Three-dimensional gaussian scene simulation method and system for automatic driving end-to-end test

Abstract

This invention relates to a 3D Gaussian scene simulation method and system for end-to-end testing of autonomous driving. It aims to address the difficulty of simultaneously achieving geometric consistency, multi-agent interaction, and real-time rendering performance in existing autonomous driving simulation technologies. In the offline phase, a time-invariant static background is reconstructed, and a compact dynamic vehicle asset library is built using a semantic-aware compression process. In the online phase, dynamically planned trajectories are received, and vehicles are accurately registered to a unified 3D coordinate system through map-guided geometric registration. A real-time illumination reconstruction module based on agents extracts local ambient light probes, achieving seamless integration of vehicle photometric consistency. Finally, closed-loop sensor images are output through unified rasterization. Compared with existing technologies, this invention can significantly reduce memory usage while maintaining high-fidelity visual effects, supports real-time generation of multi-agent highly interactive scenes, and meets the large-scale closed-loop safety testing requirements of end-to-end autonomous driving algorithms.

Description

Technical Field

[0001] This invention relates to the intersection of autonomous driving simulation testing and neural rendering, and in particular to a three-dimensional Gaussian scene simulation method and system for end-to-end testing of autonomous driving. Background Technology

[0002] In recent years, autonomous driving (AD) has been rapidly shifting from the classic modular pipeline to an end-to-end (E2E) learning paradigm, which directly maps raw sensor observations into control commands. Although end-to-end algorithms reduce error accumulation, their safety and driving strategies must undergo rigorous closed-loop validation across a wide range of demanding scenarios.

[0003] However, existing methods for microscopic driving behavior modeling and scene generation inherently involve trade-offs and compromises regarding these requirements. Specifically, existing technologies can be mainly categorized into the following three types and their limitations: First, traditional game engine-based simulators offer mature scene graphs and real-time rasterization pipelines, boasting strong interactivity and high performance. However, the engineering cost of building diverse high-fidelity assets is extremely high, and the over-reliance on artificially created materials and approximate physical models results in synthetic images with obvious artificial appearances and a significant simulation-to-real domain offset problem between them and real sensor data.

[0004] Secondly, 2D generative models achieve remarkable photorealism in scene synthesis. However, because these methods are inherently 2D and lack an intrinsic, multi-view physically consistent 3D representation, they struggle to address issues of timing inconsistencies and high inference latency when dealing with the high frame rate demands of the control loop.

[0005] Finally, neural 3D reconstruction and neural rendering methods can directly reconstruct geometry and view-dependent appearance in a 3D coordinate system, effectively reducing domain offset problems and supporting new viewpoint synthesis. However, applying traditional 3DGaussian Splatting (3DGS) techniques to interactive end-to-end testing encountered significant bottlenecks. Because it typically assumes a static scene and lacks the modular decomposition capability of perceptual maps, it cannot clearly separate persistent static infrastructure from mobile traffic participants. When attempting to seamlessly insert dynamic vehicles into a static scene, the following three levels of representation conflicts arise: (1) Efficiency conflict: Storing and rendering multiple high-fidelity vehicle instances will exhaust video memory (Out of Memory, OOM), resulting in a severe drop in frame rate when multiple agents exist in the scene at the same time, which cannot meet the real-time requirements of large-scale closed-loop simulation; (2) Geometric conflict: Unstructured neural coordinates cannot be aligned with the semantic metric map required for planning and control, resulting in inconsistent trajectory registration and alignment in closed-loop simulation; (3) Photometric conflict: In neural scene representation, ambient lighting is usually coupled with geometry and appearance, causing newly inserted dynamic assets to be unable to automatically inherit consistent ambient lighting, contact shadows and ambient reflections.

[0006] In summary, existing methods cannot meet the comprehensive requirements of large-scale, rigorous end-to-end algorithm closed-loop training pipelines. Therefore, there is an urgent need for a 3D Gaussian scene simulation method and system to resolve the aforementioned representation conflicts and provide a closed-loop testing environment for autonomous driving that combines real-time interaction with photorealistic simulation. Summary of the Invention

[0007] The purpose of this invention is to overcome the defects of the prior art and provide a three-dimensional Gaussian scene simulation method and system for end-to-end testing of autonomous driving, so as to solve or partially solve the problems that storing and rendering multiple high-fidelity vehicle instances will exhaust video memory, resulting in a serious drop in frame rate when multiple intelligent agents exist in the scene at the same time.

[0008] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving, comprising an offline phase and an online phase: Offline phase: Acquire real driving sensor data and reconstruct a static background 3D Gaussian point cloud; Vehicle asset data in Gaussian form is acquired, and semantically aware asset compression is performed through importance scoring, adaptive pruning, and attribute vectorization to construct a vehicle asset library. Online phase: The two-dimensional planning trajectory during the simulation process is obtained, and the two-dimensional planning trajectory is mapped and registered to a preset map through planar registration and opacity accumulation three-dimensional grounding. Based on the static background 3D Gaussian point cloud and the vehicle asset library, synthetic sensor images are output through lighting reconstruction and rendering for end-to-end testing of autonomous driving.

[0009] As a preferred technical solution, the semantically aware asset compression includes the following steps: For each primitive in the vehicle asset data, an importance score is calculated based on the visibility of the training view sample, the color contrast with the local background contrast, and the texture entropy of the texture complexity in the primitive's neighborhood. Based on the importance score and the preset threshold, adaptive pruning is performed on all Gaussian units to obtain the filtered Gaussian units. The high-dimensional covariance and color spherical harmonic coefficients of the filtered Gaussian elements are replaced with integer indices from the learnable codebook to achieve attribute vectorization.

[0010] As a preferred technical solution, the importance score is calculated using the following formula: in, For the first The importance score of each high-level element. , , These are, respectively, visibility estimation based on training view sampling, color contrast term measuring contrast with local background, and texture entropy term capturing texture complexity within the neighborhood of primitives. , , These are the corresponding weighting coefficients.

[0011] As a preferred technical solution, the attribute vectorization is implemented using the following formula: in, The first Covariance and color spherical harmonic coefficients after Gaussian approximation (integer index) , These are a learnable shape codebook and a learnable color codebook, respectively. , The first The original high-dimensional covariance and color spherical harmonic coefficients of Gaussian elements.

[0012] As a preferred technical solution, the planar registration includes the following steps: Extract vectorized lane centerlines from a bird's-eye view with a static background; Construct a cost matrix that simultaneously penalizes Euclidean distance and heading error, and calculate the optimal alignment path between the two-dimensional programming trajectory and the lane line through dynamic time warping; The globally optimal two-dimensional rigid body transformation is calculated through orthogonal Proukt analysis, thus achieving planar registration.

[0013] As a preferred technical solution, the opacity accumulation three-dimensional grounding includes the following steps: For anchor points defined at the bottom corner of the asset bounding box Calculate its vertical height ,in, and Vertical columns Opacity and height of intersecting background Gaussian elements; By fitting the local ground plane using the least squares method, the vertical position, pitch angle, and roll angle attitude of the vehicle asset's wheels in physical contact with the terrain are obtained, achieving three-dimensional grounding with accumulated opacity.

[0014] As a preferred technical solution, the illumination reconstruction includes the following steps: The spherical harmonic coefficients of the background are used as a dense illumination probe field, and the Gaussian features of the adjacent background are aggregated to construct a local ambient illumination descriptor. The spherical harmonic coefficients of the standard vehicle are updated using an offline pre-calibrated linear modulation operator; Extract the dominant light intensity from the ambient light descriptor, dynamically modulate the intensity of the parameterized shadow mask projected onto the local ground plane, and generate a contact shadow that matches the environment.

[0015] As a preferred technical solution, the local ambient lighting descriptor is: in, This is a local ambient lighting descriptor. , These are the spherical harmonic coefficient and visibility, respectively. For the neighboring background Gaussian elements, For the spatial location of the primitive, For distance scale parameters, For the first The visibility and distance weighting of the contribution of each neighboring background Gaussian element to local illumination. This refers to the local spatial location from which ambient lighting information needs to be extracted. The updated spherical harmonic coefficients are calculated using the following formula: in, , These are the updated spherical harmonic coefficients and the vehicle's own spherical harmonic coefficients, respectively. and The transformation matrix is ​​optimized using offline least squares regression. Matrix multiplication is a linear modulation process applied to the spherical harmonic coefficients of a vehicle.

[0016] As a preferred technical solution, the rendering includes the following steps: 3D Gaussian point cloud of static background and the volume of dynamic intelligent agents The global scene is obtained by taking the union of the spatial sets. ,in, The number of dynamic intelligent agents. This represents a 6-DOF rigid body transformation of a dynamic intelligent agent from a local coordinate system to a world coordinate system. This represents the operation of mapping a gauge volume to the world coordinate system using a rigid body transformation; After rotating and aligning the intrinsic spherical harmonic coefficients of the dynamic vehicle using the Wigner D-matrix, all reconstructed Gaussian elements are merged into a unified depth sorting list, and the sensor image is synthesized by continuously outputting alpha mixing from front to back.

[0017] Another aspect of the present invention provides a three-dimensional Gaussian scene simulation system for end-to-end testing of autonomous driving, used to implement the aforementioned three-dimensional Gaussian scene simulation method. The system includes an offline preparation unit and an online real-time simulation unit. The offline preparation unit includes: The static background reconstruction module is used to process real driving sensor data and reconstruct a time-invariant static background 3D Gaussian model. The dynamic asset compression module is used to process the original high-dimensional vehicle assets using a semantically aware compression process to build a compact dynamic vehicle asset library. The online real-time simulation unit includes: The geometric registration module is used to obtain the two-dimensional trajectory input of the dynamic planner and perform map-guided geometric registration based on high-precision maps or extracted semantic topology to register dynamic vehicle assets into a unified three-dimensional coordinate system. The real-time illumination reconstruction module is used to extract local ambient illumination information and perform agent-based photometric integration and illumination reconstruction on the registered dynamic vehicle assets to match the lighting conditions of the background environment. The unified rendering module spatially merges time-invariant static backgrounds with all transformed dynamic vehicle assets and outputs synthetic sensor images for end-to-end testing via unified rasterization rendering.

[0018] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Improved operational efficiency and scalability: This invention reconstructs a static background 3D Gaussian point cloud from real driving sensor data, and compresses the Gaussian trigonometric vehicle asset data into semantically perceptual assets through importance scoring, adaptive pruning, and attribute vectorization to construct a vehicle asset library. Thanks to decoupling the scene into a persistent background and compressing the standardized assets through vector quantization, the exponential memory growth caused by the traditional 3D Gaussian method with the increase of the number of agents is avoided. In extreme high-pressure scenarios with up to 50 dynamic agents, the memory usage of this system increases linearly and reaches a maximum of only about 2.3GB, while maintaining a usable rendering frame rate of up to 15.4FPS. In scenarios with 10 agents, it can reach 38.5FPS, providing sufficient computing power for large-scale multi-agent closed-loop simulation.

[0019] (2) High realism of illumination: This invention acquires the two-dimensional planning trajectory during the simulation process, maps and registers the two-dimensional planning trajectory to a preset map through planar registration and opacity accumulation three-dimensional grounding, achieving an extremely low average trajectory displacement error of only 0.05m and a ground penetration rate of 0.2%, fundamentally eliminating dynamic vehicle drift and clipping problems. Furthermore, real-time illumination reconstruction can synthesize accurate contact shadows and specular reflections under various complex lighting conditions, achieving a PAE of only 0.05m. The lowest peak angle error and the highest luminous fidelity eliminate visual incongruity.

[0020] (3) Strong adaptability: The simulation environment constructed by this invention can be directly used as a high-fidelity sensor data engine closed-loop access to mainstream end-to-end autonomous driving algorithms. By deploying high-density interactive and adversarial lighting scenarios, it can effectively and reliably trigger the extreme safety boundary behavior of the planning algorithm, successfully expose the potential vulnerabilities of the end-to-end planner, and provide a principled and robust stress testing scheme for autonomous driving strategies. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a 3D Gaussian scene simulation for end-to-end testing of autonomous driving in the embodiment. Figure 2 This is a schematic diagram of the scenario for the end-to-end autonomous driving simulation test results in the embodiment; Figure 3 This is a diagram demonstrating the decoupling effect between dynamic assets and static background in the example. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] Example 1 To address the problems existing in the aforementioned prior art, this embodiment provides a 3D Gaussian scene simulation method for end-to-end testing of autonomous driving. The aim is to decouple the scene into a high-fidelity static background and a manipulable dynamic intelligent agent, and introduce specific compression, registration and lighting reconstruction processes to achieve large-scale, interactive closed-loop scene simulation testing with high rendering speed, high geometric accuracy and high photometric consistency.

[0024] See Figure 1 The method employs a two-stage design: an offline preparation stage and an online real-time simulation stage. The former includes background reconstruction and semantically aware asset compression, while the latter includes map-guided registration, agent-based illumination reconstruction, and unified synthesis closed-loop evaluation.

[0025] The specific implementation steps are as follows: Step S1: Offline preparation stage: Static environment and asset initialization.

[0026] In this invention, "vehicle assets" specifically refers to three-dimensional Gaussian (3DGS) point cloud data in its original high-dimensional state, which includes high-dimensional attributes such as the vehicle's three-dimensional spatial coordinates, covariance (representing shape and scale), opacity, and color spherical harmonic coefficients. It is defined within a standardized local coordinate system as an instantiable prefabricated component of a dynamic intelligent agent.

[0027] First, a time-invariant static background 3D Gaussian (3DGS) model is reconstructed using real driving sensor data.

[0028] Secondly, semantically aware asset compression is performed, transforming the high-dimensional original 3D Gaussian vehicle assets into a compact, canonical volume. The specific compression process is as follows: Figure 2 As shown: 1) Explicit importance score.

[0029] Instead of relying on black-box neural networks, scalar importance scores are explicitly computed for each Gaussian unit. .in For visibility estimation based on training view sampling, The color contrast term is used to measure the contrast with the local background. The texture entropy term is used to capture the texture complexity within the neighborhood of the primitive.

[0030] 2) Adaptive cropping.

[0031] Redundant primitives are eliminated according to a strict threshold set in this embodiment. The empirical value is set to 0.005, so that computing resources are concentrated on preserving important vehicle structural details such as wheels and chassis boundaries.

[0032] 3) Attribute vector quantization.

[0033] To further minimize memory footprint, the high-dimensional covariance and color spherical harmonic coefficients of the retained primitives are replaced with a learnable codebook, such as... and Integer index in, i.e. and This discretization process effectively prevents memory exhaustion when loading dozens of vehicles.

[0034] Step S2: Online real-time geometric registration stage.

[0035] like Figure 3 As shown, this stage is responsible for strictly mapping and registering the dynamic two-dimensional trajectory output by the planner to a unified three-dimensional scene map coordinate system.

[0036] 1) Semantic topology and planar registration.

[0037] When high-resolution maps are unavailable, algorithms such as MapTRv2 are used to extract vectorized lane centerlines from a static bird's-eye view. The system constructs a cost matrix that simultaneously penalizes Euclidean distance and heading errors, and uses a dynamic time warping algorithm to find the optimal alignment path between the 2D planned trajectory and the lane lines. Subsequently, orthogonal Procrustes are used to calculate the globally optimal 2D rigid body transformation. To eliminate systematic lateral offset.

[0038] 2) Opacity accumulation three-dimensional grounding.

[0039] Because 3DGS lacks a continuous surface mesh, the system uses opacity-weighted depth accumulation for vehicle height estimation. For anchor points defined at the bottom corners of the asset bounding box... Its vertical height is calculated as Then, the least squares method is used to fit the local ground surface to determine the final vertical position, pitch angle, and roll angle attitude of the vehicle asset, ensuring that the wheels are physically in reasonable contact with the terrain.

[0040] Step S3: Online lighting reconstruction and rendering stage.

[0041] This stage does not require online neural network inference and can directly achieve lightweight photometric fusion and image output.

[0042] 1) Local probe sampling and linear spherical harmonic transfer.

[0043] The spherical harmonic coefficients of the background are considered as a dense field of illumination probes. By aggregating the spherical harmonic coefficients of neighboring background Gaussians And with its visibility A local ambient lighting descriptor is constructed by weighting spatial distance. In the formula, and Vertical columns The opacity and height of the intersecting background Gaussian elements are determined; then, using the aforementioned anchor points, the local ground plane is fitted using the least squares method to obtain the asset's final vertical position, pitch angle, and roll angle attitude. Subsequently, the spherical harmonic coefficients of the canonical vehicle are updated using offline pre-calibrated linear modulation operators. In the formula, and The transformation matrix is ​​optimized via offline least squares regression, enabling the baking of local illumination context without online neural network inference.

[0044] 2) Agent-based contact shadow generation.

[0045] Extract the dominant light intensity from the ambient light descriptor Dynamically modulated parametric shadow mask projected onto a local ground plane The intensity of the mask is dynamically adjusted to control the background Gaussian intensity, generating realistic contact shadows that match the environment.

[0046] 3) Unified rasterization and output.

[0047] The entire global scene is defined as a static background. and the volume of dynamic intelligent agents (i.e., the spatial union of the compact dynamic vehicle specification model constructed from the vector quantization codebook after semantically-aware compression of the original vehicle assets in step S1) After applying the Wigner D-matrix to rotate and align the intrinsic spherical harmonic coefficients of the dynamic vehicle, the system merges all reconstructed primitives into a unified depth-sorted list and synthesizes sensor images through continuous front-to-back alpha mixing output. The generated images are then directly fed into the end-to-end autonomous driving network as perception inputs for closed-loop testing.

[0048] Table 1: Comparison Results of Asset Compression Methods and Mainstream Methods Test results verification: Experimental verification shows that the simulation system built in this embodiment maintains a stable memory usage of approximately 2.3GB and a rendering throughput of 15.4 FPS in concurrent scenarios with up to 50 dynamic agents. Meanwhile, geometric registration reduces the average displacement error (ADE) to as low as 0.05m, and lighting reconstruction reduces the peak angle error (PAE) to 6.8°. This not only overcomes the problem of memory overflow under high concurrency but also significantly improves the luminosity and geometric realism of the closed-loop generated scene.

[0049] Example 2 Based on Example 1, this example provides a decoupled 3D Gaussian scene simulation system for end-to-end testing of autonomous driving, including an offline preparation unit and an online real-time simulation unit.

[0050] (1) The offline preparation unit includes.

[0051] 1) Static background reconstruction module: Used to process real driving sensor data and reconstruct a time-invariant static background 3D Gaussian model.

[0052] 2) Dynamic Asset Compression Module: This module utilizes a semantically aware compression process to process the original high-dimensional vehicle assets and construct a compact dynamic vehicle specification asset library. Specifically, this module includes three stages: importance scoring, adaptive pruning, and attribute vector quantization. First, a scalar importance score is explicitly calculated for each Gaussian element. ,in For visibility estimation, For color contrast, For texture entropy; secondly, through a set strict threshold. Redundant primitives below a threshold are adaptively removed; finally, the high-dimensional properties of the retained primitives are vector-quantized, and the covariance and color spherical harmonic coefficients are replaced with compact integer indices of the most recent entries in the learnable codebook, resulting in a discretized representation. and This significantly reduces video memory usage.

[0053] (2) The online real-time simulation unit includes: 1) Geometric Registration Module: This module acquires the 2D trajectory input from the dynamic planner and performs strict alignment guided by the map. In the planar registration stage, the Dynamic Time Warping (DTW) algorithm is used to calculate the optimal sequential alignment path between the planned trajectory and the semantic map topology (e.g., lane centerlines), and orthogonal Procrustes are used to calculate 2D rigid body transformations to eliminate systematic lateral drift. In the 3D height estimation stage, an opacity-weighted depth accumulation method is used to calculate the vertical height of the anchor points. The least squares method is used to fit the local ground surface to ensure that the vehicle wheels maintain physical contact with the terrain.

[0054] 2) Real-time illumination reconstruction module: This module achieves seamless photometric integration through a proxy-based illumination mechanism, eliminating the need for online neural network inference. First, the background spherical harmonic coefficients are treated as a dense illumination probe field. Then, the spherical harmonic coefficients of neighboring background Gaussians are aggregated to extract local ambient illumination descriptors. Then, a pre-calibrated linear modulation operator is used. The ambient light is shifted; finally, the dominant light intensity of the light descriptor is extracted. To dynamically adjust the intensity of the parameterized mask projected onto the ground surface, in order to generate realistic contact shadows.

[0055] 3) Unified rendering module: the entire global scene exist A moment is defined as a time-invariant static background field. and The volume of a dynamic intelligent agent canonical Spatial union: By applying the Wigner D-matrix to rotate and align the intrinsic spherical harmonic coefficients, the background and the transformed dynamic assets are merged into a unified aggregation list. The synthesized sensor image is then calculated and output through continuous forward-backward alpha blending.

[0056] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving, characterized in that, Includes offline and online phases: Offline phase: Acquire real driving sensor data and reconstruct a static background 3D Gaussian point cloud; Vehicle asset data in Gaussian meta-form is acquired, and semantically aware asset compression is performed through importance scoring, adaptive pruning, and attribute vectorization to construct a vehicle asset library. Online phase: The two-dimensional planning trajectory during the simulation process is obtained, and the two-dimensional planning trajectory is mapped and registered to a preset map through planar registration and opacity accumulation three-dimensional grounding. Based on the static background 3D Gaussian point cloud and the vehicle asset library, synthetic sensor images are output through lighting reconstruction and rendering to conduct end-to-end testing of autonomous driving. The semantically aware asset compression includes the following steps: For each primitive in the vehicle asset data, an importance score is calculated based on the visibility of the training view sample, the color contrast with the local background contrast, and the texture entropy of the texture complexity in the primitive's neighborhood. Based on the importance score and the preset threshold, adaptive pruning is performed on all Gaussian units to obtain the filtered Gaussian units. The high-dimensional covariance and color spherical harmonic coefficients of the filtered Gaussian elements are replaced with integer indices in the learnable codebook to achieve attribute vectorization. The opacity accumulation three-dimensional grounding includes the following steps: For anchor points defined at the bottom corner of the asset bounding box Calculate its vertical height ,in, and Vertical columns Opacity and height of intersecting background Gaussian elements; The least squares method is used to fit the local ground plane to obtain the vertical position, pitch angle and roll angle attitude of the vehicle asset's wheels in physical contact with the terrain, thus achieving three-dimensional grounding with accumulated opacity. The illumination reconstruction includes the following steps: The spherical harmonic coefficients of the background are used as a dense illumination probe field, and the Gaussian features of the adjacent background are aggregated to construct a local ambient illumination descriptor. The spherical harmonic coefficients of the standard vehicle are updated using an offline pre-calibrated linear modulation operator; Extract the dominant light intensity from the ambient light descriptor, dynamically modulate the intensity of the parameterized shadow mask projected onto the local ground plane, and generate a contact shadow that matches the environment.

2. The three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving according to claim 1, characterized in that, The importance score is calculated using the following formula: in, For the first The importance score of each high-level element. , , These are, respectively, visibility estimation based on training view sampling, color contrast term measuring contrast with local background, and texture entropy term capturing texture complexity within the neighborhood of primitives. , , These are the corresponding weighting coefficients.

3. The three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving according to claim 1, characterized in that, The attribute vectorization is implemented using the following formula: in, The first Covariance and color spherical harmonic coefficients after Gaussian approximation (integer index) , These are a learnable shape codebook and a learnable color codebook, respectively. , The first The original high-dimensional covariance and color spherical harmonic coefficients of Gaussian elements.

4. The three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving according to claim 1, characterized in that, The planar registration includes the following steps: Extract vectorized lane centerlines from a bird's-eye view with a static background; Construct a cost matrix that simultaneously penalizes Euclidean distance and heading error, and calculate the optimal alignment path between the two-dimensional programming trajectory and the lane line through dynamic time warping; The globally optimal two-dimensional rigid body transformation is calculated through orthogonal Proukt analysis, thus achieving planar registration.

5. The three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving according to claim 1, characterized in that, The local ambient lighting descriptor is: in, This is a local ambient lighting descriptor. , These are the spherical harmonic coefficient and visibility, respectively. For the neighboring background Gaussian elements, For the spatial location of the primitive, For distance scale parameters, For the first The visibility and distance weighting of the contribution of each neighboring background Gaussian element to local illumination. This refers to the local spatial location from which ambient lighting information needs to be extracted. The updated spherical harmonic coefficients are calculated using the following formula: in, , These are the updated spherical harmonic coefficients and the vehicle's own spherical harmonic coefficients, respectively. and The transformation matrix is ​​optimized using offline least squares regression. Matrix multiplication is a linear modulation process applied to the spherical harmonic coefficients of a vehicle.

6. The three-dimensional Gaussian scene simulation method for end-to-end testing of autonomous driving according to claim 1, characterized in that, The rendering process includes the following steps: 3D Gaussian point cloud of static background and the volume of dynamic intelligent agents The global scene is obtained by taking the union of the spatial sets. ,in, The number of dynamic intelligent agents. This represents a 6-DOF rigid body transformation of a dynamic intelligent agent from a local coordinate system to a world coordinate system. This represents the operation of mapping a gauge volume to the world coordinate system using a rigid body transformation; After rotating and aligning the intrinsic spherical harmonic coefficients of the dynamic vehicle using the Wigner D-matrix, all reconstructed Gaussian elements are merged into a unified depth sorting list, and the sensor image is synthesized by continuously outputting alpha mixing from front to back.

7. A three-dimensional Gaussian scene simulation system for end-to-end testing of autonomous driving, characterized in that, To implement the three-dimensional Gaussian scene simulation method as described in any one of claims 1-6, the system includes an offline preparation unit and an online real-time simulation unit. The offline preparation unit includes: The static background reconstruction module is used to process real driving sensor data and reconstruct a time-invariant static background 3D Gaussian model. The dynamic asset compression module is used to process the original high-dimensional vehicle assets using a semantically aware compression process to build a compact dynamic vehicle asset library. The online real-time simulation unit includes: The geometric registration module is used to obtain the two-dimensional trajectory input of the dynamic planner and perform map-guided geometric registration based on high-precision maps or extracted semantic topology to register dynamic vehicle assets into a unified three-dimensional coordinate system. The real-time illumination reconstruction module is used to extract local ambient illumination information and perform agent-based photometric integration and illumination reconstruction on the registered dynamic vehicle assets to match the lighting conditions of the background environment. The unified rendering module spatially merges time-invariant static backgrounds with all transformed dynamic vehicle assets and outputs synthetic sensor images for end-to-end testing via unified rasterization rendering.

Images