A gaussian ball clustering method for three-dimensional reconstruction of expressway traffic accident scene

Abstract

The application provides a Gaussian ball clustering method for three-dimensional reconstruction of a highway traffic accident site, comprising the following steps: S1, collecting multi-view images of the accident site, solving camera parameters and initializing a three-dimensional Gaussian ball scene; S2, constructing a rendering training loop based on the initialized three-dimensional Gaussian ball scene, and adaptively adjusting the Gaussian ball density through cloning or splitting operations; S3, extracting the spatial and color features of the three-dimensional Gaussian ball, and constructing a multi-modal physical feature set; S4, calculating an adaptive discrete step length based on the multi-modal physical feature set, and mapping the continuous features into discrete coordinates; S5, generating a multi-modal hash index based on the discrete coordinates, and realizing continuous rearrangement of the same type of Gaussian ball in the display memory through GPU parallel sorting; S6, performing physical conservation formula parameter aggregation on the same type of Gaussian ball continuously stored in the display memory, releasing redundant display memory and simplifying the three-dimensional Gaussian ball set; and S7, completing all training iterations, and exporting the final simplified three-dimensional Gaussian ball set into a general PLY format file.

Description

Technical Field

[0001] This invention belongs to the field of traffic accident scene survey and three-dimensional reconstruction technology, and in particular relates to a Gaussian sphere clustering method for three-dimensional reconstruction of highway traffic accident scenes. Background Technology

[0002] 3D reconstruction technology is a method to recover the geometric and apparent properties of a 3D scene by analyzing multi-view 2D images. In this way, the accident scene is digitized into a model that can be surveyed comprehensively in virtual space. Because highway traffic accidents typically involve a wide accident scene (often stretching hundreds of meters) and fragmented evidence (including glass shards, scattered metal parts, tire skid marks, etc.), high-fidelity 3D reconstruction has irreplaceable judicial value for accident liability determination and collision trajectory reconstruction, and helps improve the efficiency of traffic police in clearing and surveying the scene. Traditional 3D reconstruction methods (such as photogrammetry or early Neural Radiation Field (NeRF) methods) suffer from drawbacks such as slow rendering speed, difficulty in real-time editing, and extremely high demands on graphics card computing power. Furthermore, different algorithms vary significantly in their ability to preserve fine edges. 3D Gaussian Splatting (3DGS) is a new paradigm for 3D reconstruction based on explicit point clouds. Initially proposed in recent years, it has now become an analytical tool for implementing digital twins and immersive scene reconstruction. The 3DGS system employs a method called "Adaptive Density Control (ADC)" to quantify the density of a scene. The physical elements used include spatial coordinates, opacity, rotational quaternions, and spherical harmonics (color), and then combined with differentiable rasterization technology to achieve real-time rendering. While this method offers fast rendering speeds, it generates tens of millions of Gaussian spheres to fit minute scratches or debris on a highway, resulting in extremely high parameter counts and memory consumption, which can easily cause surveying terminals to crash.

[0003] Clustering analysis is a common method for reducing Gaussian spheres. It involves unsupervised learning of attributes (such as spatial distance) in unlabeled data to divide redundant Gaussian points into multiple clusters and then merge them. Currently, various clustering compression algorithms are used in 3DGS, with one of the most classic methods being spatial distance clustering based on voxel grids or K-Means. These algorithms can find Gaussian spheres with similar locations by calculating coordinate distances and then averaging the parameters. Later, to address the memory explosion problem in large scenes, a series of improved algorithms based on hierarchical bounding boxes (BVH) emerged. However, these algorithms only focus on solving the problem of spatial location partitioning and do not improve the fundamental principle of clustering algorithms, namely, characterizing the joint boundary of physical materials and photometric semantics. Therefore, in certain extreme situations at highway accident scenes, such as when there are significant differences between large areas of homogeneous asphalt pavement (large clusters) and small white skid marks or transparent glass shards close to the pavement (small clusters), traditional spatial clustering is prone to forcibly merging Gaussian spheres that are very close to each other but have completely different materials, resulting in serious "semantic collapse" and loss of physical evidence, and thus failing to obtain satisfactory judicial-grade survey results.

[0004] Therefore, given the shortcomings of the existing technologies described above, it is necessary to apply a new multimodal manifold clustering algorithm to the 3D reconstruction of traffic accident scenes to solve the problems of existing technologies relying solely on spatial distance and having poor clustering fidelity for datasets with significant differences and many redundant points between large clusters of highways and small clusters of physical evidence. Summary of the Invention

[0005] The purpose of this invention is to address the characteristics of numerous redundant points on highway surfaces and detailed collision center evidence by proposing a multimodal hash clustering method. This method integrates photometric attributes and physical conservation constraints on top of traditional spatial location attributes, and through adaptive step size control, it achieves the goal of simplifying road surface redundancy while retaining key debris.

[0006] Technical Solution: To solve the above-mentioned technical problems and achieve the above-mentioned objectives, this invention proposes a Gaussian sphere clustering method for 3D reconstruction of highway traffic accident scenes, comprising the following steps:

[0007] S1: Acquire multi-view images of the accident scene, calculate camera parameters, and initialize a 3D Gaussian sphere scene;

[0008] S2 constructs a rendering training loop based on an initialized 3D Gaussian sphere scene, and adaptively adjusts the Gaussian sphere density through cloning or splitting operations.

[0009] S3, extract the spatial and color features of all three-dimensional Gaussian spheres and construct a multimodal physical feature set;

[0010] S4, based on the multimodal physical feature set, calculates the adaptive discrete step size, mapping continuous features to discrete coordinates;

[0011] S5 generates a multimodal hash index based on discrete coordinates and achieves continuous rearrangement of similar Gaussian sphere memory through GPU parallel sorting;

[0012] S6 performs physical conservation parameter aggregation on the Gaussian spheres of the same type stored contiguously in the video memory, releasing redundant video memory and simplifying the three-dimensional Gaussian sphere set.

[0013] S7 completes all training iterations and exports the final simplified 3D Gaussian sphere set as a universal PLY format file.

[0014] Furthermore, in step S1, multi-view images of the accident scene are acquired, camera parameters are calculated, and a three-dimensional Gaussian sphere scene is initialized. The specific method is as follows:

[0015] S1.1, Use drones or handheld devices to collect video footage of traffic accident scenes, extract keyframes from the collected video, and obtain multi-view two-dimensional image sequences. ,in, , Total number of images;

[0016] S1.2, for the input two-dimensional image sequence Perform feature point detection and extract the first feature point. Two-dimensional feature point coordinates of all feature points on the image plane ,in For the first Local index of feature points within an image;

[0017] S1.3 calculates the similarity of feature points in different images using a feature matching operator, and finds two-dimensional feature points belonging to the same real space point observed from different viewpoints based on the rule of equal similarity. These two-dimensional feature points corresponding to the same real-space point are combined together to form a set of common-view feature points unique to that real-space point. ,in, This is the global index of the real-world point. , For each set, the total number of real-world points calculated in the scene is represented. It contains two-dimensional feature points of the spatial point on multiple different images; denoted as... For including the first The set of image indices for real points, then for any image indices... , All of them contain the first The image corresponding to the first Feature points of a real point ;

[0018] S1.4, set all extracted common-view feature points Input Structure from Motion (SfM) algorithm, perform Bundle Adjustment (BA) optimization, and output each frame of image. Corresponding camera intrinsic parameter matrix Rotation matrix Translation vector ;

[0019] S1.5, Extract Set pixel coordinates Based on the corresponding camera parameters The two-dimensional feature points on each frame of the image are projected as a ray in three-dimensional space. Due to image noise and camera parameter errors, the rays will not intersect perfectly. A reprojection error function is constructed as follows:

[0020]

[0021] in, For perspective projection functions, For the first The three-dimensional coordinates of a real-world point to be solved. For including the first A set of image indices for real points;

[0022] S1.6, Minimize the error function using the nonlinear least squares method. Solve and output the optimal 3D intersection point of multiple rays. That is, the coordinates of the real object in three-dimensional space, and the RGB color values ​​of the corresponding original pixels in the image are extracted. Combined with coordinates, output the initial sparse 3D point cloud set. ;

[0023] S1.7, using sparse 3D point cloud sets As the input basis, each real space point This generates a corresponding three-dimensional Gaussian sphere, thus generating a set of three-dimensional Gaussian spheres. ,in, The total number of three-dimensional Gaussian spheres, initially During training, the dynamic increase in cloning and splitting operations affects the set. Each three-dimensional Gaussian sphere The following initial parameter assignment operations are performed in parallel:

[0024] Assign center coordinates : , which represents the physical center position of the Gaussian sphere in three-dimensional space;

[0025] Calculate the spherical harmonic coefficients : Color values Perform inverse spherical projection transformation, output and assign initial spherical harmonic coefficients. , used to represent the diffuse color of a Gaussian sphere;

[0026] Assigning opacity : , indicating the transparency of the Gaussian sphere;

[0027] Assignment rotation quaternion : This indicates that the Gaussian sphere initially has no rotation;

[0028] Calculate scaling Calculate the current center point To point cloud collection The Euclidean distance between the three nearest neighbors is taken as the average of these three distances. Assign the initial scaling vector to... This indicates that the Gaussian sphere is in , , Initial size along the three axes.

[0029] Furthermore, in step S2, a rendering training loop is constructed based on the initialized 3D Gaussian sphere scene, and the Gaussian sphere density is adaptively adjusted through cloning or splitting operations, as follows:

[0030] S2.1, in each training iteration, first from all One real photograph is randomly selected as the current training frame, and the corresponding camera intrinsic parameter matrix is ​​extracted. With external references ;

[0031] S2.2, set the three-dimensional Gaussian spheres All three-dimensional Gaussian spheres The world coordinates are converted to relative coordinates with the current camera lens as the origin. Then, three-dimensional Gaussian spheres are selected that fall within the camera's view frustum, have a Z-axis depth distance between 0.1 meters and 1000 meters in the camera coordinate system, an opacity of not less than 0.001, and a size of not less than 1 pixel after projection onto the image plane. These three-dimensional Gaussian spheres are then used to form the set of visible three-dimensional Gaussian spheres in the current frame's image viewpoint. ;

[0032] S2.3, set The center coordinates, scaling factor, and rotation quaternion of each 3D Gaussian sphere are projected onto the 2D image plane of the current frame. Alpha-blending is then performed according to the depth along the Z-axis of the camera coordinate system from far to near, rendering and outputting the 2D predicted image of the current frame. ;

[0033] S2.4, Extract the two-dimensional prediction image of the current frame. Corresponding real two-dimensional image Calculate the pixel-by-pixel absolute error between the two. and structural similarity loss That is, SSIM loss, construct the joint loss function:

[0034]

[0035] In the formula, Preset weighting coefficients;

[0036] S2.5, the calculated global loss Perform gradient backpropagation to obtain the gradient values ​​of all physical parameters of each visible Gaussian sphere. ,in, Represents any parameter, including center coordinates. spherical harmonic coefficient Opacity Rotation Quaternions Scaling The gradient of an invisible Gaussian sphere is 0, and the parameters remain unchanged.

[0037] S2.6, based on the Adam gradient descent optimizer, adjusts the above parameters. , representing any parameter to be optimized , , , and Perform fully parallel synchronous updates:

[0038] For each class of trainable parameters, maintain two independent internal state variables, where the first moment estimate... Used to accumulate historical gradient information, smooth the parameter update process, and accelerate convergence; second-order moment estimation This is used to accumulate historical gradient squared information and adaptively adjust the learning rate for each parameter.

[0039] First-moment estimation of parallel parameter updates:

[0040] Second-moment estimation of parallel parameter updates:

[0041] Parallel bias correction is applied to the first and second moments of the parameters to eliminate statistical bias caused by initial values ​​of 0.

[0042] ,

[0043] The final update of all parameters is calculated in parallel:

[0044]

[0045] Perform iterative updates on all original parameters: ;

[0046] In the formula, This represents the current training iteration number. , This is an internal state variable of the Adam optimizer, used only for parameter update calculations, and is unrelated to the 3D Gaussian sphere index. , These are the first and second moment decay coefficients of the Adam optimizer; The learning rate; It is the numerical stability constant;

[0047] S2.7, for the updated rotation quaternion Perform a normalization operation to ensure that its modulus is 1, in order to maintain normal rotation, as shown below:

[0048]

[0049] S2.8 updates and overwrites the 3D Gaussian sphere set with all updated parameters. The corresponding visible set Gaussian sphere parameters;

[0050] S2.9, after each fixed number of iterations, calculate the field-of-view position gradient vector of the spatial center coordinates of each Gaussian sphere. And find its second norm. Compare it with the set position gradient threshold Compare and determine:

[0051] when At that time, based on the scaling scale of the corresponding Gaussian sphere The size of the 3D Gaussian sphere is determined, and corresponding memory topology operations are performed: if the volume is smaller than the scale threshold, a cloning operation is performed to generate a new 3D Gaussian sphere with identical parameters, which is then superimposed at the same location to enhance the reconstruction details; if the volume is larger than the scale threshold, a splitting operation is performed, cutting the sphere into two smaller sub-3D Gaussian spheres along the direction of the largest operation error, which are then fitted with details from different regions respectively; the newly generated 3D Gaussian spheres are then merged into a set. ;

[0052] Repeat steps S2.1 to S2.9 until the current iteration step. Reaching the preset clustering trigger round .

[0053] Furthermore, in step S3, the spatial and color features of all three-dimensional Gaussian spheres are extracted to construct a multimodal physical feature set. The specific method is as follows:

[0054] S3.1, when training iteration step Reaching the preset clustering trigger round At this point, global gradient backpropagation is paused during rendering, where a unique integer index of all Gaussian spheres is given. , , The total number of current 3D Gaussian spheres, using this global index. For addressing indexes, traverse the set of three-dimensional Gaussian spheres in the current video memory. Extract and output the three-dimensional spatial center coordinate vector of the full-field three-dimensional Gaussian sphere. ;

[0055] S3.2, based on the exact same global sequence number Synchronous traversal and synchronous extraction of the collection Spherical harmonic coefficients corresponding to the Gaussian sphere Extract its zeroth-order spherical harmonic coefficient components Perform a linear transformation operation in the color space, calculate and output the diffuse color vector of the corresponding Gaussian sphere. :

[0056]

[0057] S3.3 Extract the spatial center coordinate vector obtained from the above calculation. With diffuse color vector Perform tensor concatenation operation along the channel dimension to output the multimodal physical feature vector of the single Gaussian sphere. :

[0058]

[0059] S3.4, integrating all Generate and output a global multimodal physical feature set from the multimodal physical feature vectors of a three-dimensional Gaussian sphere. , which serves as the input basis for subsequent discrete quantization and mapping algorithms.

[0060] Furthermore, in step S4, an adaptive discrete step size is calculated based on the multimodal physical feature set, mapping continuous features to discrete coordinates. The specific method is as follows:

[0061] S4.1 Extract the set of 3D Gaussian spheres in the current video memory. The global bounding box, i.e., the coordinates of the centers of all Gaussian spheres. Calculate the bounding box of the cuboid formed by the maximum and minimum values ​​of . , , The coordinate ranges along the three axes are taken, and the maximum value of the three ranges is multiplied by a preset spatial resolution scaling factor to calculate and output the base spatial deviation range. ;

[0062] S4.2, from the global multimodal physical feature set Extract all A diffuse color vector of a three-dimensional Gaussian sphere Calculate the global mean of all colors in the field. and global standard deviation :

[0063]

[0064] The current iteration step is calculated using a linear time decay strategy. Corresponding time decay factor and preset lower limit of photometric step size Perform the following scalar operations to calculate and output the dynamic photometric distance from the walk. :

[0065]

[0066] In the formula, This represents the current training iteration number. This is the preset total number of training iterations; The initial attenuation coefficient is 3.0. The preset minimum luminosity distance is set to 0.01.

[0067] S4.3, by serial number Using the index, traverse the set of three-dimensional Gaussian spheres in the current video memory. Parallel extraction of multimodal physical feature vectors for each three-dimensional Gaussian sphere Extracted spatial coordinates With color vector Divide by the corresponding distance from the walk length respectively and And perform the floor operation:

[0068]

[0069] The two sets of discrete coordinates calculated are concatenated by channel to output the multimodal discrete coordinate vector corresponding to the three-dimensional Gaussian sphere. Integrate all data to output a global discrete coordinate set. .

[0070] Furthermore, in step S5, a multimodal hash index is generated based on discrete coordinates, and the same type of Gaussian sphere memory is continuously rearranged through GPU parallel sorting. The specific method is as follows:

[0071] S5.1, by serial number Use the index to traverse the set of discrete coordinates Extract the multimodal discrete coordinate vectors of each three-dimensional Gaussian sphere. Introduce a pre-defined set of independent large prime numbers. With global hash table capacity parameter Perform component multiplication, component XOR, and global modulo operations sequentially on the discrete coordinate components:

[0072]

[0073] In the formula, This indicates that vectors are multiplied by their components. Represents the XOR operation on vector components; a set of independent large prime numbers. It consists of 6 distinct large prime numbers, with values ​​{73856093, 19349663, 83492791, 1299709, 3162277, 5732533}. The global hash table capacity parameter is... Take the integer power of 2, and the value is... or ;

[0074] Calculate and output the one-dimensional scalar hash key corresponding to the three-dimensional Gaussian sphere. The integrated calculation results output a global hash index set. ;

[0075] S5.2 invokes the graphics processor's parallel radix sort algorithm, using the hash keys of all three-dimensional Gaussian spheres. Using the sort key, the memory addresses of all 3D Gaussian spheres in video memory are globally rearranged in ascending order; after rearrangement, an ordered hash set corresponding one-to-one with the video memory storage order is output. ,in, Local index of the elements in this set This corresponds one-to-one with the storage order of the Gaussian spheres in the sorted video memory. After rearrangement, the global three-dimensional Gaussian sphere set is complete. They have been stored contiguously in video memory in ascending order of hash value, so that Gaussian balls with the same hash value form contiguous memory blocks;

[0076] S5.3, Extracting the Ordered Hash Set Perform parallel prefix sum scan operations on it to calculate the difference determination identifier between adjacent hash key values. :

[0077]

[0078] in, At that time, the default It is infinite, that is ;

[0079] When the difference identifier When the cluster boundary marker is triggered: three-dimensional Gaussian spheres with the same hash key value form a cluster, and this position is the starting offset address of a new cluster in video memory;

[0080] The total number of clusters obtained from this clustering is obtained by summing all the difference labels:

[0081]

[0082] Then, calculate and output the length set of all clusters based on the address difference between two adjacent boundary markers. That is, the first The number of three-dimensional Gaussian spheres contained in each cluster.

[0083] Furthermore, in step S6, for all Each cluster performs a parallel merging operation, independently generating a new conserved 3D Gaussian sphere for each cluster, releasing redundant GPU memory and simplifying the 3D Gaussian sphere set. The specific method is as follows:

[0084] S6.1, for the first Extracting clusters within a cluster A set of opacity parameters for a three-dimensional Gaussian sphere Merge the transparency parameters, where, For cluster local indexes, Based on the Beer-Lambert law, logarithmic probabilistic aggregation is performed to calculate and output the aggregated opacity scalar. :

[0085]

[0086] S6.2 Extract the set of rotation quaternions of the 3D Gaussian sphere within the cluster. The rotation angle parameters were merged, and the quaternion of the three-dimensional Gaussian sphere with the highest opacity within the cluster was selected as the alignment reference. Introducing symbolic functions Perform hemispherical manifold correction and normalization operations, calculate and output the aggregated rotation quaternion vector. :

[0087]

[0088] S6.3 Extract the set of spatial center coordinates of the three-dimensional Gaussian sphere within the cluster. Diffuse color set With scaling scale set Merge coordinate, color, and size parameters, and adjust the corresponding opacity. For each weighting factor, perform a linear weighted average operation:

[0089]

[0090]

[0091]

[0092] Calculate and output the coordinates of the aggregation center. Aggregate diffuse color With aggregation scaling scale To aggregate colors Perform reverse color transformation Output the updated spherical harmonic coefficients ;

[0093] S6.4, Combine the new physical parameters obtained from the above calculations. Write the starting physical address of the current cluster in the video memory, overwrite the original first Gaussian sphere parameter at that address, and generate a single conserved three-dimensional Gaussian sphere.

[0094] S6.5, the consecutive addresses following the starting address of this cluster in video memory. One address space is marked as free and forcibly released; the remaining addresses within the cluster are deleted. Storage space occupied by the old three-dimensional Gaussian sphere;

[0095] S6.6, Update the total number of global 3D Gaussian spheres. The memory address offset of all subsequent clusters is adjusted, the gradient backpropagation pause state of training is released, and the updated simplified three-dimensional Gaussian sphere set is used to return to step S2 to continue iterative adjustment. Each time the iteration step reaches the set iteration round, a cluster merging operation is automatically triggered until the preset maximum training round is reached.

[0096] Furthermore, in step S7, after completing all training iterations, the final simplified 3D Gaussian sphere set is exported as a universal PLY format file, as follows:

[0097] S7.1 When the number of iterations reaches the preset maximum number of rounds, stop all gradient updates and clustering operations, and freeze the parameters of all current 3D Gaussian spheres;

[0098] S7.2, Traverse all the final simplified versions in the video memory For each 3D Gaussian sphere, its center coordinates, scaling factor, rotation quaternion, opacity, and spherical harmonics are extracted sequentially. Following the order of these parameters and strictly adhering to the vertex attribute format of the standard PLY file, all parameters of each Gaussian sphere are organized into a single vertex attribute data entry.

[0099] S7.3, create an empty PLY file, and define the total number of vertices of the 3D Gaussian sphere in the PLY file. And the type and order of all physical properties corresponding to each vertex;

[0100] S7.4, organize The vertex attribute data of each Gaussian sphere is continuously written into the vertex data segment of the PLY file, and exported as the final complete 3D reconstruction PLY model file of the accident scene. The PLY model file is then input into the SIBR_viewer streaming software for 3D reconstruction.

[0101] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0102] (1) By integrating the multimodal joint clustering mechanism of spatial location and photometric attributes, it can accurately distinguish three-dimensional Gaussian spheres that are close in distance but different in material. While simplifying the redundancy of large-area homogeneous regions, it can fully preserve the key fragmented features in the scene, fundamentally solving the semantic collapse problem that traditional spatial clustering relies solely on distance division, and meeting the requirements of judicial-level high-precision surveying.

[0103] (2) By embedding the clustering and merging operation into the training loop of 3D Gaussian reconstruction, online lightweight processing is achieved during the training process. No separate post-processing step is required. Redundant video memory is dynamically released, which effectively reduces the peak video memory usage during the training process and significantly shortens the overall reconstruction process time.

[0104] (3) Clustering is completed by GPU fully parallel hash sorting and parallel prefix sum scanning, which significantly improves the overall computational efficiency of clustering and reconstruction and can meet the needs of rapid on-site processing.

[0105] (4) The output lightweight three-dimensional Gaussian sphere collection has a significantly reduced volume and can be rendered in real time on various general-purpose terminal devices. It supports fast model transmission, remote viewing and interactive operation, and has good practicality and promotion value. Attached Figure Description

[0106] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0107] The following will provide a more detailed description of this embodiment in conjunction with the accompanying drawings used in the prior art.

[0108] The above description is a further detailed explanation of the present invention in conjunction with specific embodiments, and it should not be considered that the specific implementation of the present invention is limited to this. For those skilled in the art to which the present invention pertains and related fields, any extensions, operation methods, and data substitutions made based on the technical solution concept of the present invention should fall within the protection scope of the present invention.

[0109] like Figure 1 As shown, this invention proposes a Gaussian sphere clustering method for 3D reconstruction of highway traffic accident scenes, comprising the following steps:

[0110] S1: Acquire multi-view images of the accident scene, calculate camera parameters, and initialize a 3D Gaussian sphere scene;

[0111] S2 constructs a rendering training loop based on an initialized 3D Gaussian sphere scene, and adaptively adjusts the Gaussian sphere density through cloning or splitting operations.

[0112] S3, extract the spatial and color features of all three-dimensional Gaussian spheres and construct a multimodal physical feature set;

[0113] S4, based on the multimodal physical feature set, calculates the adaptive discrete step size, mapping continuous features to discrete coordinates;

[0114] S5 generates a multimodal hash index based on discrete coordinates and achieves continuous rearrangement of similar Gaussian sphere memory through GPU parallel sorting;

[0115] S6 performs physical conservation parameter aggregation on the Gaussian spheres of the same type stored contiguously in the video memory, releasing redundant video memory and simplifying the three-dimensional Gaussian sphere set.

[0116] S7 completes all training iterations and exports the final simplified 3D Gaussian sphere set as a universal PLY format file.

[0117] Furthermore, in step S1, multi-view images of the accident scene are acquired, camera parameters are calculated, and a three-dimensional Gaussian sphere scene is initialized. The specific method is as follows:

[0118] S1.1, Use drones or handheld devices to collect video footage of traffic accident scenes, extract keyframes from the collected video, and obtain multi-view two-dimensional image sequences. ,in, , Total number of images;

[0119] S1.2, for the input two-dimensional image sequence Perform feature point detection and extract the first feature point. Two-dimensional feature point coordinates of all feature points on the image plane ,in For the first Local index of feature points within an image;

[0120] S1.3 calculates the similarity of feature points in different images using a feature matching operator, and finds two-dimensional feature points belonging to the same real space point observed from different viewpoints based on the rule of equal similarity. These two-dimensional feature points corresponding to the same real-space point are combined together to form a set of common-view feature points unique to that real-space point. ,in, This is the global index of the real-world point. , For each set, the total number of real-world points calculated in the scene is represented. It contains two-dimensional feature points of the spatial point on multiple different images; denoted as... For including the first The set of image indices for real points, then for any image indices... , All of them contain the first The image corresponding to the first Feature points of a real point ;

[0121] S1.4, set all extracted common-view feature points Input Structure from Motion (SfM) algorithm, perform Bundle Adjustment (BA) optimization, and output each frame of image. Corresponding camera intrinsic parameter matrix Rotation matrix Translation vector ;

[0122] S1.5, Extract Set pixel coordinates Based on the corresponding camera parameters The two-dimensional feature points on each frame of the image are projected as a ray in three-dimensional space. Due to image noise and camera parameter errors, the rays will not intersect perfectly. A reprojection error function is constructed as follows:

[0123]

[0124] in, For perspective projection functions, For the first The three-dimensional coordinates of a real-world point to be solved. For including the first A set of image indices for real points;

[0125] S1.6, Minimize the error function using the nonlinear least squares method. Solve and output the optimal 3D intersection point of multiple rays. That is, the coordinates of the real object in three-dimensional space, and the RGB color values ​​of the corresponding original pixels in the image are extracted. Combined with coordinates, output the initial sparse 3D point cloud set. ;

[0126] S1.7, using sparse 3D point cloud sets As the input basis, each real space point This generates a corresponding three-dimensional Gaussian sphere, thus generating a set of three-dimensional Gaussian spheres. ,in, The total number of three-dimensional Gaussian spheres, initially During training, the dynamic increase in cloning and splitting operations affects the set. Each three-dimensional Gaussian sphere The following initial parameter assignment operations are performed in parallel:

[0127] Assign center coordinates : , which represents the physical center position of the Gaussian sphere in three-dimensional space;

[0128] Calculate the spherical harmonic coefficients : Color values Perform inverse spherical projection transformation, output and assign initial spherical harmonic coefficients. , used to represent the diffuse color of a Gaussian sphere;

[0129] Assigning opacity : , indicating the transparency of the Gaussian sphere;

[0130] Assignment rotation quaternion : This indicates that the Gaussian sphere initially has no rotation;

[0131] Calculate scaling Calculate the current center point To point cloud collection The Euclidean distance between the three nearest neighbors is taken as the average of these three distances. Assign the initial scaling vector to... This indicates that the Gaussian sphere is in , , Initial size along the three axes.

[0132] Furthermore, in step S2, a rendering training loop is constructed based on the initialized 3D Gaussian sphere scene, and the Gaussian sphere density is adaptively adjusted through cloning or splitting operations, as follows:

[0133] S2.1, in each training iteration, first from all One real photograph is randomly selected as the current training frame, and the corresponding camera intrinsic parameter matrix is ​​extracted. With external references ;

[0134] S2.2, set the three-dimensional Gaussian spheres All three-dimensional Gaussian spheres The world coordinates are converted to relative coordinates with the current camera lens as the origin. Then, three-dimensional Gaussian spheres are selected that fall within the camera's view frustum, have a Z-axis depth distance between 0.1 meters and 1000 meters in the camera coordinate system, an opacity of not less than 0.001, and a size of not less than 1 pixel after projection onto the image plane. These three-dimensional Gaussian spheres are then used to form the set of visible three-dimensional Gaussian spheres in the current frame's image viewpoint. ;

[0135] S2.3, set The center coordinates, scaling factor, and rotation quaternion of each 3D Gaussian sphere are projected onto the 2D image plane of the current frame. Alpha-blending is then performed according to the depth along the Z-axis of the camera coordinate system from far to near, rendering and outputting the 2D predicted image of the current frame. ;

[0136] S2.4, Extract the two-dimensional prediction image of the current frame. Corresponding real two-dimensional image Calculate the pixel-by-pixel absolute error between the two. and structural similarity loss That is, SSIM loss, construct the joint loss function:

[0137]

[0138] In the formula, Preset weighting coefficients;

[0139] S2.5, the calculated global loss Perform gradient backpropagation to obtain the gradient values ​​of all physical parameters of each visible Gaussian sphere. ,in, Represents any parameter, including center coordinates. spherical harmonic coefficient Opacity Rotation Quaternions Scaling The gradient of an invisible Gaussian sphere is 0, and the parameters remain unchanged.

[0140] S2.6, based on the Adam gradient descent optimizer, adjusts the above parameters. , representing any parameter to be optimized , , , and Perform fully parallel synchronous updates:

[0141] For each class of trainable parameters, maintain two independent internal state variables, where the first moment estimate... Used to accumulate historical gradient information, smooth the parameter update process, and accelerate convergence; second-order moment estimation This is used to accumulate historical gradient squared information and adaptively adjust the learning rate for each parameter.

[0142] First-moment estimation of parallel parameter updates:

[0143] Second-moment estimation of parallel parameter updates:

[0144] Parallel bias correction is applied to the first and second moments of the parameters to eliminate statistical bias caused by initial values ​​of 0.

[0145] ,

[0146] The final update of all parameters is calculated in parallel:

[0147]

[0148] Perform iterative updates on all original parameters: ;

[0149] In the formula, This represents the current training iteration number. , This is an internal state variable of the Adam optimizer, used only for parameter update calculations, and is unrelated to the 3D Gaussian sphere index. , These are the first and second moment decay coefficients of the Adam optimizer; The learning rate; It is the numerical stability constant;

[0150] S2.7, for the updated rotation quaternion Perform a normalization operation to ensure that its modulus is 1, in order to maintain normal rotation, as shown below:

[0151]

[0152] S2.8 updates and overwrites the 3D Gaussian sphere set with all updated parameters. The corresponding visible set Gaussian sphere parameters;

[0153] S2.9, after each fixed number of iterations, calculate the field-of-view position gradient vector of the spatial center coordinates of each Gaussian sphere. And find its second norm. Compare it with the set position gradient threshold Compare and determine:

[0154] when At that time, based on the scaling scale of the corresponding Gaussian sphere The size of the 3D Gaussian sphere is determined, and corresponding memory topology operations are performed: if the volume is smaller than the scale threshold, a cloning operation is performed to generate a new 3D Gaussian sphere with identical parameters, which is then superimposed at the same location to enhance the reconstruction details; if the volume is larger than the scale threshold, a splitting operation is performed, cutting the sphere into two smaller sub-3D Gaussian spheres along the direction of the largest operation error, which are then fitted with details from different regions respectively; the newly generated 3D Gaussian spheres are then merged into a set. ;

[0155] Repeat steps S2.1 to S2.9 until the current iteration. Reaching the preset clustering trigger round .

[0156] Furthermore, in step S3, the spatial and color features of all three-dimensional Gaussian spheres are extracted to construct a multimodal physical feature set. The specific method is as follows:

[0157] S3.1, when training iteration step Reaching the preset clustering trigger round At this point, global gradient backpropagation is paused during rendering, where a unique integer index for all Gaussian spheres is given. , , The total number of current 3D Gaussian spheres, using this global index. For addressing indexing, traverse the set of three-dimensional Gaussian spheres in the current video memory. Extract and output the three-dimensional spatial center coordinate vector of the entire three-dimensional Gaussian sphere. ;

[0158] S3.2, based on the exact same global sequence number Synchronous traversal and synchronous extraction of the collection Spherical harmonic coefficients corresponding to the Gaussian sphere Extract its zeroth-order spherical harmonic coefficient components Perform a linear transformation operation in the color space, calculate and output the diffuse color vector of the corresponding Gaussian sphere. :

[0159]

[0160] S3.3 Extract the spatial center coordinate vector obtained from the above calculation. With diffuse color vector Perform tensor concatenation operation along the channel dimension to output the multimodal physical feature vector of the single Gaussian sphere. :

[0161]

[0162] S3.4, integrating all Generate and output a global multimodal physical feature set from the multimodal physical feature vectors of a three-dimensional Gaussian sphere. , which serves as the input basis for subsequent discrete quantization and mapping algorithms.

[0163] Furthermore, in step S4, an adaptive discrete step size is calculated based on the multimodal physical feature set, mapping continuous features to discrete coordinates. The specific method is as follows:

[0164] S4.1 Extract the set of 3D Gaussian spheres in the current video memory. The global bounding box, i.e., the coordinates of the centers of all Gaussian spheres. Calculate the bounding box of the cuboid formed by the maximum and minimum values ​​of . , , The coordinate ranges along the three axes are taken, and the maximum value of the three ranges is multiplied by a preset spatial resolution scaling factor to calculate and output the base spatial deviation range. ;

[0165] S4.2, from the global multimodal physical feature set Extract all A diffuse color vector of a three-dimensional Gaussian sphere Calculate the global mean of all colors in the field. and global standard deviation :

[0166]

[0167] The current iteration step is calculated using a linear time decay strategy. Corresponding time decay factor and preset lower limit of photometric step size Perform the following scalar operations to calculate and output the dynamic photometric distance from the walk. :

[0168]

[0169] In the formula, This represents the current training iteration number. This is the preset total number of training iterations; The initial attenuation coefficient is 3.0. The preset minimum luminosity distance is set to 0.01.

[0170] S4.3, by serial number Using the index, traverse the set of three-dimensional Gaussian spheres in the current video memory. Parallel extraction of multimodal physical feature vectors for each three-dimensional Gaussian sphere Extracted spatial coordinates With color vector Divide by the corresponding distance from the walk length respectively and And perform the floor operation:

[0171]

[0172] The two sets of discrete coordinates calculated are concatenated by channel to output the multimodal discrete coordinate vector corresponding to the three-dimensional Gaussian sphere. Integrate all data to output a global discrete coordinate set. .

[0173] Furthermore, in step S5, a multimodal hash index is generated based on discrete coordinates, and the same type of Gaussian sphere memory is continuously rearranged through GPU parallel sorting. The specific method is as follows:

[0174] S5.1, by serial number Use the index to traverse the set of discrete coordinates Extract the multimodal discrete coordinate vectors of each three-dimensional Gaussian sphere. Introduce a pre-defined set of independent large prime numbers. With global hash table capacity parameter Perform component multiplication, component XOR, and global modulo operations sequentially on the discrete coordinate components:

[0175]

[0176] In the formula, This indicates that vectors are multiplied by their components. Represents the XOR operation on vector components; a set of independent large prime numbers. It consists of 6 distinct large prime numbers, with values ​​{73856093, 19349663, 83492791, 1299709, 3162277, 5732533}. The global hash table capacity parameter is... Take the integer power of 2, and the value is... or ;

[0177] Calculate and output the one-dimensional scalar hash key corresponding to the three-dimensional Gaussian sphere. The integrated calculation results output a global hash index set. ;

[0178] S5.2 invokes the graphics processor's parallel radix sort algorithm, using the hash keys of all three-dimensional Gaussian spheres. Using the sort key, the memory addresses of all 3D Gaussian spheres in video memory are globally rearranged in ascending order; after rearrangement, an ordered hash set corresponding one-to-one with the video memory storage order is output. ,in, Local index of the elements in this set This corresponds one-to-one with the storage order of the Gaussian spheres in the sorted video memory. After rearrangement, the global three-dimensional Gaussian sphere set is complete. They have been stored contiguously in video memory in ascending order of hash value, so that Gaussian balls with the same hash value form contiguous memory blocks;

[0179] S5.3, Extracting the Ordered Hash Set Perform parallel prefix sum scan operations on it to calculate the difference determination identifier between adjacent hash key values. :

[0180]

[0181] in, At that time, the default It is infinite, that is ;

[0182] When the difference identifier When the cluster boundary marker is triggered: three-dimensional Gaussian spheres with the same hash key value form a cluster, and this position is the starting offset address of a new cluster in video memory;

[0183] The total number of clusters obtained from this clustering is obtained by summing all the difference labels:

[0184]

[0185] Then, calculate and output the length set of all clusters based on the address difference between two adjacent boundary markers. That is, the first The number of three-dimensional Gaussian spheres contained in each cluster.

[0186] Furthermore, in step S6, for all Each cluster performs a parallel merging operation, independently generating a new conserved 3D Gaussian sphere for each cluster, releasing redundant GPU memory and simplifying the 3D Gaussian sphere set. The specific method is as follows:

[0187] S6.1, for the first Extracting clusters within a cluster A set of opacity parameters for a three-dimensional Gaussian sphere Merge the transparency parameters, where, For cluster local indexes, Based on the Beer-Lambert law, logarithmic probabilistic aggregation is performed to calculate and output the aggregated opacity scalar. :

[0188]

[0189] S6.2 Extract the set of rotation quaternions of the 3D Gaussian sphere within the cluster. The rotation angle parameters were merged, and the quaternion of the three-dimensional Gaussian sphere with the highest opacity within the cluster was selected as the alignment reference. Introducing symbolic functions Perform hemispherical manifold correction and normalization operations, calculate and output the aggregated rotation quaternion vector. :

[0190]

[0191] S6.3 Extract the set of spatial center coordinates of the three-dimensional Gaussian sphere within the cluster. Diffuse color set With scaling scale set Merge coordinate, color, and size parameters, and adjust the corresponding opacity. For each weighting factor, perform a linear weighted average operation:

[0192]

[0193]

[0194]

[0195] Calculate and output the coordinates of the aggregation center. Aggregate diffuse color With aggregation scaling scale To aggregate colors Perform reverse color transformation Output the updated spherical harmonic coefficients ;

[0196] S6.4, Combine the new physical parameters obtained from the above calculations. Write the starting physical address of the current cluster in the video memory, overwrite the original first Gaussian sphere parameter at that address, and generate a single conserved three-dimensional Gaussian sphere.

[0197] S6.5, the consecutive addresses following the starting address of this cluster in video memory. One address space is marked as free and forcibly released; the remaining addresses within the cluster are deleted. Storage space occupied by the old three-dimensional Gaussian sphere;

[0198] S6.6, Update the total number of global 3D Gaussian spheres. The memory address offset of all subsequent clusters is adjusted, the gradient backpropagation pause state of training is released, and the updated simplified three-dimensional Gaussian sphere set is used to return to step S2 to continue iterative adjustment. Each time the iteration step reaches the set iteration round, a cluster merging operation is automatically triggered until the preset maximum training round is reached.

[0199] Furthermore, in step S7, after completing all training iterations, the final simplified 3D Gaussian sphere set is exported as a universal PLY format file, as follows:

[0200] S7.1 When the number of iterations reaches the preset maximum number of rounds, stop all gradient updates and clustering operations, and freeze the parameters of all current 3D Gaussian spheres;

[0201] S7.2, Traverse all the final simplified versions in the video memory For each 3D Gaussian sphere, its center coordinates, scaling factor, rotation quaternion, opacity, and spherical harmonics are extracted sequentially. Following the order of these parameters and strictly adhering to the vertex attribute format of the standard PLY file, all parameters of each Gaussian sphere are organized into a single vertex attribute data entry.

[0202] S7.3, create an empty PLY file, and define the total number of vertices of the 3D Gaussian sphere in the PLY file. And the type and order of all physical properties corresponding to each vertex;

[0203] S7.4, organize The vertex attribute data of each Gaussian sphere is continuously written into the vertex data segment of the PLY file, and exported as the final complete 3D reconstruction PLY model file of the accident scene. The PLY model file is then input into the SIBR_viewer streaming software for 3D reconstruction.

[0204] Table 1. Comparison of the 3DGS method and the method of this invention.

[0205]

[0206] This embodiment uses the CO3D dataset as the test sequence, which includes three car scenes, denoted as Scene 1 to Scene 3. The rendering performance and resource overhead of the method provided in this invention are compared with existing standard 3DGS methods. The evaluation metrics are: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Perceptual Similarity (LPIPS), number of Gaussian spheres, and training time. Higher PSNR and SSIM values ​​indicate higher rendered image quality, while lower LPIPS values ​​indicate smaller perceptual errors. Lower values ​​for the number of Gaussian spheres and training time indicate better geometric compactness and lower computational and storage overhead.

[0207] As shown in Table 1, compared with existing standard 3DGS methods, this invention achieves significant compression of the number of Gaussian spheres and a substantial reduction in training time while maintaining or even slightly improving rendering quality (excellent performance in PSNR and SSIM metrics). This verifies that the method effectively eliminates redundant point clouds, greatly optimizes storage space, and improves engineering rendering efficiency while ensuring high-quality new perspective synthesis.

Claims

1. A Gaussian sphere clustering method for 3D reconstruction of highway traffic accident scenes, characterized in that, The method includes the following steps: S1: Acquire multi-view images of the accident scene, calculate camera parameters, and initialize a 3D Gaussian sphere scene; S2 constructs a rendering training loop based on an initialized 3D Gaussian sphere scene, and adaptively adjusts the Gaussian sphere density through cloning or splitting operations. S3, extract the spatial and color features of all three-dimensional Gaussian spheres and construct a multimodal physical feature set; S4, based on the multimodal physical feature set, calculates the adaptive discrete step size, mapping continuous features to discrete coordinates; S5 generates a multimodal hash index based on discrete coordinates and achieves continuous rearrangement of similar Gaussian sphere memory through GPU parallel sorting; S6 performs physical conservation parameter aggregation on the Gaussian spheres of the same type stored contiguously in the video memory, releasing redundant video memory and simplifying the three-dimensional Gaussian sphere set. S7 completes all training iterations and exports the final simplified 3D Gaussian sphere set as a universal PLY format file for 3D reconstruction.

2. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 1, characterized in that, In step S1, multi-view images of the accident scene are acquired, camera parameters are calculated, and a 3D Gaussian sphere scene is initialized. The specific method is as follows: S1.1, Use drones or handheld devices to collect video footage of traffic accident scenes, extract keyframes from the collected video, and obtain multi-view two-dimensional image sequences. ,in, , Total number of images; S1.2, for the input two-dimensional image sequence Perform feature point detection and extract the first feature point. Two-dimensional feature point coordinates of all feature points on the image plane ,in, For the first Local index of feature points within an image; S1.3 calculates the similarity of feature points in different images using a feature matching operator, and finds two-dimensional feature points belonging to the same real space point observed from different viewpoints based on the rule of equal similarity. These two-dimensional feature points corresponding to the same real-space point are combined together to form a set of common-view feature points unique to that real-space point. ,in, This is the global index of the real-world point. , For each set, the total number of real-world points calculated in the scene is represented. It contains two-dimensional feature points of the spatial point on multiple different images; denoted as... For including the first The set of image indices for real points, then for any image indices... , Contains the first The image corresponding to the first Feature points of a real point ; S1.4, set all extracted common-view feature points Input the Structure for Motion (SfM) algorithm, perform bundle adjustment (BA) optimization, and output each frame of image. Corresponding camera intrinsic parameter matrix Rotation matrix Translation vector ; S1.5, Extract Set pixel coordinates Based on the corresponding camera parameters The two-dimensional feature points on each frame of the image are projected as a ray in three-dimensional space. Due to image noise and camera parameter errors, the rays will not intersect perfectly. A reprojection error function is constructed as follows: ； in, For perspective projection functions, For the first The three-dimensional coordinates of a real-world point to be solved. For including the first A set of image indices for real points; S1.6, Minimize the error function using the nonlinear least squares method. Solve and output the optimal 3D intersection point of multiple rays. That is, the coordinates of the real object in three-dimensional space, and the RGB color values ​​of the corresponding original pixels in the image are extracted. Combined with coordinates, output the initial sparse 3D point cloud set. ; S1.7, using sparse 3D point cloud sets As the input basis, each real space point This generates a corresponding three-dimensional Gaussian sphere, thus generating a set of three-dimensional Gaussian spheres. ,in, The total number of three-dimensional Gaussian spheres, initially During training, the dynamic increase in cloning and splitting operations affects the set. Each three-dimensional Gaussian sphere The following initial parameter assignment operations are performed in parallel: Assign center coordinates : , which represents the physical center position of the Gaussian sphere in three-dimensional space; Calculate the spherical harmonic coefficients : Color values Perform inverse spherical projection transformation and output the initial spherical harmonic coefficients. , used to represent the diffuse color of a Gaussian sphere; Assigning opacity : , indicating the transparency of the Gaussian sphere; Assignment rotation quaternion : This indicates that the Gaussian sphere initially has no rotation; Calculate scaling Calculate the current center point To point cloud collection The Euclidean distance between the three nearest neighbors is taken as the average of these three distances. Assign the initial scaling vector to... This indicates that the Gaussian sphere is in , , Initial size along the three axes.

3. The Gaussian sphere clustering method for 3D reconstruction of highway traffic accident scenes according to claim 2, characterized in that, In step S2, a rendering training loop is constructed based on the initialized 3D Gaussian sphere scene. The Gaussian sphere density is adaptively adjusted through cloning or splitting operations. The specific method is as follows: S2.1, in each training iteration, first from all One real photograph is randomly selected as the current training frame, and the corresponding camera intrinsic parameter matrix is ​​extracted. With external references ; S2.2, set the three-dimensional Gaussian spheres All three-dimensional Gaussian spheres The world coordinates are converted to relative coordinates with the current camera lens as the origin. Then, three-dimensional Gaussian spheres are selected that fall within the camera's view frustum, have a Z-axis depth distance between 0.1 meters and 1000 meters in the camera coordinate system, an opacity of not less than 0.001, and a size of not less than 1 pixel after projection onto the image plane. These three-dimensional Gaussian spheres are then used to form the set of visible three-dimensional Gaussian spheres in the current frame's image viewpoint. ; S2.3, set The center coordinates, scaling factor, and rotation quaternion of each 3D Gaussian sphere are projected onto the 2D image plane of the current frame. Alpha-blending is then performed according to the depth along the Z-axis of the camera coordinate system from far to near, rendering and outputting the 2D predicted image of the current frame. ; S2.4, Extract the two-dimensional prediction image of the current frame. Corresponding real two-dimensional image Calculate the pixel-by-pixel absolute error between the two. and structural similarity loss That is, SSIM loss, construct the joint loss function: ； In the formula, Preset weighting coefficients; S2.5, the calculated global loss Perform gradient backpropagation to obtain the gradient values ​​of all physical parameters of each visible Gaussian sphere. ,in, Represents any parameter, including center coordinates. spherical harmonic coefficient Opacity Rotation Quaternions Scaling The gradient of an invisible Gaussian sphere is 0, and the parameters remain unchanged. S2.6, based on the Adam gradient descent optimizer, adjusts the above parameters. , representing any parameter to be optimized , , , and Perform fully parallel synchronous updates: For each class of trainable parameters, maintain two independent internal state variables, where the first moment estimate... Used to accumulate historical gradient information, smooth the parameter update process, and accelerate convergence; second-order moment estimation This is used to accumulate historical gradient squared information and adaptively adjust the learning rate for each parameter. First-moment estimation of parallel parameter updates: ; Second-moment estimation of parallel parameter updates: ; Parallel bias correction is applied to the first and second moments of the parameters to eliminate statistical bias caused by initial values ​​of 0. ， ； The final update of all parameters is calculated in parallel: ； Perform iterative updates on all original parameters: ; In the formula, This represents the current training iteration number. , This is an internal state variable of the Adam optimizer, used only for parameter update calculations, and is unrelated to the 3D Gaussian sphere index. , These are the first and second moment decay coefficients of the Adam optimizer; The learning rate; It is the numerical stability constant; S2.7, for the updated rotation quaternion Perform a normalization operation to ensure that its modulus is 1, in order to maintain normal rotation, as shown below: ； S2.8 updates all parameters and overwrites the 3D Gaussian sphere set. The corresponding visible set Gaussian sphere parameters; S2.9, after each fixed number of iterations, calculate the field-of-view position gradient vector of the spatial center coordinates of each Gaussian sphere. And find its second norm. Compare it with the set position gradient threshold Compare and determine: when At that time, based on the scaling scale of the corresponding Gaussian sphere The size of the 3D Gaussian sphere is determined, and corresponding memory topology operations are performed: if the volume is smaller than the scale threshold, a cloning operation is performed to generate a new 3D Gaussian sphere with identical parameters, which is then superimposed at the same location to enhance the reconstruction details; if the volume is larger than the scale threshold, a splitting operation is performed, cutting the sphere into two smaller sub-3D Gaussian spheres along the direction of the largest operation error, which are then fitted with details from different regions respectively; the newly generated 3D Gaussian spheres are then merged into a set. ; Repeat steps S2.1 to S2.9 until the current iteration. Reaching the preset clustering trigger round .

4. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 3, characterized in that, In step S3, the spatial and color features of all three-dimensional Gaussian spheres are extracted to construct a multimodal physical feature set. The specific method is as follows: S3.1, when training iteration step Reaching the preset clustering trigger round At this point, global gradient backpropagation is paused during rendering, where a unique integer index for all Gaussian spheres is given. , , The total number of current 3D Gaussian spheres, using this global index. For addressing indexing, traverse the set of three-dimensional Gaussian spheres in the current video memory. Extract and output the three-dimensional spatial center coordinate vector of the entire three-dimensional Gaussian sphere. ; S3.2, based on the exact same global sequence number Synchronous traversal and synchronous extraction of the collection Spherical harmonic coefficients corresponding to the Gaussian sphere Extract its zeroth-order spherical harmonic coefficient components Perform a linear transformation operation in the color space, calculate and output the diffuse color vector of the corresponding Gaussian sphere. : ； S3.3 Extract the spatial center coordinate vector obtained from the above calculation. With diffuse color vector Perform tensor concatenation operation along the channel dimension to output the multimodal physical feature vector of the single Gaussian sphere. : ； S3.4, integrating all Generate and output a global multimodal physical feature set from the multimodal physical feature vectors of a three-dimensional Gaussian sphere. , which serves as the input basis for subsequent discrete quantization and mapping algorithms.

5. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 4, characterized in that, In step S4, the adaptive discrete step size is calculated based on the multimodal physical feature set, mapping continuous features to discrete coordinates. The specific method is as follows: S4.1 Extract the set of 3D Gaussian spheres in the current video memory. The global bounding box, i.e., the coordinates of the centers of all Gaussian spheres. Calculate the bounding box of the cuboid formed by the maximum and minimum values ​​of . , , The coordinate ranges along the three axes are taken, and the maximum value of the three ranges is multiplied by a preset spatial resolution scaling factor to calculate and output the base spatial deviation range. ; S4.2, from the global multimodal physical feature set Extract all A diffuse color vector of a three-dimensional Gaussian sphere Calculate the global mean of all colors in the field. and global standard deviation : ； The current iteration step is calculated using a linear time decay strategy. Corresponding time decay factor and preset lower limit of photometric step size Perform the following scalar operations to calculate and output the dynamic photometric distance from the walk. : ； In the formula, This represents the current training iteration number. This is the preset total number of training iterations; The initial attenuation coefficient is 3.

0. The preset minimum luminosity distance is set to 0.

01. S4.3, by serial number Using the index, traverse the set of three-dimensional Gaussian spheres in the current video memory. Parallel extraction of multimodal physical feature vectors for each three-dimensional Gaussian sphere Extracted spatial coordinates With color vector Divide by the corresponding distance from the walk length respectively and And perform the floor operation: ； The two sets of discrete coordinates calculated are concatenated by channel to output the multimodal discrete coordinate vector corresponding to the three-dimensional Gaussian sphere. Integrate all data to output a global discrete coordinate set. .

6. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 5, characterized in that, In step S5, a multimodal hash index is generated based on discrete coordinates, and the same type of Gaussian sphere memory is continuously rearranged through GPU parallel sorting. The specific method is as follows: S5.1, by serial number Use the index to traverse the set of discrete coordinates Extract the multimodal discrete coordinate vectors of each three-dimensional Gaussian sphere from it. Introduce a pre-defined set of independent large prime numbers. With global hash table capacity parameter Perform component multiplication, component XOR, and global modulo operations sequentially on the discrete coordinate components: ； In the formula, This indicates that vectors are multiplied by their components. Represents the XOR operation on vector components; a set of independent large prime numbers. It consists of 6 distinct large prime numbers, with values ​​{73856093, 19349663, 83492791, 1299709, 3162277, 5732533}, and the global hash table capacity parameter. Take the integer power of 2, and the value is... or ; Calculate and output the one-dimensional scalar hash key corresponding to the three-dimensional Gaussian sphere. The integrated calculation results output a global hash index set. ; S5.2 invokes the graphics processor's parallel radix sort algorithm, using the hash keys of all three-dimensional Gaussian spheres. Using the sort key, the memory addresses of all 3D Gaussian spheres in video memory are globally rearranged in ascending order; after rearrangement, an ordered hash set corresponding one-to-one with the video memory storage order is output. ,in, Local index of the elements in this set This corresponds one-to-one with the storage order of the Gaussian spheres in the sorted video memory. After rearrangement, the global three-dimensional Gaussian sphere set is complete. They have been stored contiguously in video memory in ascending order of hash value, so that Gaussian balls with the same hash value form contiguous memory blocks; S5.3, Extracting the Ordered Hash Set Perform parallel prefix sum scan operations on it to calculate the difference determination identifier between adjacent hash key values. : ;in, At that time, the default It is infinite, that is ; When the difference identifier When the cluster boundary marker is triggered: three-dimensional Gaussian spheres with the same hash key value form a cluster, and this position is the starting offset address of a new cluster in video memory; The total number of clusters obtained from this clustering is obtained by summing all the difference labels: ； Then, calculate and output the length set of all clusters based on the address difference between two adjacent boundary markers. That is, the first The number of three-dimensional Gaussian spheres contained in each cluster.

7. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 6, characterized in that, In step S6, for all Each cluster performs a parallel merging operation, independently generating a new conserved 3D Gaussian sphere for each cluster, releasing redundant GPU memory and simplifying the 3D Gaussian sphere set. The specific method is as follows: S6.1, for the first Extracting clusters within a cluster A set of opacity parameters for a three-dimensional Gaussian sphere Merge the transparency parameters, where, For cluster local indexes, Based on the Beer-Lambert law, logarithmic probabilistic aggregation is performed to calculate and output the aggregated opacity scalar. : ； S6.2 Extract the set of rotation quaternions of the 3D Gaussian sphere within the cluster. The rotation angle parameters were merged, and the quaternion of the three-dimensional Gaussian sphere with the highest opacity within the cluster was selected as the alignment reference. Introducing symbolic functions Perform hemispherical manifold correction and normalization operations, calculate and output the aggregated rotation quaternion vector. : ； S6.3 Extract the set of spatial center coordinates of the three-dimensional Gaussian sphere within the cluster. Diffuse color set With scaling scale set Merge coordinate, color, and size parameters, and adjust the corresponding opacity. For each weighting factor, perform a linear weighted average operation: ； ； ； Calculate the coordinates of the aggregation center. Aggregate diffuse color With aggregation scaling scale , will aggregate colors Perform reverse color transformation Output the updated spherical harmonic coefficients ; S6.4, Combine the new physical parameters obtained from the above calculations. Write the starting physical address of the current cluster in the video memory, overwrite the original first Gaussian sphere parameter at that address, and generate a single three-dimensional Gaussian sphere. S6.5, the consecutive addresses following the starting address of this cluster in video memory. One address space is marked as free and forcibly released; the remaining addresses within the cluster are deleted. Storage space occupied by the old three-dimensional Gaussian sphere; S6.6, Update the total number of global 3D Gaussian spheres. The memory address offset of all subsequent clusters is adjusted, the gradient backpropagation pause state of training is released, and the updated simplified three-dimensional Gaussian sphere set is used to return to step S2 to continue iterative adjustment. Each time the iteration step reaches the set iteration round, a cluster merging operation is automatically triggered until the preset maximum training round is reached.

8. The Gaussian spherical clustering method for 3D reconstruction of highway traffic accident scenes according to claim 6, characterized in that, In step S7, after completing all training iterations, the final simplified 3D Gaussian sphere set is exported as a general PLY format file. The specific method is as follows: S7.1 When the number of iterations reaches the preset maximum number of rounds, stop all gradient updates and clustering operations, and freeze the parameters of all current 3D Gaussian spheres; S7.2, Traverse all the final simplified versions in the video memory For each 3D Gaussian sphere, extract the center coordinates, scaling factor, rotation quaternion, opacity and spherical harmonic coefficient in sequence. Following the above parameter order and the vertex attribute format specification of the standard PLY file, organize all parameters of each Gaussian sphere into a vertex attribute data. S7.3, create an empty PLY file, and define the total number of vertices of the 3D Gaussian sphere in the PLY file. And the type and order of all physical properties corresponding to each vertex; S7.4, organize The vertex attribute data of each Gaussian sphere is continuously written into the vertex data segment of the PLY file, and exported as the final complete 3D reconstruction PLY model file of the accident scene. The PLY model file is then input into the SIBR_viewer streaming software for 3D reconstruction.

Images