Adaptive 4D gaussian splatting high-precision three-dimensional reconstruction system and method

Abstract

The application discloses a self-adaptive 4D Gaussian splashing high-precision three-dimensional reconstruction system and method, and belongs to the technical field of computer graphics processing. In order to solve the problems of realizing high-precision automatic registration of multi-source heterogeneous data and improving the calculation efficiency, the application comprises the following steps: collecting multi-source heterogeneous data; constructing a multi-modal fusion registration method, including realizing spatial distribution geometric rough registration by combining satellite image data and low-altitude oblique photography data, realizing photometric fine registration by fusing low-altitude laser radar point cloud data and ground acquisition vehicle laser radar point cloud data, and establishing semantic feature auxiliary registration to obtain registered multi-source data; initializing a 4D Gaussian primitive and performing self-adaptive splashing reconstruction to obtain an optimized 4D Gaussian splashing model; designing a cloud-edge collaborative computing architecture for the 4D Gaussian splashing reconstruction, performing distributed parallel processing on the obtained optimized 4D Gaussian splashing model; performing quality evaluation and self-adaptive optimization, and outputting a final self-adaptive 4D Gaussian splashing model.

Description

An Adaptive 4D Gaussian Splash High-Precision 3D Reconstruction System and Method Technical Field

[0001] This invention belongs to the field of computer graphics processing technology, specifically relating to an adaptive 4D Gaussian splash high-precision 3D reconstruction system and method. Background Technology

[0002] Current 3D reconstruction technologies face numerous challenges when handling large-scale, multi-source heterogeneous data fusion, including: insufficient data fusion accuracy (traditional methods struggle to effectively fuse multi-source data such as satellite imagery, aerial photography, and LiDAR point clouds, with registration errors typically ranging from 5-30 centimeters, failing to meet high-precision application requirements); low computational efficiency (existing reconstruction algorithms have high computational complexity, taking days to weeks to process large-scale city-level scenes, making it difficult to meet real-time or near-real-time application needs); weak dynamic scene processing capabilities (traditional static reconstruction methods cannot effectively handle dynamic scenes containing moving targets, resulting in artifacts and inconsistencies in reconstruction results); poor scalability (the lack of an effective distributed processing architecture makes it difficult to fully utilize cloud and edge computing resources, with processing power limited by single-machine performance); and low automation (requiring significant manual intervention for parameter adjustment and quality control, hindering fully automated industrial applications). In recent years, emerging technologies such as Neural Radiation Field (NeRF) and Gaussian Splatting have brought breakthroughs to 3D reconstruction, but problems such as long training times, high memory consumption, and insufficient multi-source data fusion capabilities still exist. Summary of the Invention

[0003] The problem this invention aims to solve is to achieve high-precision automatic registration of multi-source heterogeneous data, control the registration error to the millimeter level and improve computational efficiency, and proposes an adaptive 4D Gaussian splash high-precision three-dimensional reconstruction system and method.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An adaptive 4D Gaussian splash high-precision 3D reconstruction method includes the following steps:

[0006] S1. Collect multi-source heterogeneous data, including satellite imagery data, low-altitude oblique photography data, low-altitude lidar point cloud data, low-altitude video stream data, ground acquisition vehicle lidar point cloud data, and ground acquisition vehicle video stream data, and then perform preprocessing, including data source standardization processing and multi-source device spatiotemporal reference alignment processing.

[0007] S2. For the multi-source heterogeneous data preprocessed in step S1, construct a multi-modal fusion registration method, including combining satellite imagery data and low-altitude oblique photography data to achieve coarse spatial distribution geometric registration, fusing low-altitude lidar point cloud data and ground acquisition vehicle lidar point cloud data to achieve fine photometric registration, and establishing semantic feature-assisted registration to obtain registered multi-source data;

[0008] S3. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives and perform adaptive splash reconstruction to obtain an optimized 4D Gaussian splash model;

[0009] S4. For the optimized 4D Gaussian splash model obtained in step S3, design a cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction, and perform distributed parallel processing on the optimized 4D Gaussian splash model.

[0010] S5. Perform quality assessment and adaptive optimization on the optimized 4D Gaussian splash model obtained after distributed parallel processing in step S4, and output the final adaptive 4D Gaussian splash model.

[0011] Furthermore, the specific implementation method of step S1 includes the following steps:

[0012] S1.1. Perform data source standardization processing on the acquired multi-source heterogeneous data, including radiometric and geometric correction on the acquired satellite image data, color balance and radial distortion correction on the low-altitude oblique photography data; noise filtering and coordinate transformation on the low-altitude lidar point cloud data and the ground acquisition vehicle lidar point cloud data; and optical flow stabilization and keyframe extraction on the low-altitude video stream data and the ground acquisition vehicle video stream data.

[0013] S1.2. Multi-source device spatiotemporal reference alignment processing;

[0014] S1.2.1. Time base unification and synchronization, including GPS time base conversion, multi-source device time synchronization, and timestamp interpolation alignment;

[0015] S1.2.2. Multi-source positioning information fusion and alignment, including GPS / RTK location fusion, BeiDou / GPS dual-system fusion, and inertial navigation-assisted alignment;

[0016] S1.2.3. Equipment attitude alignment correction, including constructing a standardized attitude matrix, performing Euler angle transformation and quaternion interpolation alignment;

[0017] S1.2.4. Perform spatiotemporal registration verification on the multi-source data processed in steps S1.2.1-S1.2.3, evaluate the time synchronization accuracy and spatial alignment accuracy, and set a comprehensive alignment quality index.

[0018] Furthermore, the specific implementation method of step S2 includes the following steps:

[0019] S2.1. Construct a geometric coarse registration method, including GPS / RTK pose alignment, SIFT+SuperGlue fusion, and RANSAC robust transform estimation;

[0020] S2.2. Construct a photometric registration method, including using bundle adjustment optimization for low-altitude oblique photography and using the iterative nearest point algorithm (ICP) to optimize lidar point cloud data;

[0021] S2.3. Construct a semantic feature-assisted registration method and set semantic consistency constraints. The expression is:

[0022]

[0023] in, These are the i-th semantic tag and the j-th semantic tag, respectively. , Let i and j be the semantic feature vectors, respectively. Spatial weights, For indicator functions;

[0024] Perform multimodal fusion registration.

[0025]

[0026] in, To achieve the target value for multimodal fusion registration accuracy, For geometric registration error, This is due to photometric registration error. , and These are the balance weight parameters for geometric error, photometric error, and semantic constraints, respectively. .

[0027] Furthermore, the specific implementation method of step S3 includes the following steps:

[0028] S3.1. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives;

[0029] S3.1.1. Initialize 3D Gaussian primitives from LiDAR point cloud data and perform initial Gaussian function seeding, including setting initial position, covariance, color, and opacity;

[0030] S3.1.2. The 3D Gaussian primitives initialized in step S3.1.1 are expanded to 4D spatiotemporal Gaussian primitives, and time parameters are added to parameterize the initial position and opacity over time.

[0031] S3.1.3. Construct an adaptive initialization strategy for the 4D spatiotemporal high-order primitives obtained in step S3.1.2;

[0032] The expression for the density-adaptive initialization strategy is:

[0033]

[0034] in, The density value obtained after adaptation. Based on density, For adaptive coefficients, , For image Laplacian operator;

[0035] The expression for the scale-adaptive initialization strategy is:

[0036]

[0037] in, For the Gaussian scale, The distance to the i-th nearest neighbor. The average distance, The global average neighbor distance is used as the normalization benchmark. For the scaling adjustment index, set ;

[0038] S3.2. For the 4D Gaussian primitives initialized in step S3.1, perform adaptive differentiable rendering optimization iteration for adaptive splash reconstruction. During the iterative process, the density is adaptively controlled by the adaptive splitting criterion and the adaptive deletion criterion to obtain the optimized 4D Gaussian splash model.

[0039] Furthermore, step S4, which designs the specific implementation method of the cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction, includes the following steps:

[0040] S4.1. Perform edge-end intelligent preprocessing on the optimized 4D Gaussian splash model obtained in step S3;

[0041] First, SLAM state estimation and keyframe pose optimization are performed;

[0042] Then, information entropy evaluation is used to construct keyframe selection criteria, and adaptive keyframes are selected.

[0043] Finally, multimodal data compression and point cloud compression optimization are performed to obtain a 4D Gaussian splash model after intelligent preprocessing at the edge.

[0044] S4.2. Perform parallel depth reconstruction of the 4D Gaussian splash model obtained from the edge-end intelligent preprocessing in step S4.2.1 in the cloud.

[0045] S4.2.1. Calculate load assessment and perform intelligent load balancing allocation. The expression is:

[0046]

[0047] in, Let i be the load evaluation function for node i. Let i be the amount of data at node i. Let be the computational complexity of node i. Let i be the memory usage of node i. , , These are the weighting coefficients for the data volume of node i, the computational complexity of node i, and the memory usage of node i, respectively.

[0048] Then, task allocation optimization is performed, expressed as follows:

[0049]

[0050] in, To optimize the target value for task scheduling, Let be the total time cost of the i-th node. Let be the processing time for node i. For communication overhead, For communication weights;

[0051] S4.2.2. Utilize data parallelism strategies for parallel GPU cluster processing;

[0052] S4.2.3. Distributed result fusion is performed on the 4D Gaussian splash model after intelligent preprocessing at the edge after parallel GPU cluster processing, including spatial partition fusion and overlapping region processing, to obtain the optimized 4D Gaussian splash model after distributed parallel processing.

[0053] Furthermore, the specific implementation method of step S5 includes the following steps:

[0054] S5.1. Construct a geometric accuracy evaluation method, including calculating point cloud distance error, normal vector consistency, and curvature fidelity;

[0055] S5.2. Construct a texture fidelity evaluation method, including calculating the structural similarity index, perceptual hash similarity, and multi-scale texture consistency;

[0056] S5.3. Construct an integrity assessment method, including calculating volume overlap rate, surface coverage, and detail retention;

[0057] S5.4. Construct a time consistency assessment method, including calculating motion continuity, deformation consistency, and spatiotemporal smoothness;

[0058] S5.5. Construct a comprehensive quality assessment index, perform a weighted comprehensive assessment, and obtain the expression:

[0059]

[0060] in, As a weighted comprehensive evaluation indicator, This refers to the point cloud distance error. For multi-scale texture consistency, For surface coverage, For the continuity of motion, , , , The weights of point cloud distance error, multi-scale texture consistency, surface coverage, and motion continuity in the comprehensive quality evaluation index are respectively:

[0061] S5.6. Based on the weighted comprehensive evaluation index obtained in step S5.5, perform adaptive weight adjustment and confidence interval evaluation to complete adaptive optimization and output the final adaptive 4D Gaussian splash model.

[0062] A system for predicting and analyzing the energy-saving effect of a drone's climb process includes a processor, a memory, and a computer program stored in the memory and run on the processor. When the computer program runs, it implements the steps of the method for predicting and analyzing the energy-saving effect of a drone's climb process.

[0063] The beneficial effects of this invention are:

[0064] This invention presents an adaptive 4D Gaussian splash high-precision 3D reconstruction method, proposing a three-level registration strategy of coarse registration, fine registration, and global optimization. Combining geometric feature matching, photometric consistency constraints, and semantic feature assistance, it achieves millimeter-centimeter level precision registration of multi-source heterogeneous data. This solves the problems of insufficient accuracy and poor robustness of traditional registration methods. Registration accuracy is improved by 5-10 times, laying the foundation for subsequent high-precision reconstruction.

[0065] This invention discloses an adaptive 4D Gaussian splash high-precision 3D reconstruction method that extends Gaussian splash technology to the 4D spatiotemporal domain, introducing a time dimension parameter and an adaptive density control mechanism to support continuous reconstruction of dynamic scenes. It overcomes the limitations of traditional static reconstruction, enabling the handling of complex dynamic scenes containing moving targets, and significantly improving reconstruction quality and efficiency.

[0066] This invention discloses an adaptive 4D Gaussian splash high-precision 3D reconstruction method. It designs a three-layer collaborative architecture of edge preprocessing, cloud-based deep reconstruction, and result distribution. Combining intelligent load balancing based on computing power type and load, and parallel task computing, it achieves dynamic scheduling and optimization of computing resources. By fully utilizing cloud and edge computing resources, processing capacity can be dynamically expanded according to data scale, improving processing efficiency by 10-50 times.

[0067] This invention discloses an adaptive 4D Gaussian splashing high-precision 3D reconstruction method. It optimizes the parallel characteristics of the Gaussian splashing algorithm, designs a CUDA-specific kernel and memory management strategy, and achieves efficient utilization of GPU computing resources. This significantly improves computational efficiency, increasing GPU utilization to over 90%, providing technical support for large-scale real-time 3D reconstruction applications.

[0068] This invention discloses an adaptive 4D Gaussian splash high-precision 3D reconstruction method. It establishes a multi-dimensional quality assessment system and an adaptive parameter optimization mechanism, achieving a fully automated processing flow from data acquisition to model output. This significantly reduces the need for manual intervention, increasing the automation level by over 90%, and meeting the reliability and stability requirements of industrial applications. Attached Figure Description

[0069] Figure 1 is a flowchart of an adaptive 4D Gaussian splash high-precision 3D reconstruction method according to the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0071] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0072] To further understand the invention's content, features, and effects, the following specific embodiments are provided, and detailed descriptions are given in conjunction with Figure 1:

[0073] Example 1:

[0074] An adaptive 4D Gaussian splash high-precision 3D reconstruction method includes the following steps:

[0075] S1. Collect multi-source heterogeneous data, including satellite imagery data, low-altitude oblique photography data, low-altitude lidar point cloud data, low-altitude video stream data, ground acquisition vehicle lidar point cloud data, and ground acquisition vehicle video stream data, and then perform preprocessing, including data source standardization processing and multi-source device spatiotemporal reference alignment processing.

[0076] Furthermore, the specific implementation method of step S1 includes the following steps:

[0077] S1.1. Perform data source standardization processing on the acquired multi-source heterogeneous data, including radiometric and geometric correction on the acquired satellite image data, color balance and radial distortion correction on the low-altitude oblique photography data; noise filtering and coordinate transformation on the low-altitude lidar point cloud data and the ground acquisition vehicle lidar point cloud data; and optical flow stabilization and keyframe extraction on the low-altitude video stream data and the ground acquisition vehicle video stream data.

[0078] Furthermore, the formula for radiometric correction of the acquired satellite image data is as follows: ,in, To correct the radiance, Original radiance, Dark current, Atmospheric transmittance, To observe the zenith angle, The distance between the Earth and the Sun. Solar irradiance, The solar zenith angle;

[0079] The formula for geometric correction is: ,in, The homography matrix is ​​3×3, and it is solved using the least squares method of ground control points.

[0080] The formula for color balance is: ,in, , The formula for radial distortion correction is: ,in, , Radial distortion coefficient;

[0081] The formula for statistical filtering of noise is: , The filtering conditions are: ,in This is the threshold coefficient;

[0082] The formula for the coordinate transformation matrix is: ,in, Let be a rotation matrix. It is a translation vector;

[0083] The formula for optical flow stabilization is: ,in, For optical flow vector, For regularization parameters;

[0084] The formula for keyframe extraction is: ,when The time frame is selected as the keyframe, where, An adaptive threshold;

[0085] S1.2. Multi-source device spatiotemporal reference alignment processing;

[0086] S1.2.1. Time base unification and synchronization, including GPS time base conversion, multi-source device time synchronization, and timestamp interpolation alignment;

[0087] Furthermore, the formula for GPS time base conversion is: ,in, For leap second correction, The time offset is determined by GPS; the formula for multi-device time synchronization is: ,in, Let i be the time offset of device i. The clock drift rate is used; the formula for timestamp interpolation alignment is: , in, The sampling interval is... For interpolation position parameters;

[0088] S1.2.2. Multi-source positioning information fusion and alignment, including GPS / RTK location fusion, BeiDou / GPS dual-system fusion, and inertial navigation-assisted alignment;

[0089] Furthermore, the formula for GPS / RTK location fusion is: ,in, These are the positioning accuracy of GPS and RTK, respectively; the formula for BeiDou / GPS dual-system fusion is: The formula for calculating the weights is: The formula for inertial navigation-assisted alignment is: ,in, Given the initial position and velocity, It is the acceleration vector;

[0090] S1.2.3. Equipment attitude alignment correction, including constructing a standardized attitude matrix, performing Euler angle transformation and quaternion interpolation alignment;

[0091] Furthermore, the formula for normalizing the attitude matrix is: ,in, The rotation matrix is ​​calibrated for the equipment; the formula for Euler angle transformation is: ,in, For the elements of the rotation matrix; the formula for quaternion interpolation alignment is: ,in, The angle between quaternions;

[0092] S1.2.4. Perform spatiotemporal registration verification on the multi-source data processed in steps S1.2.1-S1.2.3, evaluate the time synchronization accuracy and spatial alignment accuracy, and set a comprehensive alignment quality index;

[0093] Time synchronization accuracy The expression for evaluation is:

[0094]

[0095] in, Let i be the reference timestamp for the i-th time. Let N be the synchronization timestamp of the i-th time, and N be the total number of times.

[0096] Spatial alignment accuracy The expression for evaluation is:

[0097]

[0098] in, Let j be the position of the j-th reference space. Let j be the position of the alignment space. This represents the total number of spaces.

[0099] Comprehensive alignment quality indicators The expression is:

[0100]

[0101] in, For time weighting coefficients, Spatial weighting coefficient, This is the time precision threshold. Set the spatial precision threshold. , ms, mm.

[0102] Furthermore, the key parameters are explained below:

[0103] Time synchronization accuracy: Multi-device time synchronization error < 1ms, clock drift rate < 10^{-6} / s, interpolation time accuracy < 0.1ms; Positioning system fusion: GPS accuracy 5-50m, RTK accuracy 1-2cm, BeiDou / GPS dual system improves reliability by 30%; Equipment calibration parameters: Camera intrinsic parameter calibration accuracy < 0.1 pixels, IMU calibration error < 0.01°, LiDAR calibration accuracy < 1mm; Verification index thresholds: Time synchronization accuracy < 1ms, spatial alignment accuracy < 20cm, overall alignment quality > 0.95.

[0104] S2. For the multi-source heterogeneous data preprocessed in step S1, construct a multi-modal fusion registration method, including combining satellite imagery data and low-altitude oblique photography data to achieve coarse spatial distribution geometric registration, fusing low-altitude lidar point cloud data and ground acquisition vehicle lidar point cloud data to achieve fine photometric registration, and establishing semantic feature-assisted registration to obtain registered multi-source data;

[0105] Furthermore, the specific implementation method of step S2 includes the following steps:

[0106] S2.1. Construct a geometric coarse registration method, including GPS / RTK pose alignment, SIFT+SuperGlue fusion, and RANSAC robust transform estimation;

[0107] Furthermore, the formula for GPS / RTK pose alignment is: ,in, Let be the rigid body transformation matrix. For GPS location measurement, Used as reference coordinates;

[0108] The SIFT feature descriptor in the SIFT+SuperGlu fusion is:

[0109] ;

[0110] The formula for SuperGlue matching probability is: ,in, For feature descriptors, The weight matrix is ​​for learning.

[0111] The formula for RANSAC robust transform estimation is: ,in, Here is the Huber loss function. Let be the set of interior points. The RANSAC threshold;

[0112] S2.2. Construct a photometric registration method, including using bundle adjustment optimization for low-altitude oblique photography and using the iterative nearest point algorithm (ICP) to optimize lidar point cloud data;

[0113] Furthermore, the formula for minimizing the reprojection error in bundle adjustment optimization is as follows: ,in, For the rotation and translation of the j-th camera, For the i-th 3D point, For projection functions;

[0114] The formula for the objective function of point cloud registration using the Iterative Closest Point (ICP) optimization algorithm is as follows: The convergence condition is: Maximum number of iterations ;

[0115] S2.3. Construct a semantic feature-assisted registration method and set semantic consistency constraints. The expression is:

[0116]

[0117] in, These are the i-th semantic tag and the j-th semantic tag, respectively. , Let i and j be the semantic feature vectors, respectively. Spatial weights, For indicator functions;

[0118] Perform multimodal fusion registration.

[0119]

[0120] in, To achieve the target value for multimodal fusion registration accuracy, Geometric registration error includes errors in feature point coordinates or residuals obtained from coarse registration. Photometric registration error, including image reprojection error or point cloud ICP residual, originates from the coarse and fine registration residuals obtained in S2.1 and S2.2 above. , and These are the balanced weighting parameters for geometric error, photometric error, and semantic constraints, respectively. .

[0121] Furthermore, the key parameters are explained below:

[0122] Scale-based strategy: coarse registration handles global transformation (meter-level accuracy), fine registration handles local optimization (millimeter-level accuracy).

[0123] Multimodal fusion weights: geometric features dominate (α=0.5), photometric features provide texture information (β=0.3), and semantic features enhance robustness (γ=0.2);

[0124] Convergence threshold: The ICP convergence threshold is set to 10^-6 to ensure sub-millimeter registration accuracy; RANSAC threshold: set to 0.01 pixels to balance computational efficiency and registration accuracy;

[0125] S3. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives and perform adaptive splash reconstruction to obtain an optimized 4D Gaussian splash model;

[0126] Furthermore, the specific implementation method of step S3 includes the following steps:

[0127] S3.1. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives;

[0128] S3.1.1. Initialize 3D Gaussian primitives from LiDAR point cloud data and perform initial Gaussian function seeding, including setting initial position, covariance, color, and opacity;

[0129] The Gaussian unit of 3D Gaussian is defined as: ,in, For spatial location, Let covariance matrix be the variance matrix. Opacity;

[0130] The initial position is set as follows: ,in, Given the coordinates of the i-th point in the point cloud; the initial covariance estimate is: ,in, Let i be the k nearest neighbors. For regularization parameters, It is the identity matrix;

[0131] S3.1.2. The 3D Gaussian primitives initialized in step S3.1.1 are expanded to 4D spatiotemporal Gaussian primitives, and time parameters are added to parameterize the initial position and opacity over time.

[0132] The 4D spatiotemporal Gaussian unit is defined as: The 4D covariance matrix is: ,in, For spatial covariance, Standard deviation over time;

[0133] Position-time parameterization is as follows: ,in, Basic position, For the range of motion, Angular frequency, For phase; opacity time parameterization is: ,in, Based on opacity, At peak time, For time-diffusion parameters;

[0134] S3.1.3. Construct an adaptive initialization strategy for the 4D spatiotemporal high-order primitives obtained in step S3.1.2;

[0135] The expression for the density-adaptive initialization strategy is:

[0136]

[0137] in, The density value obtained after adaptation. Based on density, For adaptive coefficients, , For image Laplacian operator;

[0138] The expression for the scale-adaptive initialization strategy is:

[0139]

[0140] in, For the Gaussian scale, The distance to the i-th nearest neighbor, for example, the distance from a point to its 5th nearest neighbor, is used to measure local density. The average distance, The global average neighbor distance is used as the normalization benchmark. For the scaling adjustment index, set ;

[0141] Furthermore, the key parameters are explained as follows: 4D Extension Innovation: For the first time, Gaussian splashing is extended from 3D to 4D, adding a temporal dimension to handle dynamic scenes; Temporal Parameterization: A temporal variation model for position and opacity supports periodic and non-periodic motion; Adaptive Initialization: Automatically adjusts the number and scale of Gaussian elements based on image features and point cloud density; Spatiotemporal Consistency: Ensures spatiotemporal continuity through the covariance matrix in the temporal dimension; Computational Efficiency Optimization: Regularization parameters... and adaptive coefficients Optimize memory usage and computation speed;

[0142] S3.2. For the 4D Gaussian primitives initialized in step S3.1, perform adaptive differentiable rendering optimization iteration for adaptive splash reconstruction. During the iterative process, the density is adaptively controlled by the adaptive splitting criterion and the adaptive deletion criterion to obtain the optimized 4D Gaussian splash model.

[0143] Furthermore, the 4D rendering equation for multi-view differentiable rendering is as follows:

[0144]

[0145] in, For pixel coordinates, For back projection function, The color changes over time;

[0146] The formula for multi-view consistency constraints is: ,in, For the number of viewpoints, For the pixel domain of the j-th viewpoint;

[0147] The comprehensive loss function designed for multimodal loss function is:

[0148]

[0149] Among them, photometric loss ,in, For Huber's losses, Geometric consistency loss ,in, For reference point cloud location, For smoothing regularization coefficients;

[0150] Time consistency loss ,in, To predict speed, For time intervals;

[0151] The momentum update formula for Adam optimizer parameter updates is: ;

[0152] The formula for updating the second moment is: ;

[0153] The formula for parameter update is: in: ;

[0154] The density evaluation index for adaptive density control is:

[0155]

[0156] in, For position The neighborhood, Density kernel bandwidth;

[0157] The formula for the adaptive splitting criterion is: ,in, This is the splitting threshold;

[0158] The formula for the adaptive deletion criterion is: ,in, The deletion threshold;

[0159] Furthermore, the key parameters are explained below:

[0160] Multimodal loss weights: (Luminosity) (geometry), (time), (regular expression);

[0161] Optimizer parameters: learning rate momentum coefficient Ensuring stable convergence; Adaptive frequency control: Density control is performed once every 100 iterations to balance computational efficiency and reconstruction quality; Threshold setting: Split threshold. and deletion threshold Ensure dynamic balance of Gaussian elements; Temporal consistency: ensure spatiotemporal continuity of 4D reconstruction through velocity prediction and time interval constraints;

[0162] S4. For the optimized 4D Gaussian splash model obtained in step S3, design a cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction, and perform distributed parallel processing on the optimized 4D Gaussian splash model.

[0163] Furthermore, step S4, which designs the specific implementation method of the cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction, includes the following steps:

[0164] S4.1. Perform edge-end intelligent preprocessing on the optimized 4D Gaussian splash model obtained in step S3;

[0165] First, SLAM state estimation and keyframe pose optimization are performed;

[0166] The formula for SLAM state estimation is: ,in, Let be the pose at time t. For observation, For observation models, For map points;

[0167] The formula for keyframe pose optimization is: , in, A set of keyframes For smoothing regularization coefficients;

[0168] Then, information entropy evaluation is used to construct keyframe selection criteria, and adaptive keyframes are selected.

[0169] The information entropy evaluation formula for adaptive keyframe selection is: ,in, The pixel intensity histogram probability; the formula for measuring motion amplitude is: ,in, These are the rotation weighting coefficients;

[0170] The keyframe selection criteria are as follows: ,in, (Information entropy threshold) (Motion threshold) Frame (minimum interval);

[0171] Finally, multimodal data compression and point cloud compression optimization are performed to obtain a 4D Gaussian splash model after intelligent preprocessing at the edge.

[0172] The formula for controlling the image compression rate in multimodal data compression is: ,in, dB ;

[0173] The formula for point cloud compression optimization is: ,in, m is the voxel size; the formula for the overall compression ratio is: (Target compression ratio);

[0174] S4.2. Perform parallel depth reconstruction of the 4D Gaussian splash model obtained from the edge-end intelligent preprocessing in step S4.2.1 in the cloud.

[0175] S4.2.1. Calculate load assessment and perform intelligent load balancing allocation. The expression is:

[0176]

[0177] in, Let i be the load evaluation function for node i. Let i be the amount of data at node i. Let be the computational complexity of node i. Let i be the memory usage of node i. , , These are the weighting coefficients for the data volume of node i, the computational complexity of node i, and the memory usage of node i, respectively.

[0178] Furthermore, ;

[0179] Then, task allocation optimization is performed, expressed as follows:

[0180]

[0181] in, To optimize the target value for task scheduling, Let be the total time cost of the i-th node. Let be the processing time for node i. For communication overhead, For communication weights;

[0182] S4.2.2. Utilize data parallelism strategies for parallel GPU cluster processing;

[0183] The formula for the data parallelism strategy is: ,in, To share model parameters;

[0184] The formula for gradient aggregation is: ,in, This refers to the number of GPU nodes.

[0185] The formula for parameter synchronization is: ,in, The learning rate;

[0186] S4.2.3. Distributed result fusion is performed on the 4D Gaussian splash model after intelligent preprocessing at the edge after parallel GPU cluster processing, including spatial partition fusion and overlapping region processing, to obtain the optimized 4D Gaussian splash model after distributed parallel processing.

[0187] The formula for spatial partition fusion is: ,in, Partition the space of node i;

[0188] The formula for handling overlapping regions is: ,in, , To integrate core bandwidth;

[0189] Furthermore, the key parameters are explained below:

[0190] Edge processing parameter configuration: keyframe selection rate 20%, compression ratio 30%, SLAM optimization frequency 20Hz to ensure real-time performance; load balancing weight: data volume ratio 40%, calculation ratio 40%, memory usage ratio 20%, dynamic resource scheduling and allocation; parallel processing configuration: GPU cluster size 4-8 nodes, gradient aggregation uses AllReduce algorithm to ensure parameter consistency; fusion parameter settings: spatial partition overlap rate 15% to ensure smooth transition of boundary regions;

[0191] S5. Perform quality assessment and adaptive optimization on the optimized 4D Gaussian splash model obtained after distributed parallel processing in step S4, and output the final adaptive 4D Gaussian splash model.

[0192] Furthermore, the specific implementation method of step S5 includes the following steps:

[0193] S5.1. Construct a geometric accuracy evaluation method, including calculating point cloud distance error, normal vector consistency, and curvature fidelity;

[0194] The formula for point cloud distance error is: ,in, For reference point cloud, To reconstruct the points in the point cloud;

[0195] The formula for normal vector consistency is: ,in, and These are the reconstructed and reference normal vectors, respectively.

[0196] The formula for curvature fidelity is: ,in, Principal curvature, It is a numerically stable term;

[0197] S5.2. Construct a texture fidelity evaluation method, including calculating the structural similarity index, perceptual hash similarity, and multi-scale texture consistency;

[0198] The formula for the structural similarity index is: ,in, , , Range of pixel values;

[0199] The formula for perceptual hash similarity is: ;

[0200] The formula for multi-scale texture consistency is: in: For scale quantity, The image is at the s-th scale;

[0201] S5.3. Construct an integrity assessment method, including calculating volume overlap rate, surface coverage, and detail retention;

[0202] The formula for volume overlap ratio is: ,in, and These are the reconstructed and reference volumes, respectively.

[0203] The formula for surface coverage is: ,in, For surface point set, m is the surface distance threshold;

[0204] The formula for detail retention is: ,in, The number of edge points. For the edge;

[0205] S5.4. Construct a time consistency assessment method, including calculating motion continuity, deformation consistency, and spatiotemporal smoothness;

[0206] The formula for the continuity of motion is: ,in, Let be the velocity vector at time t. ;

[0207] The formula for consistent deformation is: ,in, The expected deformation vector;

[0208] The formula for spatiotemporal smoothness is: ,in, For time gradient, To smooth out kernel bandwidth;

[0209] S5.5. Construct a comprehensive quality assessment index, perform a weighted comprehensive assessment, and obtain the expression:

[0210]

[0211] in, As a weighted comprehensive evaluation indicator, This refers to the point cloud distance error. For multi-scale texture consistency, For surface coverage, For the continuity of motion, , , , The weights of point cloud distance error, multi-scale texture consistency, surface coverage, and motion continuity in the comprehensive quality evaluation index are respectively:

[0212] S5.6. Based on the weighted comprehensive evaluation index obtained in step S5.5, perform adaptive weight adjustment and confidence interval evaluation to complete adaptive optimization and output the final adaptive 4D Gaussian splash model.

[0213] The formula for weighted comprehensive evaluation is: ;

[0214] The formula for adaptive weight adjustment is: ,in, and Let be the historical mean and variance of the i-th indicator;

[0215] The formula for confidence interval assessment is: ,in, This represents the number of samples.

[0216] Furthermore, the key parameters are explained as follows: Geometric accuracy threshold: point cloud distance error < 0.01m, normal vector consistency > 0.95, curvature fidelity > 0.9; Texture fidelity standard: SSIM > 0.85, perceptual hash similarity > 0.9, multi-scale consistency > 0.8; Integrity requirements: volume overlap > 0.95, surface coverage > 0.9, detail retention > 0.85; Temporal consistency index: motion continuity > 0.9, deformation consistency error < 0.005m, spatiotemporal smoothness > 0.85; Comprehensive evaluation weight: geometric accuracy 30%, texture fidelity 25%, integrity evaluation 25%, temporal consistency 20% - Adaptive mechanism: dynamically adjusts weights based on historical data, providing a 95% confidence interval evaluation.

[0217] Example 2:

[0218] A system for predicting and analyzing the energy-saving effect of a drone's climb process includes a processor, a memory, and a computer program stored in the memory and run on the processor. When the computer program runs, it implements the steps of the method for predicting and analyzing the energy-saving effect of a drone's climb process as described in Example 1.

[0219] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0220] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An adaptive 4D Gaussian splash high-precision 3D reconstruction method, characterized in that: The process includes the following steps: S1. Collect multi-source heterogeneous data, including satellite imagery data, low-altitude oblique photography data, low-altitude lidar point cloud data, low-altitude video stream data, ground acquisition vehicle lidar point cloud data, and ground acquisition vehicle video stream data. Then, perform preprocessing, including data source standardization processing and multi-source device spatiotemporal reference alignment processing. S2. For the multi-source heterogeneous data preprocessed in step S1, construct a multi-modal fusion registration method, including combining satellite imagery data and low-altitude oblique photography data to achieve coarse spatial distribution geometric registration, fusing low-altitude lidar point cloud data and ground-based vehicle lidar point cloud data to achieve fine photometric registration, and establishing semantic feature-assisted registration to obtain registered multi-source data; S3. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives and perform adaptive splash reconstruction to obtain an optimized 4D Gaussian splash model; The specific implementation method of step S3 includes the following steps: S3.

1. Based on the registered multi-source data obtained in step S2, initialize 4D Gaussian primitives; S3.1.

1. Initialize 3D Gaussian primitives from lidar point cloud data, and perform initial Gaussian function seeding, including setting initial position, covariance, color, and opacity; S3.1.

2. The 3D Hodgkin primitives initialized in step S3.1.1 are expanded to 4D spatiotemporal Hodgkin primitives, and time parameters are added to parameterize the initial position and opacity over time; S3.1.

3. An adaptive initialization strategy is constructed for the 4D spatiotemporal Hodgkin primitives obtained in step S3.1.2; the expression for the density adaptive initialization strategy is: ;in, The density value obtained after adaptation. The initial default primitive density, For adaptive coefficients, ， For the image Laplacian operator; the expression for the scale-adaptive initialization strategy is: ;in, For the Gaussian scale, The distance to the i-th nearest neighbor is used to measure local density. The average distance, The global average neighbor distance is used as the normalization benchmark. For the scaling adjustment index, set S3.

2. For the 4D Gaussian primitives initialized in step S3.1, perform adaptive differentiable rendering optimization iteration for adaptive splash reconstruction. During the iterative process, density is adaptively controlled through adaptive splitting and deletion criteria to obtain an optimized 4D Gaussian splash model. S4. For the optimized 4D Gaussian splash model obtained in step S3, design a cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction and perform distributed parallel processing on the obtained optimized 4D Gaussian splash model. The specific implementation method of designing the cloud-edge collaborative computing architecture for 4D Gaussian splash reconstruction in step S4 includes the following steps: S4.

1. Perform edge-end intelligent preprocessing on the optimized 4D Gaussian splash model obtained in step S3. First, perform SLAM state estimation and keyframe pose optimization. Then, use information entropy evaluation to construct a keyframe selection criterion and select adaptive keyframes. Finally, perform multimodal data compression and point cloud compression optimization to obtain the edge-end intelligent preprocessed 4D Gaussian splash model. S4.

2. The 4D Gaussian splash model obtained from the edge-end intelligent preprocessing in step S4.2.1 is then reconstructed in parallel in the cloud; S4.2.

1. Load assessment is calculated, and intelligent load balancing is performed, with the following expression: ;in, Let i be the load evaluation function for node i. Let i be the amount of data at node i. Let be the computational complexity of node i. Let i be the memory usage of node i. 、 、 These are the weight coefficients for the data volume of node i, the computational complexity of node i, and the memory usage of node i, respectively; then, task allocation optimization is performed, expressed as: ;in, To optimize the target value for task scheduling, Let be the total time cost of the i-th node. Let be the processing time for node i. For communication overhead, S4.2.

2. Perform parallel GPU cluster processing using a data parallel strategy; S4.2.

3. Perform distributed result fusion on the edge-end intelligent preprocessed 4D Gaussian splash model after parallel GPU cluster processing, including spatial partition fusion and overlapping region processing, to obtain the optimized 4D Gaussian splash model after distributed parallel processing; S5. Perform quality assessment and adaptive optimization on the optimized 4D Gaussian splash model obtained in step S4 after distributed parallel processing, and output the final adaptive 4D Gaussian splash model.

2. The adaptive 4D Gaussian splash high-precision 3D reconstruction method according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Perform data source standardization processing on the acquired multi-source heterogeneous data, including radiometric and geometric correction on the acquired satellite image data, color balance and radial distortion correction on the low-altitude oblique photography data; noise filtering and coordinate transformation on the low-altitude lidar point cloud data and the ground acquisition vehicle lidar point cloud data; optical flow stabilization processing and key frame extraction on the low-altitude video stream data and the ground acquisition vehicle video stream data; S1.

2. Multi-source device spatiotemporal reference alignment processing; S1.2.

1. Time base unification and synchronization, including GPS time base conversion, multi-source device time synchronization, and timestamp interpolation alignment; S1.2.

2. Multi-source positioning information fusion and alignment, including GPS / RTK location fusion, BeiDou / GPS dual-system fusion, and inertial navigation-assisted alignment; S1.2.

3. Equipment attitude alignment correction, including constructing a standardized attitude matrix, performing Euler angle transformation and quaternion interpolation alignment; S1.2.

4. Perform spatiotemporal registration verification on the multi-source data processed in steps S1.2.1-S1.2.3, evaluate the time synchronization accuracy and spatial alignment accuracy, and set comprehensive alignment quality indicators.

3. The adaptive 4D Gaussian splash high-precision 3D reconstruction method according to claim 2, characterized in that, The specific implementation method of step S2 includes the following steps: S2.

1. Constructing a geometric coarse registration method, including GPS / RTK pose alignment, SIFT+SuperGlue fusion, and RANSAC robust transform estimation; S2.

2. Constructing a photometric fine registration method, including using bundle adjustment optimization for low-altitude oblique photography and using the Iterative Closest Point (ICP) algorithm for LiDAR point cloud data; S2.

3. Constructing a semantic feature-assisted registration method and setting semantic consistency constraints. The expression is: ;in, These are the i-th semantic tag and the j-th semantic tag, respectively. 、 Let i and j be the semantic feature vectors, respectively. Spatial weights, As an indicator function; perform multimodal fusion registration. ;in, To achieve the target value for multimodal fusion registration accuracy, Geometric registration error includes errors in feature point coordinates or residuals obtained from coarse registration. Photometric registration error, including image reprojection error or point cloud ICP residual, originates from the coarse and fine registration residuals obtained in S2.1 and S2.2 above. 、 and These are the balance weight parameters for geometric error, photometric error, and semantic constraints, respectively. 。 4. The adaptive 4D Gaussian splash high-precision 3D reconstruction method according to claim 3, characterized in that, The specific implementation method of step S5 includes the following steps: S5.

1. Construct a geometric accuracy evaluation method, including calculating point cloud distance error, normal vector consistency, and curvature fidelity; S5.

2. Construct a texture fidelity evaluation method, including calculating structural similarity index, perceptual hash similarity, and multi-scale texture consistency; S5.

3. Construct an integrity evaluation method, including calculating volume overlap rate, surface coverage, and detail retention; S5.

4. Construct a temporal consistency evaluation method, including calculating motion continuity, deformation consistency, and spatiotemporal smoothness; S5.

5. Construct a comprehensive quality evaluation index, perform a weighted comprehensive evaluation, and obtain the expression: ;in, As a weighted comprehensive evaluation indicator, This refers to the point cloud distance error. For multi-scale texture consistency, For surface coverage, For the continuity of motion, 、 、 、 The weights of point cloud distance error, multi-scale texture consistency, surface coverage, and motion continuity in the comprehensive quality evaluation index are respectively; S5.

6. Based on the weighted comprehensive evaluation index obtained in step S5.5, adaptive weight adjustment and confidence interval evaluation are performed to complete adaptive optimization and output the final adaptive 4D Gaussian splash model.

5. An adaptive 4D Gaussian splash high-precision 3D reconstruction system, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed, implements the steps of an adaptive 4D Gaussian splash high-precision three-dimensional reconstruction method as described in any one of claims 1-4.

Images