Drainage pipeline three-dimensional imaging method and system based on neural radiation field and gaussian sputtering fusion reconstruction

Abstract

The application discloses a kind of based on neural radiation field and Gaussian sputtering fusion reconstruction drainage pipe three-dimensional imaging method and system, it is related to computer vision technology and deep learning technical field.The method is calculated to the standardized multi-view image of acquisition and is trained to generate point cloud data, is initialized sparse three-dimensional structure reconstruction using Gaussian sputtering, neural radiation field initialization is guided using sparse three-dimensional structure, respectively through Gaussian sputtering model and neural radiation field network generation rendering image in the same perspective, obtain three-dimensional imaging model by fusing two kinds of rendering images;The application aims to realize efficient, high-precision, physically credible three-dimensional reconstruction under the internal environment of uneven illumination, serious occlusion, sparse texture by technology fusion and field knowledge embedding, provide reliable data basis for pipe disease diagnosis and quantitative analysis.

Description

Technical Field

[0001] This invention relates to computer vision technology and deep learning technology, and is a method and system for high-precision three-dimensional imaging of drainage pipes that integrates neural radiation field (NeRF) and Gaussian sputtering and combines pipe physical constraints for fine optimization. Background Technology

[0002] Urban drainage pipelines are core underground infrastructure for ensuring urban water security and preventing flooding. Precise and efficient 3D imaging of their internal structure is a prerequisite for assessing pipeline health, accurately locating defects, and making scientific maintenance decisions. Currently, 3D imaging technologies applied to drainage pipeline inspection mainly include:

[0003] LiDAR scanning and ultrasonic imaging technologies: While these active detection technologies can directly acquire geometric information, they have significant limitations in practical applications for drainage pipelines. LiDAR is susceptible to obstruction by water vapor and suspended particles inside the pipe, as well as interference from deposits on the pipe wall, resulting in missing point cloud data and high noise levels. Ultrasonic imaging has limited resolution, making it difficult to clearly present the surface texture and subtle defects of the pipe. In addition, the related equipment is expensive and the data processing flow is complex, making large-scale engineering deployment difficult.

[0004] Traditional photogrammetry methods: This method is based on the stereo vision principle of multi-view images and calculates the 3D structure through feature point matching and structure-of-motion (SfM). However, the interior of drainage pipes often has problems such as uneven lighting, repetitive or missing textures (e.g., smooth concrete pipe walls) and a large number of obstructions (e.g., siltation), which can easily lead to feature matching failures, resulting in distorted and hollow 3D models or a serious decrease in accuracy.

[0005] Reconstruction methods based on Neural Radiation Fields (NeRF): NeRF implicitly represents the continuous volumetric radiation field of a scene through neural networks, enabling the synthesis of highly realistic new perspective images from multi-view images with strong detail restoration capabilities. However, it suffers from two major bottlenecks: First, the training and rendering processes are computationally expensive and time-consuming, making it difficult to meet the needs of rapid pipeline detection and real-time preview; second, in scenes with long, repetitive structures like pipelines, NeRF is prone to issues such as blurred distant details and distorted geometric structures due to limited viewpoint coverage and gradient problems, leading to difficulties in model convergence.

[0006] Reconstruction methods based on Gaussian sputtering: This emerging technology uses an explicit, optimizable set of 3D Gaussian ellipsoids to represent the scene, resulting in extremely fast rendering speeds and real-time interactivity. However, its drawbacks include relatively weak modeling capabilities for complex lighting changes. In areas with distinct light and shadow transitions within pipes, color distortion or loss of detail can easily occur. Furthermore, for low-volume defects such as thin layers of sediment or minute cracks adhering to the pipe walls, a purely geometric Gaussian representation may not be able to accurately capture their shape and boundaries.

[0007] Furthermore, existing 3D reconstruction technologies generally treat the target as a general object, failing to fully incorporate the specific prior physical knowledge of drainage pipes. For example, the pipe diameter is relatively constant in design and within a local area; the pipe wall material (such as concrete / PVC) has specific surface optical properties (reflectivity range); and the accumulation pattern of sediment is constrained by fluid dynamics principles, exhibiting a natural angle of repose. Reconstruction models lacking such constraints may produce results that violate common sense physics, such as inconsistent pipe diameters or sediment appearing suspended, reducing the credibility and engineering practical value of the reconstruction results.

[0008] Therefore, how to synergistically utilize NeRF's high-fidelity detail modeling capabilities and Gaussian sputtering's efficient geometric representation capabilities, and effectively embed pipeline domain knowledge as physical constraints, in order to overcome the shortcomings of single technologies and achieve high-precision, high-efficiency, and high-robust 3D imaging of the interior of drainage pipelines, has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] Objective: To address the shortcomings of existing 3D imaging methods for drainage pipes mentioned in the background section regarding accuracy, efficiency, robustness, and physical plausibility, this invention provides a 3D imaging method and system for drainage pipes based on the fusion and reconstruction of neural radiation fields and Gaussian sputtering. This method aims to achieve efficient, high-precision, and physically reliable 3D reconstruction of the internal environment of pipes with severe uneven illumination, occlusion, and sparse texture through technology fusion and domain knowledge embedding, providing a reliable data foundation for pipe defect diagnosis and quantitative analysis.

[0010] The technical solution of this invention:

[0011] A three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation fields and Gaussian sputtering includes:

[0012] Point cloud image data is generated by calculating and training the collected standardized multi-view images;

[0013] The point cloud data is initialized and reconstructed using a Gaussian sputtering model to obtain a sparse three-dimensional structure model.

[0014] The neural radiation field network is initialized using a sparse 3D structure model. Geometric features of Gaussian elements are extracted from the sparse 3D structure model and constructed as a feature vector F. gauss F gauss Including the spatial distribution histogram and color distribution of Gaussian elements, density field network MLP1 and color field network MLP2 of the neural radiation field network are constructed, and the feature vector F of the Gaussian elements is used to construct the neural radiation field network. gauss As a priori, the weights of MLP1 and MLP2 are initialized. The weights of the first layer of MLP1 are initialized according to the spatial distribution histogram of Gaussian elements, and the weights of the output layer of MLP2 are initialized according to the color distribution of Gaussian elements. The pipeline space is uniformly sampled, and the initial density field ρ0 and initial color field c0 are calculated for each sampling point through MLP1 and MLP2 and stored as the initial model of the neural radiation field network.

[0015] Based on the initial model of the neural radiation field network, we perform neural radiation field network modeling, training, and fusion optimization.

[0016] Rendered images are generated from the same viewpoint using a Gaussian sputtering model and a neural radiation field network, respectively. The two rendered images are then fused to obtain a three-dimensional imaging model.

[0017] When generating the rendered image, the geometric deviation between the Gaussian sputtering model and the neural radiation field network is calculated, and a joint loss function L is constructed. total Iterative optimization involves simultaneously calculating the gradients of the Gaussian distribution parameters and the weights of the neural radiation field network, and then fusing the rendered images generated by the Gaussian sputtering model and the neural radiation field network according to certain weights.

[0018] After fusing the two rendered images to obtain a 3D image, physical constraints are embedded into the 3D image for constraint loss calculation. These physical constraints include pipe diameter, material, and siltation boundary. The physical constraints are then combined with the joint loss function L. total The final optimized loss is obtained by combining the weights, the parameters are adjusted and iterative optimization is performed, and the optimized 3D imaging model is output.

[0019] By linking the 3D imaging model with GIS data, establishing a mapping relationship between the 3D model coordinate system and the geographic coordinate system, mileage annotation is performed on the 3D imaging model to identify siltation areas and defects, and visualization output and data storage are achieved.

[0020] The specific steps include:

[0021] Step 1: Multi-source pipeline scene data acquisition and calibration. Simultaneously acquire multi-view image data, equipment posture data and pipeline basic parameters inside the drainage pipeline, complete the calibration of equipment internal and external parameters, and unify them to a preset three-dimensional coordinate system with the pipeline axis as the X-axis.

[0022] Step 2: Calculate and train the collected standardized multi-view images to generate point cloud data and reduce noise. Perform feature point matching on the multi-view images to generate descriptors. Use the nearest neighbor ratio method and RANSAC algorithm to perform feature matching and outlier removal to form an initial sparse point cloud. Calculate the three-dimensional coordinates of the feature points.

[0023] Step 3: Gaussian sputtering initialization and sparse 3D structure reconstruction. Gaussian sputtering initialization optimizes the Gaussian distribution parameters by minimizing photometric loss, adaptively selects effective Gaussian elements, and outputs a sparse 3D structure model containing a set of effective Gaussian elements and their corresponding spatial location information.

[0024] Step 4: Model the fusion of neural radiation field and Gaussian sputtering. Initialize the neural radiation field using a Gaussian structure and construct the joint loss function L. total Iterative optimization, joint loss function L total This includes photometric loss, geometric consistency loss, and smoothness loss;

[0025] Step 5: Combine the model refinement optimization with pipeline physical constraints, embed the pipe diameter constraint, material reflectivity constraint and siltation boundary constraint into the joint loss function, and refine the parameters of the fusion model to ensure that the fusion model conforms to the pipeline physical characteristics;

[0026] Step 6: Generate a high-density 3D model and associate it with GIS. Perform high-resolution voxel sampling on the refined and optimized fusion model to generate a high-density point cloud and grid model. Associate it with geographic information system data and achieve multi-terminal visualization output.

[0027] Step 3 includes:

[0028] S3.1 Initialization of Gaussian Distribution Set: Randomly sample points from the dimensionality-reduced point cloud as the initial centers of Gaussian elements; calculate initial weights based on the pixel brightness and depth confidence of the sampled points; set the initial scale based on the variance of the neighboring point cloud; construct a local coordinate system based on the sampled point normal vector and the pipeline axis to determine the initial rotation parameters; and set a uniform initial transparency; determine the Gaussian center, establish the mapping relationship between Gaussian elements and point cloud sampled points, calculate the initial Gaussian parameters, and generate the initial Gaussian distribution set;

[0029] S3.2 Obtain the intrinsic and extrinsic parameters of the multi-view image. Generate virtual camera rays for each viewpoint and sample K points along the ray direction. For each sampled point, calculate its density ρ and color c using Gaussian element parameters. Calculate the rendered image using ray integration and calculate the photometric loss L between the rendered image and the original image. photo Update the parameters and iteratively optimize the Gaussian distribution set;

[0030] S3.3 Effective Gaussian Element Filtering: Set an adaptive threshold to remove redundant Gaussian elements with too low weights and output a sparse 3D structure model.

[0031] Step 4 includes:

[0032] S4.1 Gaussian Structure-Guided Neural Radiation Field Initialization: Based on the sparse three-dimensional structure model obtained in step 3, extract the geometric features of Gaussian elements to construct the feature vector F. gauss The density field network MLP1 and color field network MLP2 of the neural radiation field are constructed, and the eigenvectors F of Gaussian elements are... gauss As a priori, the weights of MLP1 and MLP2 are initialized. The weights of the first layer of MLP1 are initialized according to the spatial distribution histogram of Gaussian elements. The output layer of MLP2 is initialized according to the color distribution of Gaussian elements. The pipeline space is uniformly sampled. The initial density field ρ0 and color field c0 are calculated through MLP1 and MLP2 and stored as the initial model of the radiation field.

[0033] S4.2 Construction and Feature Fusion of Multi-Scale Neural Radiation Field Network: Based on the initial model of the neural radiation field network, neural radiation field modeling, training, and feature fusion optimization are performed. The multi-scale neural radiation field network contains 4 fully connected layers with ReLU+Sigmoid activation function. gauss The multi-dimensional features fused with multi-scale are concatenated and input into MLP1 and MLP2. The image is then divided into training and validation sets for iterative training.

[0034] S4.3 Fusion Optimization Iteration: For the same viewpoint, render images I are generated using both a Gaussian sputtering model and a neural radiation field network. gauss and I nerf ; Calculate I gauss Photometric loss L photo Calculate I gauss and I nerf Geometric deviation L geo Calculate I nerf gradient smoothness L smooth Joint loss function L total =0.6×L photo +0.3×L geo +0.1×L smooth The Adam optimizer is used to simultaneously calculate the gradients of the Gaussian distribution parameters and the neural radiation field network weights, update the parameters, and perform iterative optimization to calculate the fused rendered image I. fusion =0.4×I gauss +0.6×I nerf .

[0035] Step 5 includes:

[0036] S5.1 Pipe diameter constraint, material reflectivity constraint, and siltation boundary constraint are embedded in the loss function:

[0037] Pipe diameter constraint: Based on the pipe design diameter, an allowable range of radial distance is set. Gaussian elements or sampling points that exceed the range are penalized, and the pipe diameter constraint loss L is calculated. r ;

[0038] Material constraints: Based on the preset surface reflectivity range of the pipe material, a penalty is applied to cases where the reflectivity calculated for the rendered color exceeds this range, and the material constraint loss L is calculated. m ;

[0039] Siltation boundary constraint: Based on fluid dynamics principles, a maximum slope threshold for the siltation surface is set. A penalty is applied to areas identified as siltation regions where the surface slope exceeds this threshold, and the siltation boundary constraint loss L is calculated. m ;

[0040] Total constraint loss L const =0.4×L r +0.3×L m +0.3×L s ;

[0041] S5.2 Refined Optimization: Final Optimization Loss L final =L total +0.5×L const An adaptive momentum estimation optimizer is used to adjust parameters and perform iterative calibration. After satisfying the iterative conditions, a refined optimized fusion model is obtained.

[0042] A 3D imaging system for drainage pipelines based on the fusion reconstruction of neural radiation fields and Gaussian sputtering is provided to execute the aforementioned 3D imaging method for drainage pipelines based on the fusion reconstruction of neural radiation fields and Gaussian sputtering. The system includes a multi-source pipeline data acquisition subsystem, a data preprocessing subsystem, a fusion reconstruction subsystem, a model optimization subsystem, and a 3D imaging visualization subsystem. The system structure is divided into a data access control layer, an access control layer, and an application layer. The hardware support platform includes a computing terminal, a database server, a GIS server, and detection equipment. The fusion reconstruction subsystem includes a Gaussian sputtering initialization module, a multi-scale NeRF network module, and a joint optimization module to achieve synergistic optimization of sparse structures and high-precision radiation fields. The application layer adopts a hybrid C / S and B / S architecture, supporting 3D model browsing, sectioning annotation, and GIS-related query functions.

[0043] The beneficial effects of this invention are:

[0044] 1. High reconstruction accuracy and detail reproduction: Through the deep integration of NeRF and Gaussian sputtering, it combines NeRF's high-fidelity modeling capability for fine visual effects such as complex lighting and semi-transparent subsurface scattering, with Gaussian sputtering's accurate expression capability for explicit geometry, significantly improving the reconstruction accuracy in challenging areas such as pipe shadows and thin layers of dirt accumulation.

[0045] 2. Significantly Improved Modeling Efficiency and Practicality: The introduction of Gaussian sputtering provides NeRF with high-quality geometric initialization, greatly accelerating the training convergence speed. Simultaneously, Gaussian sputtering itself supports real-time rendering, enabling rapid previewing and interaction both during and after training, better meeting the timeliness requirements of engineering inspection.

[0046] 3. Strong Model Robustness and Generalization Ability: The fusion model reduces its dependence on the quality of input data. Gaussian geometry provides support in regions where image texture is sparse and matching is difficult; NeRF radiation field supplements in regions with blurred geometry but rich color information. The complementary advantages of both improve the success rate and stability of reconstruction in complex pipeline environments.

[0047] 4. High physical reliability of reconstruction results: The innovative introduction of pipeline physical constraints transforms industry prior knowledge into part of the optimization objective, ensuring that the generated 3D model conforms to engineering reality in terms of pipe diameter, material, and siltation morphology, avoiding the distortion phenomenon of "looking real but physically unreasonable", and making the reconstruction results directly applicable to engineering measurement and analysis.

[0048] 5. Forming a complete technical closed loop: This invention provides a complete end-to-end 3D imaging solution for drainage pipeline inspection scenarios, from multi-source data acquisition, fusion, intelligent processing to physical constraint optimization and final visualization output, with good system integration and engineering application prospects. Attached Figure Description

[0049] Figure 1 Schematic diagram of the fusion modeling of neural radiation field and Gaussian sputtering.

[0050] Figure 2 Flowchart of a three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering.

[0051] Figure 3 Comparison of 3D reconstruction results from different algorithms.

[0052] Figure 4 The processing diagram of this invention is shown. Detailed Implementation

[0053] Example 1:

[0054] Step 1: Data Acquisition and Calibration for Multi-Source Pipeline Scenarios

[0055] The system utilizes an unmanned pipeline measurement device (UPM), an IoT-based radar detection vehicle, a video surveillance robot, a lidar sensor, and a high-definition camera to simultaneously collect multi-view image data of the drainage pipeline interior, equipment attitude data (position and angle), and pipeline basic parameters (pipe diameter, material, laying slope). Camera intrinsic parameter calibration is performed using the Zhang Zhengyou calibration method, and equipment extrinsic parameter calibration is achieved through IMU and GPS fusion positioning. A unified coordinate system for the multi-source data is established (a three-dimensional coordinate system with the pipeline start-up end as the origin and the pipeline axis as the X-axis).

[0056] Step 2: Multi-source data preprocessing

[0057] 2.1 Point cloud data is generated by computationally training the acquired standardized multi-view images. Statistical filtering algorithms are used to remove outlier noise points from the point cloud, median filtering is used to remove salt-and-pepper noise from the image data, and wavelet transform is used to suppress signal interference from radar data. Preprocessing of the multi-view images is performed using feature point matching (SIFT algorithm), including grayscale conversion (converting RGB images to single-channel grayscale images, grayscale value = 0.299R + 0.587G + 0.114B) and Gaussian blur (σ = 1.6, kernel size = 5 × 5) (σ is the standard deviation of the Gaussian function in the Gaussian blur algorithm, which is the core parameter controlling the degree of blur).

[0058] 2.2 Key Point Detection: The key point detection steps are as follows:

[0059] ① Find the extreme points by using the "Gaussian difference pyramid" (6 octaves, 4 layers per group), perform 3D interpolation (sub-pixel level positioning) on ​​the extreme points, remove low contrast points (contrast threshold = 0.03) and edge points (principal curvature ratio threshold = 10) to obtain stable key points;

[0060] ② Descriptor generation: For each keypoint, a 16×16 neighborhood is constructed with it as the center, which is divided into 4×4 sub-blocks. The gradient histogram of each sub-block is calculated in 8 directions to generate a 128-dimensional descriptor, and normalization is performed to eliminate the influence of illumination.

[0061] ③ Feature matching: The "nearest neighbor ratio method" is used to calculate the Euclidean distance between the descriptors of the source image and the target image. If the ratio of the nearest distance to the second nearest distance is ≤0.8 (matching threshold, used to balance the accuracy and the number of matches), it is determined to be a valid match.

[0062] ④ Outlier Removal: False matches are removed using the RANSAC algorithm (iterations = 1000, inlier threshold = 2 pixels) to ensure a matching accuracy ≥ 95%, outputting a spatially aligned image. RANSAC (Random Sample Consensus) is an efficient method for estimating mathematical model parameters from a sample set containing outliers.

[0063] 2.3 Data Dimensionality Reduction: The registered point cloud data is reduced by voxel grid downsampling to retain key structural features; the image data is scale normalized to improve the efficiency of subsequent modeling.

[0064] Step 3: Gaussian sputtering initialization and sparse 3D structure reconstruction

[0065] 3.1 Initialization of Gaussian distribution set:

[0066] ①Startup conditions: The device reads the standardized multi-view image and point cloud data from step 2, and automatically starts initialization after confirming that the data format is correct and the key features are completely preserved;

[0067] ② Gaussian center determination: Randomly sample the point cloud data after dimensionality reduction, and randomly select sampling points from the point cloud as the initial centers of Gaussian elements; assign a unique ID to each Gaussian element, and establish a mapping relationship between Gaussian elements and point cloud sampling points to facilitate subsequent optimization and tracking;

[0068] ③ Initial calculation of Gaussian parameters (weights w, scale s, rotation parameters R, transparency α):

[0069] Weight w: Read the pixel brightness value L (0-255) of the corresponding sampling point in the multi-view image, and calculate w = L / 255 × D confidence (D) confidence Depth confidence is calculated from lidar depth data: D confidence =1-|d measured -d predicted | / d measured , where d measured For the measured depth, d predicted (Mean depth of the neighborhood); Scale s: Calculate the variance σ of the point cloud within the neighborhood (radius = 10 mm) of the sampling point, set s = (σ x ,σ y ,σ z (Variance of the point cloud in the x, y, z directions), with an initial scale range of: radial s r =3mm-5mm, circumferential sc=2mm-4mm, axial s a =4mm-6mm; Rotation parameter R: Calculate the normal vector n of the sampling point (obtained by fitting the plane through neighborhood points), construct a local coordinate system with the normal vector as the z-axis and the pipe axis as the x-axis, and generate the initial rotation matrix R to ensure that the orientation of the Gaussian element is in line with the pipe surface; Transparency α: The initial value is set to 0.8 (default is opaque, and will be adjusted through optimization later); ④ Output: Generate an initial Gaussian distribution set (containing 50,000 Gaussian elements, each Gaussian element contains parameters (x0,y0,z0,w,s,R,α)), where x0,y0,z0 are the initial three-dimensional coordinates of the center of the Gaussian element.

[0070] 3.2 Iterative optimization of Gaussian distribution parameters:

[0071] ① Differentiable rendering calculation: Reading intrinsic parameters of multi-view images (focal length f, principal point coordinates (c...)). x ,c y The system generates virtual camera rays (rays from the camera optical center to each pixel of the image) for each viewpoint, along with extrinsic parameters (pose matrix). For each ray, it samples K points along the ray direction (K=64, based on the pipeline depth range). For each sampled point, it calculates its density ρ and color c using Gaussian parameters: ρ=Σw i ×N(p;μ i ,Σ i )(w i These are the weight parameters for each Gaussian element, where N is the Gaussian probability density function, and μ is the weight parameter for each Gaussian element. i The center is a Gaussian matrix, and Σi is the scaling matrix = R. i ×diag(s i )×R i T c=Σ(w) i ×c i ×N(p;μ i ,Σ i )) / ρ(c i The color is a Gaussian pixel, determined by the color of the corresponding image pixel, R. i (where p is the rotation parameter matrix for each Gaussian element and p is the coordinate of a point); the rendered image I is calculated using ray integration. rendered :I rendered =∫0^∞ρ(t)×c(t)×exp(-∫0^tρ(τ)dτ)dt (t is the distance parameter of light propagation, ρ(t) is the sampling point density at distance t, c(t) is the color of the sampling point at distance t, exp refers to the natural number e, and the numerical integration uses the trapezoidal rule with a step size of 0.5mm); ②Photometric loss calculation and parameter update: Calculate the rendered image I rendered With the original image I original Photometric loss: L photo =(1 / HW)×Σ|I rendered (p)-I original (p)| (HW is the image height and width, p is the coordinates of a pixel); A stochastic gradient descent (SGD) optimizer is used with a learning rate of 0.001. The gradient is calculated for the parameters (x, y, z, w, s, R, α) of each Gaussian element (implemented using the automatic differentiation framework PyTorch), and the parameter θ is updated. new =θ old -lr×∇L photo ;θnew It is the new set of parameters of the Gaussian elements after iterative updates, θ old It is the set of old parameters of the Gaussian elements before the iterative update, lr is the step size control factor, ∇ is the gradient of the loss function, and L photo This represents the pixel color difference loss between the rendered image and the original image in the pipeline. During the iteration process, the current loss value is calculated every 20 iterations. If the loss decreases by less than 1×10⁻⁶ for five consecutive iterations, the loss is considered lost. -5 If the learning rate is reduced to 0.5 times its original value, then the termination condition is: when the number of iterations is ≥100 and L... photo When the value is ≤0.02 (the threshold of the original scheme) or the number of iterations is ≥200 (to prevent overfitting), stop the optimization and output the optimized Gaussian distribution set; ④ Result verification: calculate the structural similarity between the rendered image and the original image (SSIM≥0.9). If the condition is met, proceed to the next step.

[0072] 3.3 Effective Gaussian Filtering:

[0073] ① Calculate the mean μ of the weights w of all Gaussian elements. w and standard deviation σ w Set the adaptive threshold T w =μ w -0.5σ w (Used to remove redundant Gaussian units with excessively low weights); ② Traverse all Gaussian units and remove the following two types of invalid Gaussian units: weight w <T w Gaussian elements with low contribution; Gaussian elements with any dimension >20mm at scale s (abnormally large size, possibly noise); ③ After filtering, the number of Gaussian elements is kept at 30,000-40,000 (60%-80% of the original number), the Gaussian element ID index is re-established, and the sparse three-dimensional structure model (including the effective set of Gaussian elements and the corresponding spatial location information) is output.

[0074] Step 4: Modeling the fusion of neural radiation field and Gaussian sputtering

[0075] 4.1 Gaussian structure-guided NeRF initialization:

[0076] ① Read the sparse 3D structure model output from step 3 and extract the geometric features of Gaussian elements: Global features: Spatial distribution histogram of Gaussian centers (divided into 50 intervals according to the pipe axis x, radial y, and circumferential z, and the number of Gaussian elements in each interval is counted); Local features: Scale s, rotation matrix R, and weight w for each Gaussian element, constructing a feature vector F. gauss =(s x ,s y ,s z ,R 11 ,R 12 ,R 13 ,...,w)(sx ,s y ,s z These represent the scales of the Gaussian elements in the x, y, and z directions of the pipe's three-dimensional coordinate system, corresponding to the axial, radial, and circumferential coverage areas of the pipe, respectively. R 11 ,R 12 ,R 13 It is 3 The elements in the first row and first column of matrix R, the elements in the first row and second column, and the elements in the first row and third column describe the rotational components of Gaussian elements about the pipe in the axial, radial, and circumferential directions (dimension = 12).

[0077] ② NeRF Network Initialization: Construct the NeRF density field network (MLP1) and color field network (MLP2). MLP1 takes 3D coordinates (x, y, z) and global geometric features as input, and outputs density ρ and intermediate feature vectors (256-dimensional). MLP2 takes intermediate feature vectors and view direction (d) as input. x ,d y ,d z )(d x ,d y ,d z (This refers to the components of the viewing direction in the pipe's axial, radial, and circumferential directions), output color (c r ,c g ,c b )(c r ,c g ,c b (These are the red, green, and blue channel values, representing the intensity of the red, green, and blue components at that location) This relates to the geometric features of the Gaussian elements, F. gaussAs prior geometric constraints, the weights of MLP1 and MLP2 are initialized: The weights of the first layer of MLP1 are initialized according to the spatial distribution histogram of Gaussian elements (high-density regions correspond to larger weights); the output layer of MLP2 is initialized according to the color distribution of Gaussian elements, ensuring the initial color matches the pipe surface. Specifically, a sparse 3D structure model generated by Gaussian sputtering is used as prior knowledge to guide the initial state of the Neural Radiation Field (NeRF) network. The spatial locations of all Gaussian elements are extracted from the sparse structure, and their distribution histograms in the pipe's axial, radial, and circumferential directions are statistically analyzed. For the weights of the first layer of the density field network MLP1, values ​​are assigned based on this histogram: in high-density spatial regions where Gaussian elements cluster (such as the pipe wall surface), the corresponding network connection weights are assigned larger initial values, making the network focus more on these geometrically significant regions. Simultaneously, the color information (RGB distribution) of the Gaussian elements is extracted and used to initialize the output layer weights of the color field network MLP2, ensuring that the network has a reasonable estimate of the typical color of the pipe surface (such as the grayish-white of concrete) at the beginning of training, guaranteeing that the initial rendered color is basically consistent with the actual pipe wall appearance. This process injects explicit geometry and color priors into the implicit neural field, significantly accelerating the convergence speed and stability of subsequent fusion training.

[0078] ③ Radiation field parameter initialization: Uniform sampling is performed on the pipe space (sampling interval = 1mm). For each sampling point, the initial density field ρ0 and color field c0 are calculated using MLP1 and MLP2, and stored as the initial model of the Neural Radiation Field Network. This step transforms the pipe geometry prior of Gaussian sputtering into initial field data that can be directly accessed by the Neural Radiation Field (NeRF). The initial model of the Neural Radiation Field Network becomes the foundational model for subsequent NeRF network modeling, training, and fusion optimization.

[0079] 4.2 Multi-scale NeRF Network Construction and Feature Fusion: Based on the initial model of the neural radiation field network, neural radiation field modeling, training, and fusion optimization are performed, including:

[0080] ① Network Structure Details: The multi-scale NeRF network contains 4 fully connected layers (hidden layer dimension = 256), with ReLU (first 3 layers) + Sigmoid (output layer) activation functions; Multi-scale Feature Fusion Module: The input 3D coordinates (x, y, z) are scaled to 1× (original scale), 2× (scaled by 2 times), and 4× (scaled by 4 times) respectively, input into 3 parallel sub-networks, each sub-network outputs 256-dimensional features, which are concatenated to obtain 768-dimensional features, and then compressed to 256 dimensions through 1 fully connected layer as the final features; ② Geometric Feature Guidance: The Gaussian local features F extracted in step 4.1 are... gaussThe 256-dimensional features after multi-scale fusion are concatenated to obtain 268-dimensional features, which are then input into MLP1 and MLP2 to enhance the network's ability to perceive details of the pipeline (such as small silt deposits in cracks); ③ Network training preparation: Divide the training set (80% multi-view images) and the validation set (20% multi-view images), set the batch size (batchsize=8), and the number of training iterations = 20000 times.

[0081] 4.3 Integration, Optimization, and Iteration:

[0082] ① Dual-model rendering result generation: For the same viewpoint, render images I are generated separately using a Gaussian sputtering model and a NeRF network. gauss and I nerf Gaussian sputtering rendering: follows the differentiable rendering logic from step 3.2; NeRF rendering: samples 128 points along the ray (the first 64 are uniform samples, and the last 64 are importance samples), calculates ρ and c for each sample point using MLP1 and MLP2, and obtains I by ray integration. nerf ② Joint loss function calculation: photometric loss L photo Same as step 3.2, weight = 0.6; geometric consistency loss L geo : Calculate the geometric deviation L between the Gaussian sputtering model and the NeRF network geo L geo =(1 / N gauss )×Σ||μ i -μ i ^nerf||(μ i Centered at Gaussian, μ i ^nerf represents the density peak at the corresponding location in the NeRF network (weight = 0.3); smoothness loss L smooth : Calculate the gradient smoothness of the NeRF radiation field, L smooth =∫||∇ρ(x,y,z)|| 2 dV (integration range is the pipe space, weight = 0.1) (x, y, z are the spatial coordinates of any sampling point in the pipe's three-dimensional coordinate system, corresponding to the pipe's axial, radial, and circumferential directions, respectively); Total loss L total =0.6×L photo +0.3×L geo +0.1×L smooth③ Joint parameter update: The Adam optimizer is used (learning rate = 0.0001, β1 = 0.9, β2 = 0.999), and the gradients of the Gaussian distribution parameters (x, y, z, w, s, R, α) and the NeRF network weights are calculated simultaneously. The update strategy is as follows: for the first 10,000 iterations, the Gaussian parameter update step size is 0.001, and the NeRF network update step size is 0.0001; for the next 10,000 iterations, the Gaussian parameter update step size is reduced to 0.0005, while the NeRF network update step size remains unchanged (prioritizing the optimization of NeRF details); ④ Termination condition: when the number of iterations is ≥ 20,000, and the L of the validation set is... total ≤0.015, stop optimization; ⑤ Result verification: Calculate the fused rendered image (I fusion =0.4×I gauss +0.6×I nerf If the PSNR of the fused rendered image is ≥40dB (PSNR, Peak Signal-to-Noise Ratio, is a core quantitative indicator for measuring the similarity between the fused rendered image and the original pipeline), then proceed to the next step.

[0083] Step 5: Refine and optimize the model based on pipeline physical constraints

[0084] 5.1 Physical constraint embedding and constraint loss calculation:

[0085] ① Pipe diameter constraint: Read the basic pipe parameters (e.g., DN800 pipe, pipe diameter D=800mm), set the radial constraint range [D-5mm, D+5mm] (795mm-805mm); for all Gaussian elements in the fused model, calculate their radial distance r=(y 2 +z 2 ) 1 / 2 (The pipe axis is the x-axis). If r < 795 mm or r > 805 mm, calculate the constraint loss L. r =|r-800| / 800 (radial deviation normalization); if r is within the range, L r =0; ② Material constraints: Read the pipe material (concrete), query the preset material reflectivity database (concrete reflectivity range = 0.3-0.5); output the color field c=(c) of the NeRF network. r ,c g ,c b The reflectivity R is calculated to be 0.299c. r +0.587c g +0.114c b (Consistent with grayscale value calculation); If R < 0.3 or R > 0.5, calculate the constraint loss L. m =|R-0.4| / 0.4 (based on a mean reflectance of 0.4); if within the range, L m=0; ③Sedimentation boundary constraint: Based on the principle of fluid dynamics, a preset sedimentation surface slope threshold θ is used. max =0.15 (maximum slope of natural sediment deposition); For the sedimentation region in the fusion model (determined by Gaussian weight w≥0.7 and depth>5mm of pipe wall surface), calculate the surface slope θ of each sedimentation point (by fitting a plane with 3×3×3 neighboring points and calculating the angle between the plane and the horizontal plane); if θ>0.15, calculate the constraint loss L. s =|θ-0.15| / 0.15; If within the range, L s =0; ④ Total constraint loss L const =0.4×L r +0.3×L m +0.3×L s (Weights are assigned based on the importance of constraints).

[0086] 5.2 Fine-tuning of model parameters:

[0087] ①The total constraint loss L const Compared with the total loss L in step 4 total Combining these, we obtain the final optimized loss L. final =L total +0.5×L const (Constraint loss weight = 0.5, balancing reconstruction accuracy and physical consistency);

[0088] ② Adaptive momentum estimation (Adam) optimizer is adopted, and the parameters are adjusted as follows: learning rate = 0.00005 (lower than the fusion optimization stage to avoid destroying the existing optimization results), β1 = 0.95, β2 = 0.999 (β1, β2 are the momentum phase hyperparameters of the Adam optimizer);

[0089] ③ Iterative calibration: Calculate L every 1000 iterations. const If L const If the value is less than 0.005 (constraint satisfied), the constraint loss weight is reduced to 0.2; Termination condition: number of iterations ≥ 5000, and L final ≤0.01, or L const ≤0.003 (physical constraints are fully satisfied);

[0090] ④ Result verification: Randomly select 100 radial cross sections of the pipe and measure the cross section diameter. The error must be ≤ ±2mm. Randomly select 50 pipe wall points and measure the reflectivity. The error must be ≤ ±0.05. Randomly select 30 siltation points and measure the surface slope. The error must be ≤ ±0.02. If all conditions are met, output the refined and optimized fusion model (including Gaussian distribution set and NeRF network parameters).

[0091] Step 6: Generation and Visualization of 3D High-Density Model

[0092] 6.1 High-density 3D model generation:

[0093] ① Voxel Sampling: Read the optimized fusion model and set the voxel resolution to 0.5mm (10 times smaller than the original reduced-dimensional voxels based on pipeline detection accuracy requirements); divide the pipeline space (x: 0m-500m, y: -400mm-400mm, z: -400mm-400mm) into a 0.5mm×0.5mm×0.5mm voxel mesh, totaling approximately 1.6×10 12 voxel; for each voxel center (x v ,y v ,z v The density ρ and color c are calculated using a fusion model: ρ = 0.3 × ρ gauss +0.7×ρ nerf (ρ) gauss For Gaussian sputtering model output, ρ nerf (This is the output of the NeRF model, with weights biased towards high-precision NeRF); c = 0.3 × c gauss +0.7×c nerf ① Remove voxels with ρ < 0.1 (invalid empty voxels), retain valid voxels, and generate a high-density voxel model; ② Point cloud model generation: For each valid voxel, take the voxel center as a point cloud point, and store the three-dimensional coordinates (x, y, z) of the point. v ,y v ,z v ) and color (c r ,c g ,c b Generate a high-density point cloud model (point cloud density ≥ 1000 points / cm²). 2 ); ③ Mesh model generation: Using the "Poisson reconstruction algorithm", input a high-density point cloud, set the reconstruction depth = 12 (to control mesh details), and the sampling density = 1.5 (1.5 times the point cloud sampling density); smooth the reconstructed mesh (Laplace smoothing, iterations = 5, smoothing factor = 0.1), remove non-manifold meshes (meshes with ≠ 2 sides), and generate a closed mesh model; ④ Result verification: calculate the average distance error of the point cloud model (≤ 0.1 mm) and the number of triangles in the mesh model (≥ 1 × 10). 7 If the conditions are met, proceed to the next step.

[0094] 6.2 Association between 3D Models and GIS Data

[0095] ① Read the urban pipeline network GIS data (including the geographical coordinates, mileage, station number, pipe diameter, and laying year of the pipeline), and establish the mapping relationship between the three-dimensional model coordinate system and the geographic coordinate system by using the coordinates of the pipeline's starting end recorded by the GPS positioning module (such as longitude E116.4°, latitude N39.9°): Geographic coordinates (Lon, Lat, Alt) → Model coordinates (x, y, z): x = (Lon - Lon0) × 111319.9 (distance corresponding to each degree of longitude), y = (Lat - Lat0) × 111319.9 × cos(Lat0), z = Alt - Alt0; where (Lon0, Lat0, Alt0) are the geographic coordinates of the pipeline's starting end, and Lon, Lat, and Alt are the core parameters for locating the spatial position of the pipeline in the geographic coordinate system, which are abbreviations for longitude, latitude, and altitude, respectively; ② Mark the 3D model with mileage markers: Mark a mileage marker every 1m along the x-axis of the pipeline (e.g., K0+000, K0+001), and associate it with the pipeline attributes (material, pipe diameter, laying year) in the GIS; ③ Automatically identify siltation areas and defects (cracks, corrosion): Siltation areas: Areas where the z-coordinate (radial) of the point cloud points is >800mm + 5mm (5mm outside the pipe wall) are marked as "siltation", and the siltation thickness (maximum z-coordinate -800mm) and range (start and end positions of the x-axis) are recorded; Crack defects: In the mesh model, areas where the angle between the normal vectors of adjacent triangles is >30° and the crack width is >0.2mm are marked as "crack", and the crack length, width, and location are recorded; ④ Store the annotation information as an attribute file (.json), associate it with the 3D model (.ply point cloud, .obj mesh), and store it in the comprehensive management database.

[0096] 6.3 Visualization Output and Data Storage:

[0097] ① Visualization module startup conditions: After the 3D model is associated with GIS data, the device automatically responds to access requests from both C / S and B / S terminals; ② Multi-terminal access processing: C / S terminal (3D imaging operation terminal): Outputs a complete high-density point cloud / mesh model, supports real-time rendering (frame rate ≥ 30fps), provides multi-angle browsing (mouse drag rotation, scroll wheel zoom), sectioning function (custom sectioning planes, such as x=100m section, y=0mm section), and annotation editing (manually add / modify defect annotations); B / S terminal (urban pipeline management department monitoring workstation): Compresses the 3D model to a low-precision version (point cloud density = 100 points / cm²). 2 Number of mesh triangles = 1 × 10 6(1) Rendering on the browser side using WebGL technology, supporting model browsing and defect query (filtered by station number and defect type); ③ Data export function: Supports users to export the following data: 3D model files (.ply / .obj format); ④ Data storage: Synchronizes 3D model data, annotation data, and GIS associated data to the comprehensive management database server (MySQL database, partitioned storage: partitioned by pipeline ID, with each 100m pipeline as a data block), supporting real-time data updates (new inspection data overwrites old data) and queries (queried by pipeline ID, station number, and defect type).

[0098] A second aspect of the present invention provides a three-dimensional imaging system for drainage pipes based on the fusion reconstruction of neural radiation fields and Gaussian sputtering, employing the above-mentioned three-dimensional imaging method and including the following subsystems:

[0099] 1. Pipeline multi-source data acquisition subsystem, including: unmanned pipeline measurement device (UPM), unmanned aerial vehicle (UAV), radar detection vehicle based on IoT communication, high-definition video surveillance robot, lidar sensor, inertial measurement unit (IMU), GPS positioning module, and supporting data acquisition workstation. This system is used to simultaneously acquire images, point clouds, attitude, and basic parameter data inside the pipeline, and to achieve real-time data transmission through wired or wireless communication technologies.

[0100] 2. The data preprocessing subsystem includes: a data denoising module, a data registration module, and a data dimensionality reduction module. These modules employ techniques such as statistical filtering and voxel grid downsampling to remove noise, spatially align, and compress the dimensions of multi-source data, outputting standardized data for subsequent modeling.

[0101] 3. The fusion reconstruction subsystem includes: a Gaussian sputtering initialization module, a multi-scale NeRF network module, and a joint optimization module. The Gaussian sputtering initialization module is used to realize sparse 3D structure reconstruction; the multi-scale NeRF network module is used to complete high-precision radiation field learning; and the joint optimization module achieves synergistic optimization of the two by constructing a joint loss function.

[0102] 4. The model optimization subsystem includes: a physical constraint embedding module and a parameter calibration module. The physical constraint embedding module is used to introduce constraints such as pipe diameter, material, and siltation boundary; the parameter calibration module uses an adaptive optimizer to fine-tune the model parameters to improve reconstruction accuracy and robustness.

[0103] 5. The 3D imaging visualization subsystem includes: a 3D model generation module, a GIS association module, and a visualization display module. The 3D model generation module outputs high-density point cloud and mesh models; the GIS association module binds the model to geographic information; and the visualization display module provides functions such as browsing, sectioning, and annotation of the 3D model.

[0104] The system architecture is divided into a data access control layer, an access control layer, and an application layer.

[0105] Data access control layer: This includes a comprehensive management database server that connects to the standard data interface of the GIS server software, providing data storage, access control, and calling interfaces, and supporting real-time updates and queries of model data and collected data.

[0106] Access control layer: Hosted by GIS server, it provides GIS services, business query operation services, computing services, and database access relay access control services; it provides complex computing services such as fusion modeling and model optimization through C / S business system, and provides services such as visualization display and data query through B / S business system.

[0107] Application layer: Adopting a hybrid C / S and B / S architecture, it includes a 3D imaging operation terminal and a monitoring workstation for urban pipeline management departments, supporting functions such as model generation parameter configuration, 3D visualization browsing, and diagnostic data export.

[0108] The hardware support platform includes: a 3D imaging computing terminal (CPU: Intel Xeon Gold 6248, GPU: NVIDIA RTX 4090 24GB, RAM: 64GB, operating system: Ubuntu 22.04 64bit), a comprehensive management database server, an urban pipeline network GIS server, a data acquisition workstation, a monitoring and visualization workstation, and various detection instruments and sensors connected via the Internet of Things.

[0109] Example 2: Experimental verification of three-dimensional imaging of urban drainage pipes using the method described in the example.

[0110] 1. Experimental Environment and Data Acquisition

[0111] A DN800 concrete drainage pipe (500m in length, 0.003 slope) in a certain city was selected as the experimental object, and data was collected using the multi-source data acquisition subsystem of the system of this invention:

[0112] High-definition camera: 4K resolution, 30fps frame rate, 24mm lens focal length;

[0113] IMU sampling frequency 100Hz, GPS positioning accuracy ±0.5m;

[0114] A set of multi-view data was collected every 5m along the pipeline axis, for a total of 100 sets of valid data.

[0115] 2. Data Preprocessing

[0116] Point cloud data is generated by processing diverse standardized photos captured by the camera.

[0117] Statistical filtering (15 neighboring points, standard deviation threshold 2.0) was used to remove laser point cloud noise;

[0118] Voxel mesh downsampling (5mm voxel size) reduces point cloud data volume by 60% while preserving key structural features.

[0119] 3. Integration of Reconstruction and Model Optimization

[0120] Gaussian sputtering initialization: set the number of Gaussian elements to 50,000, the number of iterations to 100, and the photometric loss converges to 0.02;

[0121] Multi-scale NeRF network: It adopts 4 fully connected layers (hidden layer dimension 256), the multi-scale feature fusion scale is 1×2×4×, and the number of training iterations is 20,000;

[0122] Joint loss function weighting: photometric loss weight 0.6, geometric consistency loss weight 0.3, smoothness loss weight 0.1;

[0123] Physical constraint parameters: pipe diameter constraint range 795mm-805mm, concrete material reflectivity 0.3-0.5, siltation boundary slope constraint ≤0.15.

[0124] 4. Experimental Results and Performance Comparison

[0125] The method of this invention is compared with single NeRF, single Gaussian sputtering, and traditional LiDAR reconstruction methods. The evaluation metrics include mean absolute error (MAE), root mean square error (RMSE), training time, and detail restoration accuracy.

[0126]

[0127] The results show that the method of the present invention is superior to existing methods in reconstruction accuracy (MAE reduction of 64.6%-76.8%), training efficiency (75.9% shorter than NeRF alone), and detail restoration (5%-24.4% higher), and can accurately restore the texture accumulation morphology and micro-crack defects of pipe walls.

Claims

1. A three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation fields and Gaussian sputtering, characterized in that: include: Point cloud image data is generated by calculating and training the collected standardized multi-view images; The point cloud data is initialized and reconstructed using a Gaussian sputtering model to obtain a sparse three-dimensional structure model. The neural radiation field network is initialized using a sparse 3D structure model. Geometric features of Gaussian elements are extracted from the sparse 3D structure model and constructed as a feature vector F. gauss F gauss Including the spatial distribution histogram and color distribution of Gaussian elements, density field network MLP1 and color field network MLP2 of the neural radiation field network are constructed, and the feature vector F of the Gaussian elements is used to construct the neural radiation field network. gauss As a priori, the weights of MLP1 and MLP2 are initialized. The weights of the first layer of MLP1 are initialized according to the spatial distribution histogram of Gaussian elements, and the weights of the output layer of MLP2 are initialized according to the color distribution of Gaussian elements. The pipeline space is uniformly sampled, and the initial density field ρ0 and initial color field c0 are calculated for each sampling point through MLP1 and MLP2 and stored as the initial model of the neural radiation field network. Based on the initial model of the neural radiation field network, we perform neural radiation field network modeling, training, and fusion optimization. Rendered images are generated from the same viewpoint using a Gaussian sputtering model and a neural radiation field network, and the two rendered images are fused to obtain a three-dimensional imaging model. The specific steps include: Step 1: Multi-source pipeline scene data acquisition and calibration. Simultaneously acquire multi-view image data, equipment posture data and pipeline basic parameters inside the drainage pipeline, complete the calibration of equipment internal and external parameters, and unify them to a preset three-dimensional coordinate system with the pipeline axis as the X-axis. Step 2: Calculate and train the collected standardized multi-view images to generate point cloud data and reduce noise. Perform feature point matching on the multi-view images to generate descriptors. Use the nearest neighbor ratio method and RANSAC algorithm to perform feature matching and outlier removal to form an initial sparse point cloud. Calculate the three-dimensional coordinates of the feature points. Step 3: Gaussian sputtering initialization and sparse 3D structure reconstruction. Gaussian sputtering initialization optimizes the Gaussian distribution parameters by minimizing photometric loss, adaptively selects effective Gaussian elements, and outputs a sparse 3D structure model containing a set of effective Gaussian elements and their corresponding spatial location information. Step 4: Model the fusion of neural radiation field and Gaussian sputtering. Initialize the neural radiation field using a Gaussian structure and construct the joint loss function L. total Iterative optimization, joint loss function L total This includes photometric loss, geometric consistency loss, and smoothness loss; Step 5: Refine the model by combining pipeline physical constraints. Embed the pipe diameter constraint, material reflectivity constraint, and siltation boundary constraint into the joint loss function to refine the parameters of the fusion model and ensure that the fusion model conforms to the physical characteristics of the pipeline. Step 6: Generate a high-density 3D model and associate it with GIS. Perform high-resolution voxel sampling on the refined and optimized fusion model to generate a high-density point cloud and grid model. Associate it with geographic information system data and achieve multi-terminal visualization output.

2. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering as described in claim 1, characterized in that: When generating the rendered image, the geometric deviation between the Gaussian sputtering model and the neural radiation field network is calculated, and a joint loss function L is constructed. total Iterative optimization involves simultaneously calculating the gradients of the Gaussian distribution parameters and the weights of the neural radiation field network, and then fusing the rendered images generated by the Gaussian sputtering model and the neural radiation field network according to their weights.

3. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering as described in claim 2, characterized in that: After fusing the two rendered images to obtain a 3D image, physical constraints are embedded into the 3D image for constraint loss calculation. These physical constraints include pipe diameter, material, and siltation boundary. The physical constraints are then combined with the joint loss function L. total The final optimized loss is obtained by combining the weights, the parameters are adjusted and iterative optimization is performed, and the optimized 3D imaging model is output.

4. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering as described in claim 3, characterized in that: By linking the 3D imaging model with GIS data, establishing a mapping relationship between the 3D model coordinate system and the geographic coordinate system, mileage annotation is performed on the 3D imaging model to identify siltation areas and defects, and visualization output and data storage are achieved.

5. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering according to claim 4, characterized in that... Step 3 includes: S3.1 Initialization of Gaussian Distribution Set: Randomly sample points from the dimensionality-reduced point cloud as the initial centers of Gaussian elements; calculate initial weights based on the pixel brightness and depth confidence of the sampled points; set the initial scale based on the variance of the neighboring point cloud; construct a local coordinate system based on the sampled point normal vector and the pipeline axis to determine the initial rotation parameters; and set a uniform initial transparency; determine the Gaussian center, establish the mapping relationship between Gaussian elements and point cloud sampled points, calculate the initial Gaussian parameters, and generate the initial Gaussian distribution set; S3.2 Obtain the intrinsic and extrinsic parameters of the multi-view images, generate virtual camera rays for each viewpoint, and sample K points along the ray direction. For each sampled point, calculate its density ρ and color c using Gaussian element parameters. Calculate the rendered image using ray integration, and calculate the photometric loss L between the rendered image and the original image. photo Update the parameters and iteratively optimize the Gaussian distribution set; S3.3 Effective Gaussian Element Filtering: Set an adaptive threshold to remove redundant Gaussian elements with too low weights and output a sparse 3D structure model.

6. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering according to claim 5, characterized in that... Step 4 includes: S4.1 Gaussian Structure-Guided Neural Radiation Field Initialization: Based on the sparse three-dimensional structure model obtained in step 3, extract the geometric features of Gaussian elements to construct the feature vector F. gauss The density field network MLP1 and color field network MLP2 of the neural radiation field are constructed, and the eigenvectors F of Gaussian elements are... gauss As a priori, the weights of MLP1 and MLP2 are initialized. The weights of the first layer of MLP1 are initialized according to the spatial distribution histogram of Gaussian elements. The output layer of MLP2 is initialized according to the color distribution of Gaussian elements. The pipeline space is uniformly sampled. The initial density field ρ0 and color field c0 are calculated through MLP1 and MLP2 and stored as the initial model of the neural radiation field network. S4.2 Construction and Feature Fusion of Multi-Scale Neural Radiation Field Network: Based on the initial model of the neural radiation field network, neural radiation field modeling, training, and feature fusion optimization are performed. The multi-scale neural radiation field network contains 4 fully connected layers with ReLU+Sigmoid activation function. gauss The multi-dimensional features fused with multi-scale are concatenated and input into MLP1 and MLP2. The image is then divided into training and validation sets for iterative training. S4.3 Fusion Optimization Iteration: For the same viewpoint, render images I are generated using both a Gaussian sputtering model and a neural radiation field network. gauss and I nerf ; Calculate I gauss Photometric loss L photo Calculate I gauss and I nerf Geometric deviation L geo Calculate I nerf gradient smoothness L smooth Joint loss function L total =0.6×L photo +0.3×L geo +0.1×L smooth The Adam optimizer is used to simultaneously calculate the gradients of the Gaussian distribution parameters and the neural radiation field network weights, update the parameters, and perform iterative optimization to calculate the fused rendered image I. fusion =0.4×I gauss +0.6×I nerf .

7. The three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation field and Gaussian sputtering as described in claim 6, characterized in that... Step 5 includes: S5.1 Pipe diameter constraint, material reflectivity constraint, and siltation boundary constraint are embedded in the loss function: Pipe diameter constraint: Based on the pipe design diameter, an allowable range of radial distance is set. Gaussian elements or sampling points that exceed the range are penalized, and the pipe diameter constraint loss L is calculated. r ; Material constraints: Based on the preset surface reflectivity range of the pipe material, a penalty is applied to cases where the reflectivity calculated for the rendered color exceeds this range, and the material constraint loss L is calculated. m ; Siltation boundary constraint: Based on fluid dynamics principles, a maximum slope threshold for the siltation surface is set. A penalty is applied to areas identified as siltation regions where the surface slope exceeds this threshold, and the siltation boundary constraint loss L is calculated. m ; Total constraint loss L const =0.4×L r +0.3×L m +0.3×L s ; S5.2 Refined Optimization: Final Optimization Loss L final =L total +0.5×L const An adaptive momentum estimation optimizer is used to adjust parameters and perform iterative calibration. After satisfying the iterative conditions, a refined optimized fusion model is obtained.

8. A three-dimensional imaging system for drainage pipes based on the fusion reconstruction of neural radiation fields and Gaussian sputtering, used to perform the three-dimensional imaging method for drainage pipes based on the fusion reconstruction of neural radiation fields and Gaussian sputtering as described in any one of claims 1-7, characterized in that: The system includes a pipeline multi-source data acquisition subsystem, a data preprocessing subsystem, a fusion and reconstruction subsystem, a model optimization subsystem, and a 3D imaging visualization subsystem. The system structure is divided into a data access control layer, an access control layer, and an application layer. The hardware support platform includes a computing terminal, a database server, a GIS server, and detection equipment. The fusion and reconstruction subsystem includes a Gaussian sputtering initialization module, a multi-scale NeRF network module, and a joint optimization module to achieve synergistic optimization of sparse structures and high-precision radiation fields. The application layer adopts a hybrid C / S and B / S architecture, supporting 3D model browsing, sectioning and annotation, and GIS-related query functions.

Images