point is located inside the voxel.
[0012] All sampling points are deduplicated to remove duplicates, resulting in the final optimized point cloud. This point cloud features low noise, complete detail preservation, and controllable density, making it suitable for the initial input requirements of 3D Gaussian splashing.
[0013] Furthermore, in step 5, the optimized point cloud... As the initial input for the 3D Gaussian splashed ground, subsequent rendering and reconstruction are completed. The specific process is as follows: Optimized point cloud Each point in As an initial 3D Gaussian Center: in, It is the covariance matrix, representing the scaling and rotation of this 3D Gaussian along each coordinate axis. The initial covariance matrix... The formula is determined based on the local density of the point cloud, as follows: in, The standard deviation is determined by the mean of the local neighborhood distances of the point cloud. for The identity matrix. Using multi-view images as supervision signals, the parameters of the Gaussian set are optimized by minimizing the pixel error between the rendered image and the real image. The objective function is: in, The image output by 3D Gaussian splashing. For real multi-view images, , These represent the height and width of the image, respectively.
[0014] After training convergence, a high-quality 3D scene can be rendered in real time using a Gaussian ensemble of 3D Gaussian splashing, completing the 3D reconstruction process.
[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a voxelization-based point cloud optimization method, which can effectively solve the problems of disorder, sparseness, and high noise in the original point cloud, and achieve point cloud denoising, structural completion, and geometric regularization, outputting high-quality point cloud information.
[0016] This invention designs a 3D UNet network structure that integrates an SE feature recalibration module, which can perform weighted optimization of voxel feature channels, enhance key geometric features, suppress noise features, and significantly improve the accuracy and robustness of point clouds. This invention uses the GIoU loss function to optimize voxel-level point clouds, which can effectively constrain the global geometric overlap and spatial position of the optimized point cloud, avoid global shape distortion, and is more suitable for the initialization requirements of 3D Gaussian splashing than traditional loss functions. This invention designs a complete 3D reconstruction process, including point cloud voxelization, network refinement, voxel inverse mapping, and 3D Gaussian splash initialization. This process can optimize the initial input quality of 3D Gaussian splash from the source, thereby improving the reconstruction accuracy, detail preservation results, and training convergence efficiency of 3D Gaussian splash. Attached Figure Description
[0017] Figure 1 This is the overall flowchart provided by the present invention.
[0018] Figure 2 This is an overall flowchart of the three-dimensional Gaussian splashing technology provided by the present invention.
[0019] Figure 3 This is a general block diagram of the three-dimensional reconstruction method provided by the present invention.
[0020] Figure 4 This is a schematic diagram of a 3D CNN network structure.
[0021] Figure 5 This is a comparison chart of scene reconstruction results on the Palace dataset provided by this invention.
[0022] Figure 6 This is a comparison chart of scene reconstruction results on the truck dataset provided by this invention.
[0023] Figure 7 This is a comparison chart of scene reconstruction results on the training dataset provided by this invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and effects of this invention clearer, the technical solutions of this invention are described in detail below with reference to specific examples. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention. This embodiment A method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds, characterized by the following steps: Step 1: Input the outdoor scene image to be reconstructed to obtain a large amount of initial point cloud information. Preprocess the collected raw point cloud to remove severe outliers and redundant points, reduce the computational cost of subsequent voxelization and network optimization, and improve data quality. We prepared the BlendedMVS dataset, a large-scale static outdoor reconstruction dataset containing scenes of buildings, streets, and historical sites, demonstrating broad representativeness. The scenes also contain numerous weakly textured regions and complex geometric structures, effectively validating the robustness and effectiveness of our proposed method. This dataset is a mainstream benchmark dataset for 3D reconstruction, and the experimental results are authoritative and comparable. During training, one-eighth of the images in the BlendedMVS dataset were used for testing, with the remainder used for training.
[0025] A statistical outlier removal method is used to calculate the Euclidean distance between each point and all points in its k-neighborhood. : Calculate the mean of the Euclidean distance within the k-neighborhood. and standard deviation ,when When an outlier is identified, that neighboring point is removed. This preprocessing yields clean point cloud information, providing high-quality input for subsequent voxelization.
[0026] The preprocessed point cloud is converted into regular voxels. The purpose is to transform the disordered discrete point cloud into an ordered voxel mesh, which is compatible with the convolution operation of 3D UNet, while preserving the global geometric structure of the point cloud.
[0027] The point cloud was divided using uniform voxels with a resolution of 0.01, and a 3D voxel mesh was constructed. The mesh boundary was determined by the maximum and minimum coordinates of the preprocessed point cloud. Each voxel... For a given 3D mesh cell, its center coordinates Determined by the mesh generation rules, the eigenvalues of voxels are obtained by fusing features such as the mean coordinates and density of all points within the voxel. This compensates for the sparsity of the original point cloud and reduces noise interference in the optimization process. The fusion formula... As shown below: .
[0028] in, voxels The number of points contained within. For point The density value is calculated by the number of neighboring points. When there are no point clouds within a voxel, the feature value is set to 0, resulting in a final dimension of [value missing]. The voxel feature maps are used as inputs to the subsequent network structure.
[0029] Step 2: Perform 3D feature recalibration. Input the voxelized point cloud into a network model composed of 3D UNet and SE, and perform voxel feature weighting. This step is the core of the invention. 3D UNet enables multi-scale extraction and reconstruction of voxel features, while the SE module enhances key voxel features and suppresses noisy features. The specific process is as follows: 3D UNet employs an encoder-decoder structure. In the encoding stage, it progressively downsamples through four 3D convolutional layers and 3D max-pooling layers to extract multi-scale spatial features from voxels. In the decoding stage, it progressively upsamples through four deconvolutional layers, skipping connections between low-dimensional, high-semantic features and the scale features corresponding to those in the encoding stage to achieve feature fusion. The final output is an optimized voxel feature map with the same dimension as the input. The mathematical expression for 3D convolution is shown below: in, As input, voxel feature map, The output is a voxel feature map, where W is a 3D convolution kernel with a size of [size missing]. b is the bias term; the 3D max pooling layer achieves downsampling by taking the maximum value within the convolution window, and the 3D deconvolution layer achieves upsampling by padding with zeros.
[0030] Although voxelization has already performed preliminary feature fusion on the point cloud, noise in the original point cloud is still incorporated into the features of some voxels, resulting in a large number of invalid noise features in the voxel feature map. These noise features interfere with the feature extraction process of 3DUNet, making it difficult for the network to distinguish between real geometric features and noise features. The resulting voxels will then suffer from shape deviations and loss of detail. Therefore, an SE module is inserted after each convolutional block in the encoding and decoding stages of 3D UNet to strengthen key features. SE modules are typically used in 2D; this invention applies the SE module to 3D and combines it with 3DUNet. Its specific computation consists of three steps: I. Compression Operation: Unlike the 2D SE module, the SE module of this invention first performs global average pooling on the 3D voxel feature map, compressing the three-dimensional features of each channel into a one-dimensional feature value, and then aggregating them. Three-dimensional information indicates the global feature importance of this channel: in, This is the voxel feature map of the c-th channel. Let be the squeezing feature value of the c-th channel.
[0031] II. Stimulation Operation: Through two fully connected layers and The activation function performs a non-linear mapping on the squeezed feature values to obtain the weight coefficient for each channel. Key feature channels are assigned high weights, while noisy feature channels are assigned low weights. in, From feature dimensions Compress to , The compression factor is 1. Restore the feature dimensions to , for Activation function, ensuring weight coefficients . III. Recalibration Operation: The weight coefficients obtained from the excitation are multiplied by the original channel feature map to achieve feature recalibration. This enhances the key geometric features of the final voxel feature map and suppresses noise features. The formula is as follows: Step 3: Use GIoU to guide geometry and spatial position constraints; Traditional loss functions, such as Chamfer loss, only focus on the local positional deviation of a single point and cannot constrain the global geometry and spatial relationships of the point cloud. Therefore, this invention uses GIoU as the loss function.
[0032] GIoU loss effectively constrains the global geometry, ensuring global overlap between the optimized voxel set and the real geometry to avoid distortion. It is also robust to noise. GIoU focuses on the overlap and offset of the overall bounding box, rather than the local deviation of a single point, thus effectively preventing the network from overfitting to noisy points. Let the 3D bounding box corresponding to the real voxel set be... The 3D bounding box corresponding to the optimized voxel set output by the network is , The calculation formula is as follows: in, The intersection-union ratio (IUU) of the ground truth bounding box and the predicted bounding box reflects the degree of overlap. To be able to contain simultaneously and The minimum 3D bounding box can penalize the spatial offset between the predicted bounding box and the ground truth bounding box. This represents the volume of a 3D bounding box. The loss function is as follows: The network's total loss function is a weighted sum of the GIoU loss function and the L2 loss function, used to simultaneously constrain the voxel geometry and eigenvalue accuracy. The L2 loss function is as follows: in, To optimize the voxels eigenvalues, For real voxels eigenvalues, total loss function as follows: in , These are the weighting coefficients.
[0033] Step 4: Reverse map the voxel optimization results into a point cloud; Furthermore, in step 4, the optimized voxel feature map is... Reverse mapping to discrete point cloud This ensures that the output point cloud can be directly used as the initial input for 3DGS without additional format conversion, reducing the introduction of errors. The inverse mapping employs a voxel in-point sampling strategy, the specific process of which is as follows: The optimized voxels are filtered out, and empty voxels with an eigenvalue of 0 are removed, while voxels containing effective geometric features are retained. For each effective voxel Based on its density characteristics Determine the number of sampling points The higher the density, the more sampling points are required to ensure that the point cloud density matches the geometric details and that at least one point is sampled for each region. The formula is as follows: in, This is the sampling coefficient, used to control the overall sampling amount. In each effective voxel... Within a three-dimensional spatial range, uniform random sampling is used to generate There are discrete points, and the coordinates of each point are determined by the coordinates of the voxel center and a random offset: in, s is the voxel resolution, ensuring that the sampling point is located inside the voxel.
[0034] All sampling points are deduplicated to remove duplicates, resulting in the final optimized point cloud. This point cloud features low noise, complete detail preservation, and controllable density, making it suitable for the initial input requirements of 3D Gaussian splashing.
[0035] Step 5: Use the optimized point cloud as the initial input for 3D Gaussian splashing and perform 3D reconstruction.
[0036] Poor quality of the initial point cloud is the core bottleneck of 3D Gaussian splashing. The reconstruction effect and convergence speed of 3D Gaussian splashing are highly correlated with the quality of the initial input point cloud. However, the original point clouds acquired in practice generally have many problems: high noise levels, a large number of outliers, irregular geometry, and sparse regions. This invention optimizes the original point cloud using the methods described above, thereby avoiding poor reconstruction quality caused by the initial point cloud quality in 3D Gaussian splashing from the source.
[0037] The optimized point cloud As the initial input for the 3D Gaussian splashed ground, subsequent rendering and reconstruction are completed. The specific process is as follows: Optimized point cloud Each point in As an initial 3D Gaussian Center: in, It is the covariance matrix, representing the scaling and rotation of this 3D Gaussian along each coordinate axis. The initial covariance matrix... The formula is determined based on the local density of the point cloud, as follows: in, The standard deviation is determined by the mean of the local neighborhood distances of the point cloud. for The identity matrix. Using multi-view images as supervision signals, the parameters of the Gaussian set are optimized by minimizing the pixel error between the rendered image and the real image. The objective function is: in, The image output by 3D Gaussian splashing. For real multi-view images, , These represent the height and width of the image, respectively.
[0038] After training convergence, a high-quality 3D scene can be rendered in real time using a Gaussian ensemble of 3D Gaussian splashing, completing the 3D reconstruction process.
[0039] Compared with other point cloud optimization methods, such as PointNet++, which are designed only for point cloud denoising and completion, this invention realizes a complete process from image input to 3D reconstruction without additional operations, and is more convenient to use based on mainstream 3DGS.
[0040] Experiments were conducted on the BlendedMVS and Tanks & Temples datasets, primarily using outdoor-related data from these datasets at a resolution of 1920×1080. Model training and testing were both performed on a single RTX 4090 GPU. For each scene optimization, 30,000 iterations of training were performed using the Adam optimizer. During training, the spatial boundary for converting point clouds to voxels was set to 0.1, and the threshold for converting voxels to point clouds was set to 0.5. All other network parameters followed the settings in 3DGS. PSNR, SSIM, and LPIPS were used as evaluation metrics for the rendering results.
[0041] On the BlendedMVS dataset, the proposed method was compared with the baseline method 3DGS, and the results are shown in Table 1. The rendering effect of the proposed method is superior to that of 3DGS. Through the optimization mechanism introduced in the proposed method, the effectiveness of feature extraction and the receptive field are effectively improved, so that the model can ensure pixel-level accuracy and conform to the realism perceived by the human eye.
[0042] Table 1 shows the quantitative comparison results with other methods on the Palace dataset.
[0043] On the Tanks & Temples dataset, the proposed method was compared with the baseline method 3DGS, and the results are shown in Table 2. In the Truck and Train scenarios, the proposed method outperformed other methods in all three metrics. The reconstructed image is highly consistent with the ground truth at the pixel level, with minimal noise, and the generated image is almost perfectly aligned with the original image in terms of brightness, contrast, and object contours, far superior to 3DGS and Mip-NeRF360. The visual structure is clearer and more natural, and the reconstruction result is visually smooth and realistic.
[0044] Table 2 shows the quantitative comparison results with other methods on the Tanks & Temples dataset.
[0045] The above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.
Claims
1. A method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds, characterized in that, Includes the following steps: Step 1: Input the outdoor scene image to be reconstructed to obtain a large amount of initial point cloud information, and then voxelize the point cloud; Step 2: Perform 3D feature recalibration. Input the voxelized point cloud into the network model composed of 3D UNet and SE, and perform voxel feature weighting. Step 3, use Guide geometric and spatial constraints; Step 4: Reverse map the voxel optimization results into a point cloud; Step 5: Use the optimized point cloud as the initial input for 3D Gaussian splashing and perform 3D reconstruction.
2. The method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds according to claim 1, characterized in that: In step 1, a statistical outlier removal method is used on the collected raw point cloud to calculate the Euclidean distance between each point and all points in its k-neighborhood. : Calculate the mean of the Euclidean distance within the k-neighborhood. and standard deviation v ,when When this happens, the neighboring point is identified as an outlier and removed. The preprocessing yields clean point cloud information, providing high-quality input for subsequent voxelization; The preprocessed point cloud is converted into voxels, and the point cloud is divided using uniform voxels with a voxel resolution of 0.
01. A 3D voxel mesh is then constructed, with the mesh boundary determined by the maximum and minimum coordinates of the preprocessed point cloud. Each voxel... For a given 3D mesh cell, its center coordinates Determined by the mesh generation rules, the eigenvalues of a voxel are obtained by fusing features such as the mean coordinates and density of all points within that voxel. The fusion formula is as follows: As shown below: in, voxels The number of points contained within. For point The density value is calculated by the number of neighboring points. If there are no point clouds within a voxel, the feature value is set to 0, resulting in a final dimension of [dimensional value missing]. The voxel feature maps are used as inputs to the subsequent network structure.
3. The method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds according to claim 2, characterized in that: The specific structure of the network model used in step 2 is as follows: 3D UNet employs an encoder-decoder structure, achieving feature extraction and reconstruction through progressive downsampling and upsampling. The final output is an optimized voxel feature map with the same dimension as the input. The mathematical expression for 3D convolution is shown below: in, As input, voxel feature map, The output is a voxel feature map, where W is a 3D convolution kernel with a size of [size missing]. 'b' represents the bias term; the 3D max pooling layer downsamples by taking the maximum value within the convolution window, and the 3D deconvolution layer upsamples by padding with zeros; an SE module is inserted after each convolutional block in the encoding and decoding stages of the 3D UNet to weight and calibrate the voxel feature channels; the SE module process is as follows: Global average pooling is performed on the 3D voxel feature map to compress the three-dimensional feature of each channel into a one-dimensional feature value: in, This is the voxel feature map of the c-th channel. The squeezing feature value of the c-th channel; through two fully connected layers and The activation function performs a non-linear mapping on the squeezed feature values to obtain the weight coefficients for each channel: in, From feature dimensions Compress to , The compression factor is 1. Restore the feature dimensions to , for Activation function, ensuring weight coefficients The weight coefficients obtained from the excitation are multiplied by the original channel feature map to achieve feature recalibration. 。 4. The method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds according to claim 3, characterized in that: In step 3, to ensure that the optimized voxel set can accurately reproduce the global geometry of the original point cloud, the following is adopted: As the loss function of the network, let the 3D bounding box corresponding to the real voxel set be... The 3D bounding box corresponding to the optimized voxel set output by the network is , The calculation formula is as follows: in, This represents the intersection-union ratio (IoU) between the ground truth bounding box and the predicted bounding box. To be able to contain simultaneously and The minimum 3D bounding box, Represents the volume of a 3D bounding box; The loss function is as follows: The total loss function of the network is The weighted sum of the loss function and the L2 loss function is used to simultaneously constrain the geometry of the voxels and the accuracy of the eigenvalues. The L2 loss function is as follows: in, To optimize the voxels eigenvalues, For real voxels eigenvalues, total loss function as follows: in , These are the weighting coefficients.
5. The method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds according to claim 4, characterized in that: In step 4, the optimized voxel feature map Reverse mapping to discrete point cloud This ensures that the output point cloud can be directly used as the initial input for 3DGS without additional format conversion, reducing the introduction of errors; the inverse mapping adopts a voxel in-point sampling strategy, the specific process of which is as follows: The optimized voxels are filtered out, and empty voxels with an eigenvalue of 0 are removed, while voxels containing effective geometric features are retained. For each effective voxel Based on its density characteristics Determine the number of sampling points The higher the density, the more sampling points are required to ensure that the point cloud density matches the geometric details and that at least one point is sampled for each region. The formula is as follows: in, The sampling coefficient is used to control the overall sampling amount; in each effective voxel Within a three-dimensional spatial range, uniform random sampling is used to generate... There are discrete points, and the coordinates of each point are determined by the coordinates of the voxel center and a random offset: in, s is the voxel resolution, ensuring that the sampling point is located inside the voxel; All sampling points are deduplicated to remove duplicates, resulting in the final optimized point cloud. This point cloud features low noise, complete detail preservation, and controllable density, making it suitable for the initial input requirements of 3D Gaussian splashing.
6. The method for large-scale outdoor scene 3D reconstruction based on voxelized point clouds according to claim 5, characterized in that: In step 5, the optimized point cloud As the initial input for the 3D Gaussian splashed ground, subsequent rendering and reconstruction are completed. The specific process is as follows: Optimized point cloud Each point in As an initial 3D Gaussian Center: in, It is the covariance matrix, representing the scaling and rotation of this 3D Gaussian along each coordinate axis. The initial covariance matrix... The formula is determined based on the local density of the point cloud, as follows: in, The standard deviation is determined by the mean of the local neighborhood distances of the point cloud. for The identity matrix; using multi-view images as supervision signals, the parameters of the Gaussian set are optimized by minimizing the pixel error between the rendered image and the real image. The optimization objective function is: in, The image output by 3D Gaussian splashing. For real multi-view images, , These are the height and width of the image, respectively; After training convergence, a high-quality 3D scene can be rendered in real time using a Gaussian ensemble of 3D Gaussian splashing, completing the 3D reconstruction process.