Edge terminal-oriented 3D gaussian splat adaptive distillation compression and updating method

Abstract

The application belongs to the technical field of three-dimensional scene representation, and discloses a 3D Gaussian splat adaptive distillation compression and updating method for edge terminals. The method obtains multi-view image data and camera parameters of a target scene, trains a teacher three-dimensional Gaussian splat scene model, generates a region importance score and completes Gaussian cell region division, sets corresponding parameters based on the region importance score, performs region adaptive distillation compression to obtain a student model, determines a change region and only performs incremental distillation updating on the corresponding Gaussian cell, and deploys the student model to an edge terminal. The application combines region differentiation distillation and local incremental updating, guarantees the integrity of key regions, reduces the storage and calculation overhead of the three-dimensional Gaussian splat model, adapts to the resource constraint environment of the edge terminal, and improves the model updating efficiency of dynamic scenes.

Description

Technical Field

[0001] This invention belongs to the field of 3D scene representation technology and discloses a 3D Gaussian splash adaptive distillation compression and update method for edge terminals. Background Technology

[0002] 3D Gaussian splashing, as an explicit 3D scene representation method, is widely used in 3D reconstruction and new perspective synthesis due to its advantages of fast training and efficient rendering. However, existing 3D Gaussian splashing models have a large number of Gaussian primitives and high parameter precision, making them unsuitable for direct deployment on edge devices with limited computing power and storage. Existing compression methods all employ a globally uniform compression strategy, failing to differentiate the importance of different scene regions, leading to geometric blurring and texture loss in key areas. Furthermore, local scene changes require a full model update, which is time-consuming and resource-intensive, failing to meet the needs of real-time deployment and continuous operation of dynamic scenes on edge devices.

[0003] Based on the above problems, there is an urgent need for a three-dimensional Gaussian splash compression technology that balances the preservation of key areas, lightweight model, and rapid local updates. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies and propose a 3D Gaussian splash adaptive distillation compression and update method for edge terminals, including: S1: acquiring multi-view image data and camera parameters of the target scene, and training a teacher's 3D Gaussian splash scene model based on the multi-view image data and camera parameters; S2: Generate a region importance score based on reprojection error, view coverage, texture variation, and task attention. Divide the Gaussian primitives in the teacher's 3D Gaussian splash scene model into regions based on the region importance score. S3: For the different regions after division, based on the regional importance score, the Gaussian meta-retention ratio, attribute quantization accuracy and distillation loss weight are set respectively. The region adaptive distillation compression is performed on the teacher's 3D Gaussian splash scene model to obtain the student's 3D Gaussian splash scene model. S4: Compare the newly added observation data with the historical rendering data of the student's 3D Gaussian splash scene model to determine the changed areas of the target scene, and only perform incremental distillation updates on the Gaussian elements corresponding to the changed areas. S5: Deploy the student's 3D Gaussian splash scene model to the edge terminal.

[0005] Preferably, the regional importance score satisfies: ; Score the regional importance of the r-th region. The normalized reprojection error index. As a normalized perspective coverage indicator, This is a normalized texture variation index. To normalize the task attention index. The reprojection error weighting coefficient is used. For view coverage weighting coefficient, This is the texture variation weighting coefficient. This is the task attention weighting coefficient. .

[0006] Further optimized, the normalization processing of reprojection error, view coverage, texture variation, and task attention satisfies: , The index value is the normalized value. These are the original index values. This represents the minimum value of the original metric in the target scenario. This represents the maximum value of the original metric in the target scenario. This is the stable term in the denominator.

[0007] Further preferred, the Gaussian byte retention ratio in the r-th region satisfies: ; The percentage of Gaussian elements retained in the r-th region. To preset the minimum retention ratio, To preset the maximum retention ratio, Score the regional importance of the r-th region.

[0008] Further optimized, the quantization bit width corresponding to the attribute quantization precision of the r-th region satisfies: ; Let r be the quantization bit width of the r-th region. To preset the minimum quantization bit width, To preset the maximum quantization bit width, Score the regional importance of the r-th region.

[0009] Further preferred, the distillation loss used in the region adaptive distillation compression includes rendering consistency loss, geometric consistency loss, attribute consistency loss and regularization loss. The distillation loss is the weighted sum of the distillation loss weights of different regions and the rendering consistency loss, geometric consistency loss, attribute consistency loss of the corresponding regions, and the superimposed regularization loss.

[0010] Further optimization involves determining the area of ​​change based on a change score, where the change score satisfies the following: , Score the changes in the r-th region. For reprojection differences, For texture differences, For task semantic differences, The reprojection difference weighting coefficient, This represents the texture difference weighting coefficient. This represents the semantic difference weighting coefficient for the task.

[0011] In a further preferred embodiment, the incremental distillation update includes performing addition, correction, replacement, and deletion operations on Gaussian elements corresponding to the changed regions, while maintaining the parameters of the student 3D Gaussian splash scene model corresponding to the unchanged regions unchanged during the incremental distillation update process.

[0012] Further preferred methods for dividing the Gaussian primitives in the teacher's 3D Gaussian splash scene model include regular grid spatial partitioning and semantic region partitioning.

[0013] Further optimized, the deployment of the student 3D Gaussian splash scene model includes loading model parameters in blocks according to the regional importance score, with the loading order being high importance regional model parameters, medium importance regional model parameters, and low importance regional model parameters.

[0014] Technical Effects: This invention uses regional importance scoring as the core criterion and adopts a differentiated distillation compression strategy instead of global uniform compression to avoid detail distortion in key scene areas. At the same time, it replaces the traditional full model update method with incremental distillation updates only for changing areas, effectively solving the technical problems of existing 3D Gaussian splash models being unable to adapt to edge terminals, insufficient fidelity in key areas, and low efficiency in dynamic scene updates. While ensuring the quality of 3D scene representation, it reduces model storage and computational overhead, improves the adaptability of edge terminal deployment, and enhances the model maintenance efficiency of dynamic scenes. Attached Figure Description

[0015] Figure 1 This is a flowchart of the 3D Gaussian splashing adaptive distillation compression and update method for edge terminals according to the present invention; Figure 2 This is a connection block diagram of the 3D Gaussian splashing adaptive distillation compression and update system for edge terminals according to the present invention. Detailed Implementation

[0016] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Existing technologies use global compression without partitioning, and model updates are performed on a full scale, resulting in poor adaptability to edge terminals.

[0018] Based on this, please refer to Figures 1-2This embodiment provides a 3D Gaussian splash adaptive distillation compression and update method for edge terminals. It acquires multi-view image data and camera parameters of the target scene. The multi-view image data is a sequence of two-dimensional images of the target scene under different spatial poses. The acquisition devices include industrial cameras, vehicle cameras, AR glasses acquisition modules, and mobile robot vision sensors. The acquisition process completes 360-degree circular shooting around the target scene. The shooting path covers all visible surfaces of the target scene. Each frame of the image corresponds to unique spatial pose data. The image format is RGB three-channel format. The pixel size of a single frame image is determined by the hardware parameters of the acquisition device. The image data includes the surface texture, geometric contour, and illumination distribution information of the target scene. Camera parameters include intrinsic and extrinsic parameters. The intrinsic parameters characterize the camera's imaging optical parameters, including focal length, principal point coordinates, and pixel aspect ratio. Focal length is the distance from the optical center of the camera lens to the imaging plane. Principal point coordinates are the pixel coordinates of the center of the imaging plane. The pixel aspect ratio is the ratio of the physical size of a single pixel. The intrinsic parameters are obtained through a camera calibration process using a checkerboard calibration board. Stable intrinsic parameter values ​​are obtained by solving for the calibration board images under different poses. The extrinsic parameters characterize the camera's spatial position and rotation angle in the world coordinate system. They include translation vectors and rotation matrices. The translation vector represents the three-dimensional coordinates of the camera's optical center in the world coordinate system, and the rotation matrix represents the three-dimensional rotation state of the camera's shooting angle. The extrinsic parameters are solved using a motion reconstruction structure, calculated based on feature matching relationships from multi-view images. A 3D Gaussian splash scene model of a teacher was obtained through training using multi-view image data and camera parameters. This model consists of a defined number of Gaussian primitives (GRPs). The initial GRPs are generated by uniformly sampling the 3D point cloud of the target scene; the initial number is determined by the spatial volume of the scene, with larger scenes having more initial GRPs. Each GRP carries a spatial center position parameter, a covariance parameter, a color parameter, an opacity parameter, and a radiation characteristic parameter. The spatial center position parameter represents the GRP's 3D coordinates in the world coordinate system. The covariance parameter is a 3D matrix representing the spatial extension shape and orientation of the GRP. The color parameter is an RGB three-channel value representing the surface appearance color of the GRP. The opacity parameter is a single-channel value representing the light transmission characteristics of the GRP. The radiation characteristic parameter is a high-dimensional feature vector representing the optical radiation properties of the GRP.The training process relies on a differentiable rendering pipeline. The rendering process projects Gaussian primitives in 3D space onto a 2D image plane, generating rendered images from corresponding viewpoints. Error calculations are performed between the rendered images and the acquired multi-view images. Error types include pixel error, gradient error, and structural error. By iteratively adjusting the parameters of the Gaussian primitives, the error between the rendered result of the teacher's 3D Gaussian splash scene model and the multi-view image data is kept within a set range. The iteration terminates when the error value remains unchanged for 10 consecutive iterations. The trained teacher model possesses complete scene geometry and texture representation capabilities and can output high-definition rendered images from any viewpoint. A region importance score is generated based on reprojection error, viewpoint coverage, texture variation, and task attention. Reprojection error is the coordinate deviation of a 3D point projected onto the 2D image plane. The 3D point is the spatial center of the Gaussian primitive. The projection process is completed based on the camera intrinsic and extrinsic matrices. The coordinate deviation is the difference between the projected pixel coordinates and the coordinates of the real image feature points. The difference is calculated using Euclidean distance, and the dimension of the reprojection error is pixels. Viewpoint coverage is the number of times a target scene area is covered by the camera's viewpoint. The statistical process iterates through all multi-view images, marking the frequency of each spatial region being covered by the viewpoint. The frequency value is a non-negative integer, measured in units of "times". Texture variability is the magnitude of pixel value variation within a target scene area. The calculation process selects all corresponding image pixels within the area and calculates the variance of the pixel values. A larger variance value indicates higher texture complexity. Texture variability is measured in grayscale values. Task attention level is the level of attention given to the business task corresponding to the target scene area. The level value is set according to the needs of the business task. In industrial inspection scenarios, the level value for equipment defect areas and instrument reading areas is higher than that for background walls and floors. In digital twin scenarios, the level value for interactive areas and key equipment areas is higher than that for environmental decoration areas. Task attention level is a dimensionless level value. Based on regional importance scores, Gaussian primitives in a 3D Gaussian splash scene model of a teacher are divided into regions. The division operation uses the spatial center position parameters of the Gaussian primitives as the basis, grouping Gaussian primitives with similar spatial positions into the same region. The division boundary is determined based on a spatial coordinate threshold. The difference in the spatial center position coordinates of all Gaussian primitives within the same region must be within a set threshold range. After division, there is no spatial overlap between regions, and all regions cover the entire spatial range of the target scene. For different regions after division, Gaussian primitive retention ratio, attribute quantization precision, and distillation loss weights are set based on the regional importance scores. The Gaussian primitive retention ratio determines the ratio of the number of retained Gaussian primitives in a region to the original number; this ratio is dimensionless and ranges from 0 to 1. The attribute quantization precision determines the encoding bit depth of each Gaussian primitive parameter; the unit of encoding bit depth is bits, and a larger bit depth indicates more detail in the parameter.Distillation loss weights are used to determine the constraint strength of a region during the distillation process. The weight values ​​are dimensionless; the larger the value, the stricter the parameter constraints on the region. Adaptive distillation compression is performed on the teacher's 3D Gaussian splash scene model. The distillation compression process uses the parameters of the teacher's 3D Gaussian splash scene model as a supervision signal to initialize the Gaussian primitives of the student's 3D Gaussian splash scene model. The initial number of Gaussian primitives in the student model is determined based on the sum of the retention ratios of all regions. Initialization parameters are assigned randomly or downsampled from the teacher model parameters. The distillation process generates a rendered image of the student model through forward propagation, calculates the rendering differences, geometric differences, and attribute differences between the student and teacher models, adjusts the constraint strength of these differences based on the distillation loss weights, and updates the Gaussian primitive parameters of the student model through backpropagation. This forward and backpropagation process is iteratively executed until the loss value of the student model converges, resulting in the student's 3D Gaussian splash scene model. The number of Gaussian primitives and parameter storage in the student's 3D Gaussian splash scene model are lower than those in the teacher's 3D Gaussian splash scene model, making it adaptable to the hardware resource limitations of edge terminals. The newly added observation data is compared with the historical rendering data of the student's 3D Gaussian splash scene model. The newly added observation data consists of multi-view image data collected after scene changes, with the acquisition equipment, acquisition process, and initial data acquisition remaining consistent. The historical rendering data consists of the rendering output data of the student's 3D Gaussian splash scene model under the corresponding viewpoints. The historical rendering data is stored in the local storage unit of the edge terminal or the cloud storage server, and the retrieval process is based on the viewpoint identifier of the image to complete the matching. The changed areas of the target scene are determined through data comparison, which includes pixel-level comparison, geometric feature comparison, and semantic feature comparison. Pixel-level comparison calculates the pixel difference between the newly added observation data and the historical rendering data. Geometric feature comparison calculates the difference in spatial structure. Semantic feature comparison calculates the attribute differences of the task-related areas. The changed area is the spatial range in which scene parameters have changed, and the spatial boundary of the changed area is determined based on the difference threshold. Incremental distillation updates are performed only on the Gaussian elements corresponding to the changed areas. The incremental distillation update does not adjust the Gaussian element parameters of the unchanged areas. During the update process, the Gaussian element parameters of the changed areas are loaded, and distillation optimization is re-executed based on the newly added observation data. The optimization process only adjusts the parameters of the changed areas, while the parameters of the unchanged areas remain unchanged in their stored state. The student's 3D Gaussian splash scene model is deployed to an edge terminal, which includes an industrial inspection terminal, a mobile robot, AR glasses, and an in-vehicle edge device. The industrial inspection terminal is used for equipment inspection and defect identification in industrial scenarios, the mobile robot is used for autonomous navigation and environmental perception, the AR glasses are used for augmented reality interaction and scene visualization, and the in-vehicle edge device is used for autonomous driving environment modeling.The deployment process stores the parameters of the student's 3D Gaussian splash scene model in the storage unit of the edge terminal. The storage unit includes flash memory and RAM memory, and the parameter storage format is binary. The processing unit of the edge terminal calls the parameters to complete scene rendering and task execution. The processing unit includes CPU, GPU and embedded neural network processor. The rendering process generates scene images based on the real-time view of the terminal. Task execution includes target recognition, pose estimation, positioning and navigation and defect detection. All operations are completed locally on the edge terminal without cloud data dependence.

[0019] Existing technologies for regional importance assessment lack multi-indicator weighting rules, resulting in low score matching. The proposed regional importance score employs a linear weighted summation method. This method is based on the theoretical design of multi-indicator fusion evaluation, multiplying the values ​​of four independent indicators by their corresponding weights and then summing the results to obtain a single numerical regional importance score. The linear weighted summation calculation logic aligns with the technical principles of comprehensive evaluation of scene regions, transforming feature information from different dimensions into a unified evaluation value without complex mathematical transformations. Those skilled in the art can directly apply the calculation process to complete the score calculation. middle, Score the regional importance of the r-th region. The value is a dimensionless value, ranging from 0 to 1. The larger the value, the higher the importance level of the r-th region of the target scene. The score value is directly used as the core basis for subsequent region division, parameter setting, and distillation weight allocation, without any additional conversion process. The reprojection error weighting coefficient is used. This is a dimensionless value used to allocate the proportion of the impact of reprojection error on the importance of a region. The weight value is set according to the scene reconstruction requirements, and the higher the reconstruction difficulty of the scene, the larger the weight coefficient of the reprojection error. The normalized reprojection error index. The value is a dimensionless value obtained by standardizing the original reprojection error. It is used to reflect the difficulty level of regional reconstruction. The larger the value, the higher the difficulty of 3D reconstruction of the region and the higher its importance. For view coverage weighting coefficient, This is a dimensionless value used to assign the proportion of influence of view coverage on the importance of a region. The higher the demand for multi-view observation, the greater the weight coefficient of view coverage. As a normalized perspective coverage indicator, It is a dimensionless value obtained by standardizing the original view coverage. It is used to reflect the degree to which the area is observed. The larger the value, the higher the frequency of the area being covered by the collected view and the higher the reliability of the scene representation. This is the texture variation weighting coefficient. This is a dimensionless numerical value used to assign the proportion of influence of texture variation on the importance of a region. The higher the texture detail requirement, the larger the weight coefficient of texture variation. This is a normalized texture variation index. It is a dimensionless numerical value, obtained by standardizing the original texture variation. It is used to reflect the complexity of the texture of a region. The larger the value, the more complex the surface texture of the region and the more detailed information needs to be retained. This is the task attention weighting coefficient. This is a dimensionless value used to allocate the proportion of the impact of task attention on the importance of the region. The stronger the business task orientation, the greater the weight coefficient of task attention. To normalize the task attention index. This is a dimensionless numerical value, obtained by standardizing the original task attention level. It is used to reflect the priority of the corresponding business tasks in a region. The larger the value, the more critical the role of the region in the business tasks. The weight coefficients, summed to 1, are set based on normalization evaluation theory to maintain a stable numerical range for regional importance scores and prevent numerical overflow caused by weight overlap. The allocation of weight coefficients can be flexibly adjusted according to different application scenarios, and the same calculation process can be adapted to various scenarios such as industrial inspection, digital twins, vehicle navigation, and AR interaction. The formula derivation relies on the principle of linear algebraic vector weighting, transforming the independent characteristics of the four types of indicators into a single scoring dimension. The derivation process consists of three steps: Step 1 determines the contribution of each type of indicator to regional importance and converts the contribution into corresponding weight coefficients, ensuring that the sum of the weight coefficients is 1; Step 2 standardizes the original indicators to obtain dimensionless normalized indicators, eliminating the influence of dimensions and numerical magnitude; Step 3 multiplies the normalized indicators by their corresponding weights and sums all the product results to obtain the final regional importance score. The derivation process only includes basic arithmetic operations and does not involve complex operations such as matrix factorization, differentiation, or integration. Those skilled in the art can directly complete the calculation by following the derivation steps. The first embodiment selects an industrial inspection scenario, with the scenario objective being equipment defect detection. The reprojection error weight coefficient is set to 0.25, the view coverage weight coefficient to 0.25, the texture variation weight coefficient to 0.2, and the task attention weight coefficient to 0.3. The four normalized indices are 0.6, 0.7, 0.5, and 0.8, respectively. Substituting these values ​​into the formula yields the following results. This value corresponds to the medium-to-high importance area of ​​the target scene and is directly used as the basis for region division and parameter setting. In the second embodiment, an in-vehicle navigation scene is selected, with the scene target being road environment perception. The reprojection error weight coefficient is set to 0.3, the view coverage weight coefficient to 0.3, the texture variation weight coefficient to 0.2, and the task attention weight coefficient to 0.2. The four normalized indices are 0.4, 0.8, 0.3, and 0.5, respectively. Substituting these values ​​into the formula yields the calculated values. This value corresponds to a moderately important region in the target scene and is used to guide the parameter allocation for distillation compression. In the third embodiment, an AR interactive scene is selected, with the scene target being a virtual object overlay. The reprojection error weight coefficient is set to 0.2, the view coverage weight coefficient to 0.2, the texture variation weight coefficient to 0.3, and the task attention weight coefficient to 0.3. The four normalized indices are 0.7, 0.6, 0.8, and 0.9, respectively. Substituting these values ​​into the formula yields the calculated... This value corresponds to the high-importance areas of the target scene, which are allocated a higher primitive retention ratio and quantization bit width.

[0020] Existing technical evaluation indicators have inconsistent dimensions, making them unsuitable for direct scoring calculations. Normalization employs an extreme value standardization calculation method, designed based on data standardization theory. This method transforms indicators with different dimensions and numerical ranges into dimensionless values ​​within the same interval, ensuring equal weight for different indicators in the fusion calculation. It eliminates the interference of dimensional and numerical magnitude differences on the scoring results and is a necessary prerequisite for multi-indicator fusion calculations. middle, The index value is the normalized value. The values ​​are dimensionless and range from 0 to 1. This uniform range allows different indicators to directly participate in the weighted summation calculation, serving as input data for the regional importance score. These are the original index values. The original index values ​​for reprojection error are measured values ​​with dimensions. The original index value is in pixels, ranging from 0 to 50 pixels, representing the coordinate deviation of a 3D point projected onto the image plane. The original index value for view coverage is in times, ranging from 0 to 100 times, representing the frequency of the area being covered by the shooting view. The original index value for texture variation is in grayscale, ranging from 0 to 255 grayscale, representing the complexity of the texture of the area. The original index value for task attention is a dimensionless level value, ranging from 1 to 10 levels, representing the priority of business tasks in the area. This represents the minimum value of the original metric in the target scenario. To obtain a value with the same dimensions as the original index, the minimum value is obtained by traversing the original index values ​​of all regions of the target scene. This value is the global statistical value of the scene, ensuring the global consistency of the normalization result. This represents the maximum value of the original metric in the target scenario. To obtain a value with the same dimensions as the original indicator, the maximum value is obtained by iterating through the original indicator values ​​of all regions in the target scene and filtering. Together they constitute the numerical range of the original indicator. For the stable term in the denominator, It is a dimensionless, extremely small constant with a fixed value of 10⁻⁸, used to avoid and The calculation anomaly occurs when the denominator is 0 when the values ​​are equal. When the original index values ​​are the same across all regions of the scene, the stable denominator term ensures the normal execution of the calculation process and outputs stable normalized values. The calculation process consists of four steps: Step 1 extracts the original index value of the r-th region of the target scene. Step 2: Calculate the original metrics for all regions of the target scene and obtain the minimum value. With the maximum value Step 3: Subtract the minimum value of the original index from the original index value to obtain the relative difference of the index, and use the relative difference as the numerator; Step 4: Subtract the minimum value from the maximum value of the original index, and then add the denominator stability term to obtain the standardized denominator. Divide the numerator and denominator to obtain the normalized index value. The formula derivation relies on the principle of data preprocessing standardization. Through linear transformation, the original index value range is mapped to the interval of 0 to 1, preserving the relative change characteristics of the index. The derivation process is divided into two stages. The first stage eliminates the influence of dimensions, converting the dimensional measured values ​​into relative differences; the second stage normalizes the value range, converting the relative differences into dimensionless values ​​in the interval of 0 to 1. The derivation process only includes basic arithmetic operations of subtraction and division, and can be directly executed in the hardware processing unit without additional computational overhead. The first embodiment selects the reprojection error index, and the original index value... 15 pixels It is 5 pixels. 25 pixels The value is 10⁻⁸. Substituting this into the formula, we get... This value can be directly substituted into the regional importance scoring formula for calculation. The second embodiment selects the viewpoint coverage index, with the original index value... For 60 times, For 20 times, For 80 times, The value is 10⁻⁸. Substituting this into the formula, we get... After normalization, the numerical values ​​indicate that the observation sufficiency of this region is at a moderately high level. The third embodiment selects the texture variation index, with the original index value... It is a grayscale value of 180. The grayscale value is 30. It is a grayscale value of 210. The value is 10⁻⁸. Substituting this into the formula, we get... After normalization, the numerical value indicates that the texture complexity of this region is relatively high. The fourth embodiment selects a task attention index, with the original index value... It is level 8. Level 2 Level 10 The value is 10^-8. Substituting this into the formula, we get... The normalized values ​​indicate that the business priority in this region is relatively high.

[0021] Existing technologies use a globally fixed primitive retention ratio, resulting in insufficient primitives in important regions. The Gaussian primitive retention ratio is calculated using a linear mapping method based on differentiated resource allocation theory. This method ensures a positive correlation between the retention ratio and the region's importance score. While controlling the total number of Gaussian primitives in the model, it allocates more Gaussian primitives to important regions, guaranteeing the integrity of scene representation in these regions. Simultaneously, it reduces primitive redundancy in non-important regions, lowering the model's storage and computational overhead. middle, The percentage of Gaussian elements retained in the r-th region. It is a dimensionless numerical value, and its range is [range missing]. to The proportional value is directly used to determine the number of Gausky units retained in the region. The number of units retained is the product of the original number of units and the retention ratio. The calculation result is rounded down to the nearest integer to ensure the rationality of the number of units. To preset the minimum retention ratio, It is a dimensionless value, representing the minimum proportion of Gaussian primitives retained in all regions. Its value is fixed at 0.3 to ensure the basic expressive power of the scene and to prevent scene fragmentation due to insufficient primitives in non-important regions. To preset the maximum retention ratio, It is a dimensionless value that retains the highest proportion of Gaussian elements in all regions. Its value is fixed at 0.9 to meet the detail expression needs of important regions and ensure that the geometric and texture details of important regions are not lost. Score the regional importance of the r-th region. This is a dimensionless numerical value, ranging from 0 to 1, calculated using the regional importance scoring formula, and serves as the core adjustment factor for the retention ratio. The calculation process consists of four steps, the first step being obtaining the preset minimum retention ratio. Compared with the preset maximum retention ratio Step 2 calculates the difference between the preset maximum retention ratio and the preset minimum retention ratio to obtain the retention ratio fluctuation range; Step 3 multiplies the retention ratio fluctuation range by the regional importance score to obtain the ratio fluctuation value; Step 4 adds the ratio fluctuation value to the preset minimum retention ratio to obtain the Gaussian retention ratio of the target region. The formula derivation relies on the principle of linear function mapping, using the minimum retention ratio as a benchmark, converting the importance score into a ratio fluctuation coefficient, so that regions with higher importance obtain a larger retention ratio. The derivation process is logically intuitive, conforms to the design specifications of hardware parameter mapping, and can be directly integrated into the parameter configuration module of the edge terminal. In the first embodiment, the following settings are provided. It is 0.3. It is 0.9. The value is 0.665. Substituting this into the formula, we get... In this region, 69.9% of the original Gaussian primitives are retained. Of the original 1000 primitives, 699 are retained, thus controlling the total number of primitives while meeting the representation requirements of important regions. (In the second embodiment...) The value is 0.52. Substituting this into the formula, we get... In this region, 61.2% of the original Gausky primitives are retained, out of a total of 800 original primitives, 489 of which are retained. (Example 3) The value is 0.77. Substituting this into the formula, we get... In this region, 76.2% of the original Gausky primitives are retained, out of a total of 1200 original primitives, 915 of which are retained. (Example 4) The value is 0.2. Substituting this into the formula, we get... In this region, 42% of the original Gaussian primitives are retained. Of the original 500 primitives, 210 are retained. The number of primitives in non-critical regions is significantly reduced, thus reducing the model load.

[0022] Existing technologies use a globally uniform quantization bit width, resulting in unreasonable allocation of storage resources. Attribute quantization bit width employs a linear mapping calculation method based on storage resource optimization allocation theory. This method ensures a positive correlation between quantization bit width and regional importance scores, prioritizing the allocation of limited storage bit budget to important regions. This improves the parameter expression accuracy of important regions while reducing bit occupancy in non-important regions, thus optimizing storage resource utilization at edge terminals. middle, Let r be the quantization bit width of the r-th region. The unit is bits, which is used to determine the number of bits for encoding and storing Gaussian parameters. The Gaussian parameters include spatial center position, covariance, color, and opacity. All parameters use a uniform region quantization bit-width encoding method, which is fixed-point quantization with no floating-point precision loss. To preset the minimum quantization bit width, The unit is bits, which is the minimum number of bits used to encode all region parameters. The value is fixed at 8 bits, which is used to reduce the storage footprint of non-critical regions. 8 bits can meet the expression requirements of basic parameters. To preset the maximum quantization bit width, The unit is bits, which is the highest number of bits used to encode all region parameters. The value is fixed at 32 bits to ensure the accuracy of parameter expression in important regions. 32 bits can restore the detailed features of the parameters. Score the regional importance of the r-th region. This is a dimensionless numerical value, ranging from 0 to 1, and serves as the core adjustment factor for the quantization bit width. The calculation process consists of four steps, the first step being obtaining the preset minimum quantization bit width. With preset maximum quantization bit width Step 2 calculates the difference between the preset maximum quantization bit width and the preset minimum quantization bit width to obtain the bit width fluctuation range; Step 3 multiplies the bit width fluctuation range by the region importance score to obtain the bit width fluctuation value; Step 4 adds the bit width fluctuation value to the preset minimum quantization bit width to obtain the attribute quantization bit width of the target region. The calculation result is rounded up to an integer bit to ensure the feasibility of hardware coding. The formula derivation is based on the principle of digital signal coding quantization. Taking the minimum quantization bit width as the benchmark, the importance score is converted into a bit width fluctuation coefficient, so that important regions are allocated higher storage bit widths, which conforms to the embedded system storage design specifications and can be directly executed in the storage module of the edge terminal. In the first embodiment, it is set as follows: It is 8 bits. It is 32 bits. The value is 0.665. Substituting this into the formula, we get... The value is rounded up to 24 bits. The parameters in this region are quantized and encoded using 24 bits to balance parameter precision and storage usage. (In the second embodiment...) The value is 0.52. Substituting this into the formula, we get... Rounded up to 21 bits, the medium-importance region is encoded using 21 bits. (In the third embodiment...) The value is 0.77. Substituting this into the formula, we get... Rounded up to 27 bits, high-importance regions are encoded using 27 bits. (In the fourth embodiment...) The value is 0.2. Substituting this into the formula, we get... The data is rounded up to 13 bits, and low-importance regions are encoded using 13 bits, significantly reducing storage requirements. The quantization encoding process maps continuous parameter values ​​to discrete encoding intervals. The encoding intervals for important regions are finer, resulting in higher parameter restoration accuracy, while the encoding intervals for non-important regions are sparser, resulting in lower storage requirements. The encoded parameters for all regions can be decoded and restored normally without information loss.

[0023] Existing distillation loss techniques lack regional weights, resulting in insufficient constraint strength in key regions. Region-adaptive distillation compression employs a distillation loss system that includes rendering consistency loss, geometric consistency loss, attribute consistency loss, and regularization loss. These four types of losses collectively constitute the objective function for distillation optimization, ensuring comprehensive consistency between the student and teacher models by eliminating any missing loss type. Rendering consistency loss characterizes the difference in rendered images of the student and teacher 3D Gaussian splash scene models from the same viewpoint. The calculation process selects the same camera viewpoint, generates rendered images of both the teacher and student models, and calculates the mean square error (MSE) of the pixels in both images. A smaller MSE value indicates higher rendering consistency. Geometric consistency loss characterizes the difference in spatial depth distribution between the student and teacher 3D Gaussian splash scene models. The calculation process obtains depth values ​​at the same spatial location and calculates the mean absolute error of the depth values. A smaller error value indicates higher geometric consistency. The attribute consistency loss characterizes the differences in Gaussian meta-parameter color and opacity parameters between the student's and teacher's 3D Gaussian splash scene models. The calculation involves subtracting the corresponding Gaussian meta-parameter attribute parameters to obtain the sum of squared differences; a smaller sum of squares indicates higher attribute consistency. The regularization loss suppresses overfitting during the distillation process. It is calculated using the L2 norm of the student model's Gaussian meta-parameter parameters; a smaller norm indicates more stable parameters without drastic fluctuations. The distillation loss is a weighted sum of the distillation loss weights for different regions, along with the corresponding rendering consistency loss, geometric consistency loss, and attribute consistency loss, plus the regularization loss. The distillation loss weights are positively correlated with the region importance score; regions with higher importance are assigned larger distillation loss weights. These weights are directly mapped from the region importance score without any additional transformation process. The distillation process aims to minimize distillation loss. It iteratively adjusts the Gaussian meta-parameters of the student's 3D Gaussian splash scene model. The iteration step size is dynamically adjusted based on the loss value; a larger loss value results in a larger iteration step, and a smaller loss value results in a smaller iteration step. The iteration terminates when the change in loss value is below a set threshold for five consecutive iterations. At this point, the student model's output is consistent with the teacher model, meeting the scene representation requirements. In this embodiment, the distillation loss weight is set to 1.2 for high-importance regions, 1.0 for medium-importance regions, and 0.8 for low-importance regions. The weighted summation is then combined with a regularized loss to obtain the total distillation loss. Parameters are adjusted based on this total distillation loss, maintaining parameter consistency within a set range for high-importance regions and reducing computational load in non-critical regions while ensuring basic consistency. The distillation process is executed in the embedded processor of the edge terminal. The processor calls the model parameters from the storage unit to complete the entire process of forward propagation, loss calculation, and backpropagation. It has no external computing resource dependency, and the distillation time is lower than the global distillation method, adapting to the real-time requirements of edge terminals.

[0024] Existing technologies lack quantitative standards for determining change areas, resulting in low positioning accuracy. The change scoring method employs a linear weighted summation calculation based on multi-dimensional difference fusion theory. This method numerically represents the degree of change in a scene area, transforming geometric differences, texture differences, and task semantic differences into a single score value. This enables automatic determination of change areas without human intervention, resulting in objective and stable judgments. middle, Score the changes in the r-th region. The value is a dimensionless value. The larger the value, the greater the degree of change in the region. When the change score is higher than the set threshold, the region is determined to be a changed region. When it is lower than the threshold, it is determined to be an unchanged region. The threshold value is set according to the stability requirements of the scene. The reprojection difference weighting coefficient, This is a dimensionless value used to allocate the proportion of the impact of reprojection differences on the change score; the weight coefficient is larger for scenes that are sensitive to changes in geometric structure. For reprojection differences, The value is a dimensionless value representing the difference in reprojection error between newly observed data and historical rendering data. The difference value is obtained through normalization and represents the degree of change in the region's geometric structure. This represents the texture difference weighting coefficient. This is a dimensionless numerical value used to allocate the proportion of influence of texture differences on the change score. Scenes that are sensitive to texture changes have a larger weight coefficient. For texture differences, This is a dimensionless numerical value, representing the difference in texture variation between newly observed data and historical rendering data, characterizing the degree of change in the surface texture of the region. The semantic difference weighting coefficient for the task. This is a dimensionless numerical value used to allocate the proportion of the impact of semantic differences in tasks on the change score. The larger the weight coefficient, the more sensitive the business task is to changes. For task semantic differences, The value is a dimensionless numerical value, representing the difference in task attention between newly observed data and historical rendering data, characterizing the degree of change in regional business attributes. The calculation process consists of three steps: Step 1 multiplies the three types of difference indicators by their corresponding weight coefficients to obtain a weighted difference value; Step 2 sums the three weighted difference values ​​to obtain a regional change score; Step 3 compares the change score with a set threshold to determine whether the region is a changed region. The formula derivation relies on the multi-feature fusion evaluation principle, integrating three types of change features: scene geometry, texture, and task, and fusing difference information in a weighted manner so that the change score can comprehensively reflect the regional change status. The derivation process only includes basic multiplication and addition operations and can be executed in real time on edge terminals. In the first embodiment, it is set... It is 0.4. It is 0.3. It is 0.3. It is 0.7. It is 0.6. The value is 0.8. Substituting this into the formula, we get... If this value exceeds a set threshold of 0.5, it is determined to be a changed region and is directly used as the basis for incremental update execution. (In the second embodiment) It is 0.2. It is 0.1. The value is 0.1. Substituting this into the formula, we get... If the value is below the threshold, it is determined to be an unchanged area. In the third embodiment, a device replacement occurs in the scenario. It is 0.8. It is 0.7. The value is 0.9. Substituting this into the formula, we get... If the area is identified as a changed region, the incremental update process is initiated.

[0025] Current technology models involve full updates, but recalculation of unchanged areas results in high resource consumption. Incremental distillation updates include four types of operations: adding, correcting, replacing, and deleting Gaussian primitives corresponding to changed areas. These operations cover all types of scene changes and can adapt to various scene changes such as device additions, parameter adjustments, component replacements, and redundant removal. The adding operation adds new Gaussian primitives to the changed area. The process generates new primitives based on the spatial location and texture features of the new scene. The initial parameters of the new primitives are calculated based on the new observation data. The added primitives are incorporated into the parameter system of the student model to match the new spatial structure and texture features of the scene. The correcting operation adjusts the spatial center position, covariance, color, and opacity parameters of existing Gaussian primitives in the changed area. The adjustment process re-optimizes the parameter values ​​based on the new observation data to ensure that the parameters match the changed scene state without parameter bias. The replacing operation replaces invalid Gaussian primitives in the changed area with new ones. Invalid primitives are those whose parameters deviate from the actual scene. The replacement process deletes invalid primitives and generates new primitives to supplement the corresponding spatial locations, ensuring the integrity of the scene representation. The deletion operation removes redundant Gaussian primitives from changed regions. Redundant primitives are those corresponding to scenes that have disappeared. Removal reduces the model's storage and computational load, as there are no extra primitives consuming resources. During incremental distillation updates, the parameters of the student's 3D Gaussian splash scene model corresponding to unchanged regions remain unchanged. The number of Gaussian primitives and their parameters in unchanged regions remain in the state before the distillation update, and they do not participate in parameter adjustment or iterative calculations. The parameters of unchanged regions in the storage unit remain in a read-only state, with no write operations. In the example, if a device is replaced in a certain region of the scene, it is determined to be a changed region. Correction and addition operations are performed on the Gaussian primitives in that region, while the parameters of unchanged regions remain unchanged. The update process takes less time than a full update, reducing the model's computational load and storage usage. Edge terminals can stably run the updated model without any lag, crashes, or other abnormalities. The incremental update process can be executed cyclically. Each time a local change occurs in the scene, an incremental update of the corresponding region is initiated, while unchanged regions remain stable, forming a continuous maintenance mechanism for dynamic scenes, adapting to the needs of long-term operating edge terminals.

[0026] Existing technologies for region division are limited in their simplistic approach and have weak adaptability to different scenarios. The proposed method for region division of Gaussian elements in a 3D Gaussian splash scene model of a teacher includes both regular grid spatial partitioning and semantic region division. These two methods can be used independently and flexibly switched according to scene type without compatibility conflicts. The regular grid spatial partitioning method is based on a 3D Cartesian coordinate system, dividing the spatial range of the target scene into uniformly sized cubic grids. The side length of the cubic grids is set according to the scene's precision requirements: 0.1 meters for industrial scenes, 1 meter for vehicle scenes, and 0.05 meters for AR scenes. Each cubic grid corresponds to an independent region. Gaussian elements whose spatial center falls within the same cubic grid are grouped into the same region. The division process relies solely on spatial coordinates, without semantic information dependency, making it suitable for scenarios with regular structures. The semantic region segmentation method is based on the semantic information of the target scene, dividing the scene into regions of different semantic types, including device regions, background regions, interaction regions, navigation regions, and defect regions. Gaussian primitives belonging to the same semantic type are grouped into the same region. Semantic information is obtained through an image semantic segmentation model, which outputs pixel-level semantic labels and maps them to three-dimensional spatial regions, adapting to business-oriented scenarios. Both segmentation methods use the spatial center position parameters of Gaussian primitives as the core basis. The segmentation operation only adjusts the region affiliation of Gaussian primitives without changing their own parameters. There is no data loss during the segmentation process, and the segmentation results can be directly used for subsequent parameter settings. In the first embodiment, the indoor regular scene uses a regular grid spatial block method with a grid side length of 0.1 meters, resulting in 100 uniform regions, with a balanced number of primitives in each region. In the second embodiment, the industrial inspection scene uses a semantic region segmentation method, dividing the scene into device regions, background regions, and defect regions, with higher importance scores for the device and defect regions. In the third embodiment, the vehicle navigation scene uses a regular grid spatial block method with a grid side length of 1 meter, adapting to the segmentation needs of large-scale spaces. Neither of the two classification methods requires creative effort, and those skilled in the art can directly choose to use them based on the characteristics of the scenario.

[0027] The existing technology model loads as a whole, resulting in slow startup speeds and low resource scheduling efficiency on edge terminals. Deploying a 3D Gaussian splash scene model for students involves loading model parameters in blocks according to regional importance scores. The loading order is high-importance region model parameters, medium-importance region model parameters, and low-importance region model parameters. Block loading is prioritized based on regional importance levels, ensuring the rapid activation of core business functions. High-importance region model parameters correspond to parameters for regions with an importance score higher than a first set value (fixed at 0.7). These parameters contain geometric, texture, and attribute information of the core business region, and core functions are activated immediately after loading. Medium-importance region model parameters correspond to parameters for regions with an importance score between the first and second set values ​​(fixed at 0.4). These parameters contain information about auxiliary business regions, and their loading enhances the scene's auxiliary representation. Low-importance region model parameters correspond to parameters for regions with an importance score lower than the second set value. These parameters contain background region information, and their loading completes the scene's overall representation. The block loading process splits the parameters of the student's 3D Gaussian splash scene model into independent data blocks by region. Each data block corresponds to the parameters of one region, and the data blocks are independent of each other and can be read and stored independently. The processing unit of the edge terminal reads the data blocks sequentially and stores them in the memory unit. It prioritizes reading the data blocks of high-importance regions, and immediately activates the scene expression function of the corresponding region after reading. Subsequently, it reads the data blocks of medium-importance and low-importance regions, gradually improving the complete expression of the scene. The memory unit only stores the currently loaded data block, without occupying the memory of the overall model. In the first embodiment, when the AR glasses deploy the model, the parameters of the high-importance interactive region are loaded first to quickly activate the interactive function, and then the parameters of the medium and low-importance background regions are loaded. The terminal startup speed is improved, the memory usage is kept within a reasonable range, and the scene rendering and interactive tasks can be run stably for a long time. In the second embodiment, when the industrial inspection terminal deploys the model, the high-importance parameters of the defect detection region are loaded first to quickly activate the defect recognition function, and then the background region parameters are loaded to meet the real-time requirements of industrial inspection. In the third embodiment, when the vehicle-mounted edge device deploys the model, the high-importance parameters of the road navigation region are loaded first to quickly activate the environmental perception function and ensure the safety of autonomous driving. The chunked loading method adapts to all types of edge terminals, optimizes terminal memory usage and startup efficiency, and eliminates hardware compatibility issues.

[0028] During the execution of the student's 3D Gaussian splash scene model on the edge terminal, the latest observation data of the scene is collected in real time. This latest observation data is compared with the current model's rendering data, and the process of determining changed regions and incremental distillation updates is repeatedly executed. Model parameters are continuously iterated and optimized, ensuring that the scene representation always remains consistent with the actual state. During model execution, the processing unit monitors hardware resource usage in real time. When storage resources are insufficient, the quantization bit width of low-importance regions is automatically reduced to free up storage space. When computing resources are insufficient, the number of Gaussian elements in low-importance regions is automatically reduced to decrease computational overhead. This resource adjustment process does not affect the representation effect of high-importance regions, ensuring the stable operation of core business functions. The model's parameter storage format supports cross-device migration. The same set of student model parameters can be seamlessly migrated between industrial inspection terminals, mobile robots, AR glasses, and vehicle-mounted edge devices without parameter loss or format conversion overhead, improving the method's versatility and adaptability. All operations are completed locally on the edge terminal, without cloud data transmission or computation dependencies, avoiding data transmission latency and privacy risks. This adapts to scenarios with high real-time and security requirements, such as industrial, automotive, and AR applications. All steps of this method are implemented based on publicly available hardware devices and basic algorithms, without relying on dedicated hardware or confidential algorithms. Those skilled in the art can directly reproduce the technical solution by following the described process, formulas, parameters, and operation steps. This solves the problems of global compression distortion, inefficient full update, and poor edge adaptability in existing technologies, and enables stable deployment and dynamic maintenance of the 3D Gaussian splash model on edge terminals.

[0029] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A 3D Gaussian splashing adaptive distillation compression and update method for edge terminals, characterized in that, include: S1: Obtain multi-view image data and camera parameters of the target scene, and train a 3D Gaussian splash scene model of the teacher based on the multi-view image data and camera parameters; S2: Generate a region importance score based on reprojection error, view coverage, texture variation, and task attention. Divide the Gaussian primitives in the teacher's 3D Gaussian splash scene model into regions based on the region importance score. S3: For the different regions after division, based on the regional importance score, the Gaussian meta-retention ratio, attribute quantization accuracy and distillation loss weight are set respectively. The region adaptive distillation compression is performed on the teacher's 3D Gaussian splash scene model to obtain the student's 3D Gaussian splash scene model. S4: Compare the newly added observation data with the historical rendering data of the student's 3D Gaussian splash scene model to determine the changed areas of the target scene, and only perform incremental distillation updates on the Gaussian elements corresponding to the changed areas. S5: Deploy the student's 3D Gaussian splash scene model to the edge terminal.

2. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The regional importance score is satisfied: ； Score the regional importance of the r-th region. The normalized reprojection error index. As a normalized perspective coverage indicator, This is a normalized texture variation index. To normalize the task attention index. The reprojection error weighting coefficient is used. For view coverage weighting coefficient, This is the texture variation weighting coefficient. This is the task attention weighting coefficient. .

3. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 2, characterized in that, The normalization of reprojection error, view coverage, texture variation, and task attention satisfies the following: ， The index value is the normalized value. These are the original index values. This represents the minimum value of the original metric in the target scenario. This represents the maximum value of the original metric in the target scenario. This is the stable term in the denominator.

4. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The Gausky element retention ratio in the r-th region satisfies: ； The percentage of Gaussian elements retained in the r-th region. To preset the minimum retention ratio, To preset the maximum retention ratio, Score the regional importance of the r-th region.

5. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The quantization precision of the r-th region corresponds to the quantization bit width satisfying the following: ； Let r be the quantization bit width of the r-th region. To preset the minimum quantization bit width, To preset the maximum quantization bit width, Score the regional importance of the r-th region.

6. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The distillation loss used in region adaptive distillation compression includes rendering consistency loss, geometric consistency loss, attribute consistency loss and regularization loss. The distillation loss is a weighted sum of the distillation loss weights of different regions and the rendering consistency loss, geometric consistency loss, attribute consistency loss of the corresponding regions, and regularization loss.

7. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The area of ​​change is determined based on the change score, and the change score must satisfy: ， Score the changes in the r-th region. For reprojection differences, For texture differences, For task semantic differences, The reprojection difference weighting coefficient, This represents the texture difference weighting coefficient. This represents the semantic difference weighting coefficient for the task.

8. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, Incremental distillation updates include adding, correcting, replacing, and deleting Gaussian elements corresponding to the changed regions. During the incremental distillation update process, the parameters of the student 3D Gaussian splash scene model corresponding to the unchanged regions remain unchanged.

9. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, The methods for dividing the Gaussian primitives in the 3D Gaussian splash scene model of the teacher include regular grid spatial block method and semantic region division method.

10. The 3D Gaussian splash adaptive distillation compression and update method for edge terminals according to claim 1, characterized in that, Deploying the student's 3D Gaussian splash scene model involves loading model parameters in blocks according to the importance score of the region, with the loading order being high importance region model parameters, medium importance region model parameters, and low importance region model parameters.

Images