Vehicle model construction method and device, storage medium and program product

Abstract

The invention discloses a vehicle model construction method and device, a storage medium and a program product, and the method comprises the steps: carrying out the reconstruction of a vehicle white pattern of a target vehicle based on the collection data of the target vehicle, and extracting the parameters of a collision body bounding box of the vehicle white pattern; shape features of the target vehicle are extracted, and corresponding matched vehicle materials are determined from a pre-constructed vehicle material library according to the extracted shape features; extracting vehicle model characteristic parameters corresponding to the matched vehicle materials, and performing deformation processing on the vehicle white mold according to the vehicle model characteristic parameters to generate a derivative vehicle model; and outputting the derivative vehicle model carrying the collision body bounding box parameters. Therefore, the modeling efficiency is improved, the cost is reduced, and meanwhile, the adaptability and the accuracy of the vehicle model in a simulation system are improved.

Description

Technical Field

[0001] This application relates to the field of simulation modeling technology, and in particular to a method, device, storage medium and program product for building vehicle models. Background Technology

[0002] With the rapid development of autonomous driving technology, vehicle simulation has become an important means of testing and optimizing autonomous driving systems. High-quality, realistic vehicle models play a crucial role in autonomous driving simulation. Currently, in fields such as autonomous driving simulation, virtual environment construction, and game development, vehicle model construction techniques mainly include manual modeling, laser scanning reconstruction, and procedural generation. However, these traditional methods all have certain limitations, affecting the deployment and application of large-scale simulation systems.

[0003] On the one hand, manual modeling relies on professional 3D modeling software (such as Maya and Blender) to build vehicle models. Although this method can provide relatively accurate models, the production time for each model is long (generally 8 to 16 hours), making it difficult to meet the needs of a large number of different vehicle models in large-scale scenarios. In addition, manually modeled models mostly lack dynamic adaptability and cannot flexibly cope with deformation and collision situations in different simulation scenarios, limiting their scalability in practical applications.

[0004] On the other hand, laser scanning reconstruction technology acquires 3D data of real vehicles using high-precision laser scanners (such as the Faro Focus scanner), providing millimeter-level accuracy and making it suitable for certain high-precision scenarios. However, laser scanning reconstruction is expensive (generally costing over $5,000 per vehicle) and complex to operate, making it difficult to meet the vehicle model acquisition needs of large-scale simulations or virtual reality applications. Furthermore, the static models obtained through laser scanning are difficult to integrate effectively with dynamic simulation environments, resulting in insufficient application in dynamic adaptation and scene interaction.

[0005] On the other hand, procedural generation techniques based on methods such as Generative Adversarial Networks (GANs) have been gradually introduced into the field of vehicle modeling in recent years. Although this method has high generation efficiency and can quickly synthesize a large number of vehicle models, procedurally generated models often have topological errors, and due to the limitations of training data, the generated vehicle models cannot fully reflect the complex details of the real world in some cases. In addition, due to the lack of data, the generated models lack ground truth data, making it difficult to achieve the expected realism and accuracy when applied in simulation systems. Summary of the Invention

[0006] This application provides a vehicle model construction method, device, storage medium, and program product to at least solve one of the above-mentioned technical problems.

[0007] In a first aspect, embodiments of this application provide a vehicle model construction method, comprising: reconstructing a white model of the target vehicle based on collected data of the target vehicle, and extracting the bounding box parameters of the collision body of the white model; extracting the shape features of the target vehicle, and determining corresponding matching vehicle materials from a pre-built vehicle material library according to the extracted shape features; extracting the vehicle model feature parameters corresponding to the matching vehicle materials, and deforming the white model of the vehicle according to the vehicle model feature parameters to generate a derived vehicle model; and outputting the derived vehicle model carrying the bounding box parameters of the collision body.

[0008] Secondly, embodiments of this application provide a storage medium storing one or more programs including execution instructions, which can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform any of the vehicle model construction methods described above in this application.

[0009] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the vehicle model construction methods described above in this application.

[0010] Fourthly, embodiments of this application also provide a computer program product, the computer program product including a computer program stored on a storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to execute any of the above-described vehicle model construction methods.

[0011] The beneficial effects of the embodiments of this application are as follows: By automatically generating a white model of the vehicle based on data collected from the target vehicle, and combining this with precise vehicle shape features and model type characteristic parameters for deformation processing, a high-precision and dynamically adaptable derived vehicle model is generated. By extracting the bounding box parameters of the collider, accurate physical collision detection in the simulation environment is ensured. Simultaneously, the model's diversity and realism are enhanced by utilizing a pre-built vehicle material library and matching shape features. Furthermore, through deformation processing based on vehicle model type characteristic parameters, the model can flexibly adapt to the simulation requirements of different vehicles, supporting the construction of different vehicle categories. This improves modeling efficiency and reduces costs, while simultaneously enhancing the adaptability and accuracy of the vehicle model within the simulation system. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating an example of a vehicle model construction method according to an embodiment of this application is shown; Figure 2 This document shows an example of an operation flowchart for reconstructing a vehicle white model in a vehicle model construction method according to an embodiment of this application; Figure 3 This document shows an example of an operation flowchart illustrating the determination of matching vehicle materials in a vehicle model construction method according to an embodiment of this application. Figure 4 The diagram illustrates an example of a vehicle model construction method according to an embodiment of this application, which involves deforming a white vehicle model to generate a derived vehicle model. Figure 5 A flowchart illustrating another example of a vehicle model construction method according to an embodiment of this application is shown; Figure 6 This is a schematic diagram of the structure of an embodiment of the electronic device of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0015] It should also be noted that, in this document, the terms "comprising" or "including" include not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0016] It should be noted that in applications such as autonomous driving simulation and game development, the realism and diversity of vehicle models directly affect the system's coverage and stability. Current technologies primarily rely on three approaches: first, manual modeling, created manually using software like Maya and Blender, where a single model often requires 8-16 man-hours, making it difficult to support large-scale, multi-vehicle scenario construction; second, laser scanning reconstruction, where while high-precision scanning of real vehicles (such as Faro Focus) can obtain millimeter-level models, the cost of acquiring data per vehicle typically exceeds $5,000, making deployment too expensive; and third, procedural generation (such as GAN-based solutions), which has high generation efficiency but is prone to approximately 12% topological errors and lacks ground truth data directly usable for simulation, affecting subsequent verification and evaluation.

[0017] Besides the limitations of the modeling approach itself, several significant bottlenecks exist: insufficient diversity, with vehicle repetition rates reaching 40-60% in city-level simulations due to data and material accumulation constraints, weakening the test of algorithm generalization ability; high resource consumption, with high-precision models reaching 500MB-1GB individually, and scenarios involving tens of thousands of models approaching petabyte-level storage and transmission pressure; and weak dynamic adaptation capabilities, as collision and deformation effects often require manual configuration of physical parameters for different vehicle models, making it difficult to achieve consistent and reliable dynamic performance in complex scenarios. These issues collectively hinder the coordinated advancement of high-fidelity and large-scale simulations.

[0018] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0019] Figure 1 A flowchart illustrating an example of a vehicle model construction method according to an embodiment of this application is shown.

[0020] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities. In some examples, the method in the embodiments of this application can be integrated and configured in an electronic device or terminal through software, hardware or a combination of software and hardware. The type of terminal or electronic device can be diverse, such as mobile phone, tablet computer, desktop computer or vehicle terminal, vehicle controller, server, etc.

[0021] For example, the execution subject of the method in this application embodiment can be a vehicle model building platform. By collecting data to reconstruct the white model of the vehicle, extracting and matching shape features and vehicle model features, dynamically adjusting the shape of the vehicle model, and combining the bounding box information of the collision body, an efficient, accurate, and flexible vehicle model that can be used for large-scale simulation can be generated. It can significantly improve modeling efficiency while maintaining high accuracy and can flexibly meet the needs of different simulation scenarios.

[0022] like Figure 1 As shown, in step S110, a white model of the target vehicle is reconstructed based on the collected data of the target vehicle, and the bounding box parameters of the collision body of the white model are extracted.

[0023] Specifically, precise data acquisition of the target vehicle can be performed using one or more data acquisition methods (such as laser scanning, photography, etc.) to collect point cloud information or image information of the target vehicle. The acquired data can include the vehicle's external shape, size, surface texture, and geometric features of the wheels and body. Furthermore, the quality of the acquired data directly affects the accuracy of the generated vehicle white model. Therefore, high-precision sensor equipment and / or efficient image processing technology can be used to collect and integrate the data to ensure its comprehensiveness and accuracy.

[0024] Furthermore, utilizing the obtained vehicle sensor data, and comprehensively considering factors such as the vehicle's overall shape, size proportions, and the relative positions of the body and wheels, a 3D reconstruction algorithm is used to model the target vehicle, generating a white vehicle model. The white vehicle model is a preliminary, basic model without specific appearance details; it retains the vehicle's geometry and spatial structure but does not contain texture or material information. Simultaneously with generating the white vehicle model, the bounding box of the collider can be extracted. The bounding box can be a minimum rectangle or cube that completely encloses the vehicle model, allowing collision detection to be calculated in the simulation using the bounding box parameters.

[0025] Regarding the implementation details of the collision bounding box extraction, in some examples of the embodiments of this application, the vehicle white model can be converted into a representation suitable for geometric operations before calculating the boundary parameters. Specifically, firstly, the white model point set is denoised and outlier removed (e.g., statistical filtering / radius filtering), and ground points are removed using RANSAC segmentation of the ground plane; subsequently, principal component analysis (PCA) is performed on the vehicle body point set to estimate the vehicle's local coordinate system and major axis direction. Based on this, two types of bounding boxes are calculated: one is the axis-aligned bounding box (AABB), which takes the minimum / maximum value of the three-axis coordinates of the mesh / point set in the world coordinate system to obtain the center point, length, width, and height (extents), and orientation aligned with the XYZ axes; the other is the orientation bounding box (OBB), which takes the minimum / maximum value of the point set in the body coordinate system obtained by PCA and returns to the world system to obtain the center, three-axis dimensions, and attitude quaternions / rotation matrix. To improve robustness, morphological closing operations or local convex hulls can be introduced to suppress small protrusions such as mirror images and antennas, and a height threshold can be set at the tire contact point to avoid ground residue raising the bounding box. Finally, the output collision bounding box parameters can include: center point coordinates, length, width, height, orientation, and binding offset to the vehicle white model. In the simulation, AABB is used preferentially for wide-phase detection by default, and then OBB or a fine mesh is used for narrow-phase detection after a hit, thus balancing the efficiency and accuracy of collision detection.

[0026] Therefore, by collecting data to generate a vehicle white model and extracting the bounding box parameters of the collision body, the accuracy and adaptability of the model can be guaranteed, providing a general model framework that can be adapted to different vehicles, and providing an accurate foundation for subsequent vehicle model deformation and simulation testing.

[0027] In step S120, the shape features of the target vehicle are extracted, and the corresponding matching vehicle material is determined from the pre-built vehicle material library based on the extracted shape features.

[0028] In some implementations, the main shape features of the target vehicle are extracted from the generated vehicle white model, including at least some features such as the body shape, roof profile, body length, width, and the design of windows and headlights. For example, to accurately extract shape features, graphics-based shape analysis algorithms, such as bounding box detection, contour extraction, or surface fitting techniques, can be used to automatically extract the core design features of the vehicle from complex geometry.

[0029] Furthermore, the vehicle resource library is a pre-built database containing various types of vehicle model resources, which can be categorized according to different criteria such as vehicle type, shape, and design style. Then, the extracted shape features of the target vehicle are matched against the pre-built vehicle resource library. For example, a feature similarity-based metric is used to calculate the shape similarity between the target vehicle and each resource model in the library, in order to find the vehicle resource most similar to the target vehicle.

[0030] In step S130, the vehicle model feature parameters corresponding to the matched vehicle material are extracted, and the vehicle white model is deformed according to the vehicle model feature parameters to generate a derived vehicle model.

[0031] In one example, each vehicle image in the vehicle image library has corresponding vehicle model feature parameters; in another example, the model details of the vehicle model corresponding to the matching vehicle image can also be identified to obtain the vehicle model feature parameters of the corresponding image. The model feature parameters include a series of key features that affect the appearance and physical properties of the vehicle, such as the body structure, window layout, wheel size, and roof design, to accurately understand the model detail design of the matching image.

[0032] Next, these vehicle model feature parameters are used to deform the previously generated white vehicle model. The deformation process mainly involves adjusting the geometry of the white vehicle model using shape transformation algorithms (such as constraint-based shape deformation, surface adjustment, mesh topology optimization, etc.) to give it a structural design that matches the vehicle model details of the matching vehicle material.

[0033] It should be understood that the deformation processing can be diversified and is not limited to adjusting the geometry of the vehicle white model. It can also be combined with the structural features in the vehicle model details for fine-tuning, such as the size of the windows, the tilt angle of the front of the car, and the position of the wheels, so that the original vehicle white model can be quickly transformed into a derivative vehicle model with a specific appearance under the guidance of the vehicle model feature parameters.

[0034] For example, when the target vehicle is a truck, the corresponding vehicle model feature parameters are those related to the truck's model details. These parameters may specifically include the number of axles and wheelbase sequence, track width and tire outer diameter, cab type and A / B pillar positions, front and rear overhangs and frame longitudinal beam height, cargo box (or superstructure) length, width, and height, and assembly surfaces. Therefore, during the deformation stage, key areas of the white model are controlled and driven, such as repositioning the wheel centers and adjusting the wheel arch openings accordingly, stretching / translating the chassis frame according to the frame and front / rear overhang parameters, constraining the windshield and door / window opening ratios based on the cab type, and correcting the superstructure's outer envelope according to the cargo box dimensions and assembly surfaces. This allows for rapid convergence into a derived vehicle model that conforms to the truck's details without disrupting the topology and assembly relationships.

[0035] This allows for the flexible generation of various derivative vehicle models, improving the personalization and accuracy of the models and ensuring that the derivative models accurately reflect the external features and design style of the target vehicle.

[0036] In step S140, a derived vehicle model carrying the bounding box parameters of the collider is output.

[0037] Here, the derived vehicle model generated through deformation processing is combined with the previously extracted bounding box parameters of the collider, so that the derived vehicle model not only has a high-precision geometric shape, but also contains the necessary physical properties (such as the bounding box parameters of the collider), enabling it to perform physical simulation interaction and collision detection with other models in the simulation environment.

[0038] For example, the output derived vehicle model can be used for applications such as autonomous driving system testing, simulation verification, and virtual environment construction. Because the model contains bounding box data related to vehicle collisions, it can accurately perform calculations for key tasks such as collision detection and path planning in simulations, ensuring that the autonomous driving system can perform more stably and safely in various complex scenarios.

[0039] This application's embodiments generate a vehicle white model based on collected data and extract collision bounding box parameters, ensuring the model's accuracy and simulation adaptability. By extracting vehicle shape features and matching them to a material library, rapid identification and material selection for different vehicle types are achieved. Controlled deformation of the white model, combined with vehicle model feature parameters, ensures that the generated derived model accurately reflects the structural details and appearance features of different vehicle models. Finally, a derived vehicle model carrying collision bounding box parameters is output, enabling the model to not only possess realistic appearance but also support efficient and reliable collision detection and physical interaction in the simulation environment. This significantly improves the generation efficiency and diversity of vehicle models while balancing realism and dynamic adaptability, providing a high-quality vehicle modeling solution for autonomous driving simulation and related applications.

[0040] Regarding the application of derived vehicle models, in some examples of embodiments of this application, during dynamic simulation of the vehicle driving environment, deformation processing is performed on the derived vehicle model based on events triggered by the simulation scene. Specifically, in dynamic simulation, the derived vehicle model not only exists as a static object but also needs to undergo dynamic deformation in response to events occurring in the scene. For example, when the simulation scene triggers events such as traffic accidents, emergency braking, rollovers, or contact with obstacles, the dynamic deformation module of the vehicle model can be invoked. Based on the geometric structure and material properties of the derived vehicle model, local or overall deformation processing is performed on specific areas of the vehicle through methods such as finite element simulation, skeleton-based deformation control, or constraint-driven mesh deformation.

[0041] In some implementations, if the front of the vehicle collides with an obstacle, the deformation process of the front bumper, engine compartment, and wheel positions can be simulated. If the vehicle rolls over, the motion planning and deformation process of the vehicle model are adjusted by changing the center of gravity and stress points, so that the vehicle body exhibits a twisting or collapsing effect that conforms to physical laws in the simulation. Dynamic deformation realizes the transition of the vehicle model from static display to dynamic interaction, and the derived vehicle model can realistically reflect the simulated response of the vehicle to sudden events in complex driving scenarios.

[0042] Then, physical collision detection is performed based on the bounding box parameters of the colliders, and collision response data is generated. During dynamic deformation, the bounding box parameters pre-embedded in the vehicle model can be invoked. The bounding box, as a simplified physical boundary of the vehicle, can efficiently calculate the model's position in 3D space and its contact state with adjacent objects, and accurately locate the collision area. For example, once a collision is detected, parameters such as vehicle speed, mass, and force direction can be further combined to generate corresponding collision response data, such as real-time recorded collision point coordinates, collision force magnitude, duration of impact, and deformation energy distribution.

[0043] Furthermore, the deformation time-series data and collision response data of the derived vehicle model are correlated and recorded with the simulation scene trigger events to generate a ground truth dataset for training the autonomous driving system. Specifically, after dynamic simulation and collision detection are completed, the deformation state of the derived vehicle model throughout the entire event process can be recorded in a time series, which can include the mesh shape of the vehicle at each time frame, the displacement of key components, and the evolution trajectory of local stress concentration areas. At the same time, collision response data, such as impact force change curves and energy dissipation, can also be recorded synchronously in a time series.

[0044] By binding the aforementioned data with event parameters (such as event type, initial speed, collision angle, road conditions, etc.) of the simulated scenario, a structured ground truth dataset can be formed. This dataset is used to train the perception and decision-making models of autonomous driving systems, helping them to identify collision precursors, assess risks, and make avoidance decisions. Thus, the simulation process generates a high-precision ground truth dataset, solving the problem of the difficulty in obtaining large-scale real-world accident data and enriching the training sample space for autonomous driving systems.

[0045] In some examples of embodiments of this application, the vehicle white model is a 3D Gaussian point cloud model.

[0046] Figure 2 A flowchart illustrating an example of reconstructing a vehicle white model in a vehicle model construction method according to an embodiment of this application is shown.

[0047] like Figure 2As shown, in step S210, the camera pose parameters of the acquired data are calculated using the motion recovery structure algorithm.

[0048] It should be noted that the vehicle appearance data may be collected by devices such as fixed cameras, drones, or surround shooting equipment (such as handheld cameras or mobile data collection vehicles). Therefore, it is necessary to determine the camera pose corresponding to each frame of the image, that is, the position and orientation of the camera in the world coordinate system, to ensure the geometric consistency of subsequent 3D reconstruction.

[0049] Specifically, the Structure from Motion (SfM) algorithm can be used to process the acquired data. By utilizing feature point matching, fundamental matrix factorization, and incremental 3D reconstruction, the camera extrinsic parameters (position and pose) and some sparse point cloud structure of each image frame can be recovered. For example, feature point extraction algorithms can be combined for multi-view matching, and global or local optimization (such as bundled adjustment) can be used to reduce accumulated errors, thereby obtaining high-precision camera pose parameters and providing accurate camera geometric information.

[0050] In step S220, a 3D Gaussian kernel set is initialized based on the camera pose parameters and the acquired data, wherein the 3D Gaussian kernel includes position parameters, covariance matrix parameters, transparency parameters and color parameters.

[0051] After acquiring the camera pose, each frame of image or point cloud data can be projected into 3D space to obtain an initialized set of 3D Gaussian kernels. Each 3D Gaussian kernel is a high-dimensional parametric representation unit used to describe a local region in space. Specifically, the position parameter describes the coordinates of the center point of the Gaussian kernel in 3D space; the covariance matrix parameter describes the geometric extension range and directionality of the Gaussian kernel, thus reflecting the distribution characteristics of the local point cloud; the transparency parameter controls the visibility weight of the region during rendering or reconstruction; and the color parameter can be pixel color information extracted from the acquired image, which is used to describe the local appearance features of the vehicle surface.

[0052] By combining camera pose parameters, multi-view data is accurately projected into three-dimensional space, and a 3D Gaussian kernel set covering the vehicle's external contour is initialized. This enables a transition from raw images and / or point cloud data to a parametric 3D representation, providing a smoother, more continuous spatial distribution representation, allowing the vehicle's geometry and appearance information to be expressed in a more compact and controllable manner.

[0053] In step S230, the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted, and a white model of the target vehicle is generated based on the adjusted 3D Gaussian kernel set.

[0054] In some implementations, the parameters (position, covariance matrix, transparency, and color) of each Gaussian kernel are finely adjusted iteratively to better fit the actual shape of the target vehicle. After optimization, all Gaussian kernels together constitute a 3D Gaussian point cloud white model of the target vehicle, enabling the vehicle white model to fully preserve the three-dimensional geometry and surface features of the target vehicle, while avoiding topological defects or surface noise problems in traditional mesh reconstruction.

[0055] In this embodiment, the camera pose is recovered using the SfM algorithm, and the 3D Gaussian kernel set is initialized and optimized based on the acquired data, ultimately generating a 3D Gaussian point cloud white model of the target vehicle. This not only ensures the accuracy of the geometric structure and the realism of the appearance features, but also has good parameter adjustability.

[0056] Regarding the details of Gaussian kernel parameter adjustment in step S230, in some examples of embodiments of this application, a 3D Gaussian kernel set is projected onto an image plane to generate a rendered image, and the pixel-level error distribution between the rendered image and the acquired data is calculated. Specifically, the current 3D Gaussian kernel set is projected onto a two-dimensional image plane according to the camera pose parameters to generate a rendered image corresponding to the input acquired data. This rendering process not only considers the influence of the position and covariance matrix of each Gaussian kernel on the projected area, but also combines transparency and color parameters to calculate pixel coverage and color synthesis, thereby generating a visualization result of the vehicle's appearance at the pixel level.

[0057] Furthermore, based on the pixel-level error distribution, the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted in reverse. Specifically, the rendered image is compared pixel-by-pixel with the image in the actual acquired data to calculate the error distribution. For example, color difference metrics (such as RGB difference, perceptual consistency index) or photometric consistency loss functions can be used to reflect the deviation between the current 3D Gaussian kernel set and the real vehicle in terms of geometric structure and color appearance. The error distribution not only indicates the overall degree of difference but also reveals the underfitting of local areas (such as headlight edges, window frames, and roof surfaces). Subsequently, after obtaining the error distribution, the pixel-level error signal is progressively passed to the parameter update process of each 3D Gaussian kernel based on the backpropagation mechanism. Through iterative updates of the Gaussian kernel parameters, the error distribution tends to converge, so that the final generated 3D Gaussian point cloud white model can highly match the real vehicle in terms of geometric shape and visual appearance.

[0058] In some examples of embodiments of this application, the parameter gradient of each 3D Gaussian kernel is calculated based on the pixel-level error distribution, wherein the parameter gradient includes the position parameter gradient, the covariance matrix parameter gradient, the transparency parameter gradient, and the color parameter gradient.

[0059] Specifically, the error distribution not only reflects the overall deviation but also includes the contribution of each pixel to the reconstruction accuracy. Based on this, the pixel-level error can be mapped back to 3D space through an error backpropagation mechanism, thereby obtaining the gradient signal acting on each Gaussian kernel.

[0060] The position parameter gradient measures the direction and magnitude of the Gaussian kernel's offset from the real surface in space; when the projected position is inconsistent with the acquired image, the position parameter gradient guides the Gaussian kernel to adjust along the error direction. The covariance matrix parameter gradient reflects the morphological deviation of the Gaussian kernel in local geometric fitting; if a region has a large surface curvature but the Gaussian kernel distribution is too isotropic, the gradient will cause the covariance matrix to expand or contract in the principal curvature direction. The transparency parameter gradient is used to evaluate the effectiveness of the Gaussian kernel in rendering; when a Gaussian kernel contributes too little in multiple views or introduces occlusion errors, the gradient will reduce its transparency. The color parameter gradient is calculated from the pixel color residual and is used to guide the update of the Gaussian kernel's color components, making it more closely match the actual appearance of the vehicle surface. By deriving the parameter gradients through pixel-level error distribution, fine-grained optimization of each Gaussian kernel can be achieved.

[0061] Furthermore, the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted according to the parameter gradient, including at least one of the following: updating the position parameters of the corresponding 3D Gaussian kernel according to the position parameter gradient, and updating the covariance matrix parameters of the corresponding 3D Gaussian kernel according to the covariance matrix parameter gradient; updating the transparency parameter of the corresponding 3D Gaussian kernel according to the transparency parameter gradient, and deleting the corresponding 3D Gaussian kernel from the 3D Gaussian kernel set when the transparency parameter of the 3D Gaussian kernel is lower than the visibility threshold; updating the color parameter of the corresponding 3D Gaussian kernel according to the color parameter gradient.

[0062] Specifically, regarding position parameter updates, the spatial coordinates of the Gaussian kernel can be corrected using position gradients to ensure it gradually approximates the actual vehicle surface. For example, when a Gaussian kernel's projection falls outside the outer edge of a headlight, the position update guides it back to the correct boundary position. Regarding covariance matrix parameter updates, the expansion of the Gaussian kernel in different directions is adjusted via gradients to adaptively match local curved surfaces. For example, on the curved area of ​​the roof, the Gaussian kernel is stretched into an ellipsoid to better fit the curvature. Regarding transparency parameter updates, when the transparency gradient indicates that a Gaussian kernel contributes little to the rendering, its transparency is gradually reduced. If the transparency parameter falls below the visibility threshold, the Gaussian kernel can be removed from the set, dynamically eliminating redundant or noisy points. Regarding color parameter updates, the color components of the Gaussian kernel are corrected based on color gradients, making the projection of the vehicle surface more natural and consistent from different viewpoints, avoiding color deviations or discontinuities. Therefore, Gaussian kernel parameter updates driven by parameter gradients can refine and optimize the model locally without destroying the overall structure, maintaining modeling flexibility and accuracy.

[0063] In some examples of embodiments of this application, for each local pixel region in the pixel-level error distribution where the error value exceeds the structure optimization threshold, the 3D Gaussian kernel within the spatial range corresponding to the local pixel region is split into multiple sub-Gaussian kernels.

[0064] It should be noted that during the optimization of 3D Gaussian point cloud models, if certain local areas consistently exhibit large pixel-level errors between the rendered and acquired images, it often indicates that the existing Gaussian kernel distribution for that region is insufficient to represent the complexity of the real geometry. When such errors exceed the structural optimization threshold, the coverage area of ​​a single Gaussian kernel is too large or its shape is too simple, making it unable to fit the detailed features on the vehicle surface, such as the broken lines at the edges of headlights, the sharp transitions of window frames, or the fine stripes of the grille. Therefore, the Gaussian kernel splitting mechanism can effectively increase the expressive power of this region.

[0065] Specifically, when the error value of a local pixel region exceeds the structure optimization threshold, the corresponding range of the region in three-dimensional space can be located, the relevant Gaussian kernels within it can be marked, and these Gaussian kernels can be split.

[0066] In some implementations, the original Gaussian kernel is decomposed into multiple sub-Gaussian kernels. Each sub-Gaussian kernel inherits the core parameters of the original kernel but undergoes local perturbations and adjustments in terms of position, covariance matrix, and transparency to cover a finer-grained space. The sub-Gaussian kernels can be arranged according to the gradient direction of the error distribution or the principal local geometric direction (such as the direction of maximum curvature) to better capture the surface shape.

[0067] For example, the color parameters of the sub-Gaussian kernel can be initialized based on the color distribution of surrounding pixels to ensure natural transitions during multi-view rendering; the covariance matrix parameters are scaled according to the distribution of the original kernel to form a smaller coverage area, improving local resolution. Furthermore, to avoid blindly increasing the number of Gaussian kernels, the number of sub-Gaussian kernel splits can be controlled based on error intensity and region complexity. For instance, large planar areas of a vehicle body typically require only a few sub-kernels, while detailed areas such as the grille or the interior of headlights may be split into more sub-kernels.

[0068] By introducing a Gaussian kernel splitting mechanism, more sub-Gaussian kernels are generated in complex regions within the error set, enabling the model to accurately fit the geometric details and texture features of the vehicle. Furthermore, Gaussian kernel splitting is error-driven and does not add redundant kernels in globally irrelevant regions, improving local accuracy while maintaining modeling efficiency. Thus, the split Gaussian kernels form multi-scale coverage in the modeling region, representing both the global contour and capturing fine local structures, enhancing the consistency of the white model at both the global and local levels.

[0069] In some examples of embodiments of this application, in addition to introducing a Gaussian kernel splitting mechanism for high-error regions, a Gaussian kernel merging mechanism can also be set to control the overall number of kernels and improve modeling efficiency. Specifically, during the optimization iteration process, when multiple Gaussian kernels are detected to be spatially close, have similar covariance matrices, and have small differences in color and transparency parameters, it can be determined that these Gaussian kernels have redundant contributions to rendering. At this time, these Gaussian kernels are aggregated into a new merged Gaussian kernel. The parameters of the merged kernel can be calculated by weighted averaging or energy minimization to maintain the continuity of geometric and color representation.

[0070] This merging mechanism effectively reduces the total number of Gaussian kernels, avoiding excessive redundant kernels in locally flat or overlapping regions, thereby reducing storage and computational burden. Simultaneously, the splitting and merging mechanisms complement each other: splitting improves local accuracy in complex regions, while merging enhances overall efficiency in redundant regions, ultimately maintaining a balance between global and local representation of the vehicle white model, achieving both high precision and lightweight characteristics.

[0071] Regarding the details of extracting the bounding box parameters of the collision body from the vehicle white model, in some examples of the embodiments of this application, a mesh model corresponding to the vehicle white model is generated based on the spatial topology of the vehicle white model; then, the axial bounding box parameters of the mesh model are calculated to determine the corresponding bounding box parameters of the collision body.

[0072] It should be noted that collision detection is a crucial step in ensuring the realism of physical interactions during vehicle simulation. However, the vehicle white model itself is usually represented as a point cloud or Gaussian kernel set, which is not convenient for direct use in physics calculations. Therefore, in order to introduce the vehicle white model into the simulation engine, it is necessary to first extract its collider bounding box parameters. By approximating the model using bounding boxes, the computational load can be significantly reduced while maintaining detection accuracy.

[0073] Specifically, based on the spatial topology of the vehicle white model, surface reconstruction algorithms (such as Poisson reconstruction or Delaunay triangulation) are used to transform the point cloud or Gaussian kernel set into a continuous, closed polygonal mesh model (e.g., a mesh model). It should be understood that the continuity of local structures such as the vehicle body and wheels is ensured during mesh construction to avoid holes or non-manifold structures.

[0074] Furthermore, based on the generated mesh model, the minimum and maximum boundary points of the vehicle in each direction are calculated according to the coordinate system axes. By extracting these boundary points, the geometric parameters such as the length, width, and height of the bounding box are obtained, and the position center of the bounding box in three-dimensional space is determined. In subsequent simulation calls, the simulation engine directly uses the bounding box parameters for fast collision detection without the need for real-time parsing of complex meshes, significantly improving computational efficiency.

[0075] By converting the vehicle white model into a simplified axial bounding box, the complex mesh collision detection problem can be transformed into a simple geometric boundary calculation, significantly reducing computational overhead. Furthermore, it is independent of the specific vehicle modeling method (point cloud, mesh, or Gaussian kernel are all acceptable), making it applicable to vehicle white models from various sources. In addition, the extracted bounding box parameters can be directly used for dynamic simulation and collision detection, ensuring that the derived vehicle model possesses reliable collision detection capabilities during physical interactions.

[0076] Figure 3 A flowchart illustrating an example of determining matching vehicle materials in a vehicle model construction method according to an embodiment of this application is shown.

[0077] like Figure 3 As shown, in step S310, the collected data is preprocessed to generate a preprocessed image.

[0078] It should be noted that during the vehicle data acquisition process, the original images or video sequences may contain noise, uneven lighting, background interference, or distortion, which can affect the accuracy of subsequent 3D modeling and feature extraction. Therefore, various image preprocessing operations can be performed on the acquired data, such as image enhancement, target segmentation, or geometric correction, to generate cleaner and more stable image data (preprocessed images).

[0079] In step S320, based on the depth information in the acquired data, the preprocessed image is converted into a three-dimensional voxel mesh representation.

[0080] In some implementations, the image is projected into a three-dimensional space and voxelized by incorporating depth information from the acquired data. This is achieved by dividing the three-dimensional space into regular voxel units and mapping the depth information of the corresponding pixels to voxel coordinates, forming a dense three-dimensional voxel mesh representation.

[0081] In step S330, the shape features of the target vehicle are extracted from the three-dimensional voxel mesh representation. The shape features include at least one of the following: Hu geometric invariant moment features, aspect ratio features, number of wheel hubs features, number of contour concave and convex points features, and contour curvature features.

[0082] Here, based on the voxel mesh, shape features that characterize the vehicle's geometric appearance are extracted, focusing not only on the global appearance but also on local geometric details. For example, the Hu geometric invariant moment feature is used to describe a global shape that is insensitive to rotation, translation, and scaling, effectively distinguishing different vehicle body types; the aspect ratio feature calculates the vehicle's length, width, and height ratio based on the bounding box size of the voxel mesh, helping to identify different types such as sedans, SUVs, and trucks; the number of wheel hubs can be determined by detecting the geometric patterns of the wheel areas, effectively distinguishing two-axle, three-axle, or multi-axle vehicles; the number of contour protrusions and concavities can include the turning points and protrusion / concavity changes of the vehicle's contour lines to reflect the complexity of the vehicle's shape, such as the front grille and headlight contours; the contour curvature feature represents the smoothness and polygonal features of the body surface calculated through curvature distribution, used to distinguish between rounded and angular design styles. Thus, the vehicle's appearance can be characterized at different levels (global, local, and structural details), making the vehicle model recognition and matching more discriminative.

[0083] In step S340, the similarity between the shape features and the vehicle template features corresponding to each vehicle material in the vehicle material library is calculated, and the matching vehicle material is determined based on the calculated similarity.

[0084] In some implementations, Euclidean distance, cosine similarity, or weighted feature distance-based metrics are used to comprehensively calculate the similarity score between the target vehicle and each template. Furthermore, weighting coefficients can be set for different features (such as wheel moment and wheel hub count) to balance their contributions. Based on the similarity scores, one or more materials with the highest scores are selected as the matching results.

[0085] By matching shape features with templates in the resource library based on similarity, the system can automatically and efficiently find the closest available model for the target vehicle. Furthermore, the resource library can be expanded with new vehicle models at any time, and by continuously updating the template feature set, the accuracy and scalability of the modeling process are improved.

[0086] It should be noted that in this embodiment, shape features are not extracted from the vehicle white model for material matching. Instead, the three-dimensional voxel mesh representation of the vehicle is reconstructed using the collected data, and then shape features are extracted and material matching is performed based on this mesh representation. The vehicle white model has geometric continuity and deformability, but it may lack fine-grained structural information and spatial resolution in the early construction stage. If it is used directly for feature extraction, it is easy to cause problems such as insufficient shape representation or loss of local details.

[0087] In contrast, three-dimensional voxel mesh representation can start from depth information and use a unified spatial grid to discretize the overall appearance and local features of the vehicle, ensuring the integrity and stability of shape feature extraction. For example, geometric features such as the number of wheel hubs, contour curvature, and concave and convex points can obtain more intuitive and robust calculation results in voxel space, which can improve matching accuracy and robustness, and avoid recognition deviations caused by initial simplification or reconstruction errors of the white model.

[0088] In some examples of embodiments of this application, the vehicle material library pre-stores multiple vehicle materials and corresponding structured vehicle model parameters. Regarding the details of extracting vehicle model feature parameters, specifically, the structured vehicle model parameters corresponding to the matching vehicle materials can be retrieved from the vehicle material library to determine the corresponding vehicle model feature parameters. The structured vehicle model parameters include at least one of the following: vehicle type parameters, wheelbase feature parameters, track width feature parameters, window tilt angle feature parameters, and vehicle height feature parameters.

[0089] Specifically, vehicle type parameters, such as sedan, SUV, and truck, reflect the overall vehicle positioning; wheelbase characteristic parameters express the distance between the front and rear axles, used to determine the vehicle's longitudinal proportions; track characteristic parameters describe the distance between the left and right wheels, affecting the vehicle's lateral stability and width performance; window tilt characteristic parameters reflect the angle between the window and the vehicle body, which determines the vehicle's style characteristics (such as the difference between sporty and commercial models); and vehicle height characteristic parameters represent the height from the roof to the ground, used to differentiate the sense of space and passability of different vehicles.

[0090] Here, the vehicle material library not only stores various types of vehicle geometric models, but also assigns a set of structured vehicle model parameters to each vehicle material. These parameters are extracted from measurement data of actual vehicles or standard vehicle model databases, and have clear physical and design meanings, ensuring the accuracy and traceability of the data.

[0091] Through the embodiments of this application, pre-stored structured vehicle model parameters from the vehicle material library are introduced, and these parameters are called as the basis for deformation through matching relationships. By combining appearance matching with structural constraints, the generated derivative vehicle model not only closely resembles the target vehicle in appearance, but also maintains rationality in proportion and structure, so that the derivative modeling has stable engineering rationality and scalability.

[0092] Figure 4 The diagram illustrates an example of a vehicle model construction method according to an embodiment of this application, which involves deforming a white vehicle model to generate a derived vehicle model.

[0093] like Figure 4As shown, in step S410, when it is detected that the matching vehicle material belongs to the articulated vehicle category, the first Gaussian kernel subset and the second Gaussian kernel subset in the vehicle white model are labeled based on the semantic segmentation algorithm.

[0094] Here, the articulated vehicle category includes multiple vehicle modules, with a first Gaussian kernel subset corresponding to a first vehicle module and a second Gaussian kernel subset corresponding to a second vehicle module.

[0095] It should be noted that the biggest difference between articulated vehicles (such as trailers, articulated buses, etc.) and ordinary complete vehicles is that they are composed of multiple functional modules, and these modules have physical connections and degrees of freedom of relative movement. If the entire white model is deformed as a whole, the proportional relationship and connection method between the head and tail modules cannot be accurately represented. Therefore, different vehicle modules can be distinguished and modularly deformed in a controlled manner.

[0096] In some implementations, the current vehicle is detected to belong to the articulated vehicle category during the material matching stage, thereby triggering the modular segmentation process. Based on semantic segmentation algorithms (such as clustering segmentation methods combined with 3D Gaussian kernel distribution, or point cloud / voxel segmentation networks based on deep learning), the vehicle white model is analyzed and divided into different subsets corresponding to vehicle modules. The vehicle white model consists of a large number of 3D Gaussian kernels. By identifying feature regions such as the vehicle's centerline, wheel positions, cab, and cargo box, Gaussian kernel sets belonging to the head module (first subset) and tail module (second subset) are labeled. After segmentation, the two subsets maintain their internal geometric continuity and topological relationships, while achieving module-level separation within the overall white model.

[0097] In step S420, under the condition that the geometric shape of the constrained vehicle module is not distorted, the first Gaussian kernel subset and the second Gaussian kernel subset are subjected to separate deformation processing according to predefined deformation control parameters to generate a derived vehicle model.

[0098] Here, the modules of articulated vehicles (such as the cab and the trailer) often need to maintain their own structural integrity, but the relative positions and distances between modules may vary depending on the vehicle model. Therefore, it is necessary to perform separate deformation of the modules: allowing free adjustment between modules while constraining the internal geometry of individual modules to prevent disruption.

[0099] In some examples of embodiments of this application, the first vehicle module is a vehicle head module, and the second vehicle module is a vehicle rear module. Specifically, under the condition that the geometric shape of the vehicle module is constrained to be undistorted, taking the midpoint of the vehicle's centerline of the white vehicle model as a reference, positive and negative displacement vectors are applied to the first Gaussian kernel subset and the second Gaussian kernel subset respectively, referring to the long axis direction of the vehicle, according to predefined deformation control parameters, to generate a derived vehicle model.

[0100] Specifically, during the separation deformation, the vehicle's centerline is first determined, and its midpoint is set as a reference point. Based on predefined deformation control parameters, directional displacements are applied to the head module (first subset) and the tail module (second subset). A positive displacement vector is applied to the first subset along the vehicle's long axis; a negative displacement vector is applied to the second subset. Furthermore, rigid constraints are introduced during the displacement process to ensure that the Gaussian kernels within each subset maintain their relative positions, thereby ensuring the geometric integrity of each vehicle module. In some implementations, the magnitude of the displacement is determined by predefined control parameters, such as structured vehicle model parameters (e.g., hinge positions, module spacing ranges) stored in a vehicle material library.

[0101] In some implementations, dynamic deformation can be applied to 3D Gaussian white models of large engineering vehicles or tractors to achieve controllable dynamic separation and display of the head and tail sections. During the separation process, the system can ensure that the geometric accuracy of the head and tail sections deviates from the actual vehicle component dimensions by no more than 5%, while maintaining texture fidelity (PSNR ≥ 30dB), thus guaranteeing visualization effects and realism. The separation direction and distance can be flexibly adjusted (e.g., continuously controllable from 0-50cm) to meet the needs of separating and simulating the head and tail components of large vehicles.

[0102] Specifically, a semantic segmentation algorithm can be used to extract 2D semantic information from the input data, and combined with camera intrinsic and extrinsic parameters, it can be mapped to 3D space. The Gaussian point set is then partitioned and labeled to obtain a head Gaussian kernel subset G1 and a tail Gaussian kernel subset G2. Subsequently, independent deformation control parameters and driving rules are designed for G1 and G2 respectively, including separation direction vectors (e.g., positive / negative X-axis along the vehicle's length), separation distance thresholds, and deformation smoothing coefficients. These parameters control the spatial positions of G1 and G2, ensuring that they separate only along the set directions and controlling the continuity of motion during the separation process to avoid abrupt changes. During dynamic separation, the system uses the midpoint of the vehicle's centerline as a reference, applying positive / negative displacements to G1 and G2 respectively, and fine-tuning the covariance matrix of the Gaussian kernels to maintain local geometric consistency. Constraints are also introduced to ensure that key structures (such as the Gaussian kernel clusters of headlights and rear hooks) maintain the correct relative positions, preventing component misalignment. Therefore, through the above dynamic separation, not only can a derivative model conforming to structural logic be generated for conventional articulated vehicles, but also a controllable, accurate and stable modeling capability can be provided in the visualization and dynamic simulation of the front and rear separation of large vehicles.

[0103] Through the embodiments of this application, a separate deformation method is adopted for articulated vehicle models, which can flexibly adjust the module spacing and relative position of the articulated vehicle to generate derivative models that conform to different vehicle specifications, while keeping the geometric shape of each module intact, so that the generated model has visual realism and engineering rationality.

[0104] Figure 5 A flowchart illustrating another example of a vehicle model construction method according to an embodiment of this application is shown.

[0105] like Figure 5 As shown, in step S510, the data acquisition stage.

[0106] Specifically, multimodal data is collected from the target vehicle using modified data acquisition vehicles, XGrid handheld devices, or mobile phone cameras. For example, the data acquisition vehicle can circle the vehicle at a radius of approximately 3 meters and a speed of approximately 5 km / h, simultaneously acquiring high-resolution RGB images (e.g., 1920×1080@30fps) and LiDAR point clouds (32 lines). The obtained data provides multimodal input for subsequent 3D reconstruction and feature extraction, ensuring the integrity of spatial geometry and appearance texture.

[0107] In step S520, the white model generation stage.

[0108] Specifically, the acquired data is first preprocessed, and the intrinsic and extrinsic parameters of the camera are calculated using structure-of-motion (SfM) algorithms (such as COLMAP and vggsfm). Then, the target vehicle is modeled based on the 3D Gaussian Splatting (3DGS) algorithm.

[0109] In the initial stage of model reconstruction, radiation and density field modeling can be performed first. The density field describes the opacity of a point (x, y, z) in space; a higher value indicates a higher probability that the point belongs to the vehicle surface; when the density is 0, the point is air. Specifically, starting from a dense point cloud, the space is divided into small cubes through voxelization, and the point cloud distribution within each voxel is statistically analyzed to obtain the initial density field. The radiation field describes the color and brightness of a point in space along the viewing direction d. The processing logic combines the color information from multi-view images, infers the viewing direction for each point based on the camera pose, and assigns the color of the corresponding pixel to r(x, d), forming the initial radiation field. The joint representation of the density and radiation fields provides the basic geometric and photometric constraints for subsequent Gaussian kernel optimization.

[0110] First, an initial Gaussian kernel is generated, assigning a 3D Gaussian kernel to each point in the point cloud with a small initial covariance matrix. Then, rendering optimization is performed to minimize the pixel error between the Gaussian rendered image and the input multi-view image, updating parameters (position μ, covariance Σ, color c, transparency α) through backpropagation. Gaussian kernels with low transparency are removed, highly overlapping Gaussian kernels are merged, and kernel splitting is performed on regions with large errors to finely fit the local structure.

[0111] During the rendering stage, rasterization can be used to project and composite Gaussian kernels. Specifically, each 3D Gaussian kernel is projected onto a 2D image plane using camera intrinsic and extrinsic parameters to obtain a corresponding 2D elliptical representation. Then, sorted by depth to avoid occlusion errors, the color and transparency of the 2D ellipses are splatted pixel-by-pixel to generate the final rendered image. This rasterization mechanism ensures consistency and realism in multi-view rendering.

[0112] Finally, the 3DGS model is converted into a mesh, and the axial bounding box parameters are calculated based on the mesh for collision detection in subsequent physical simulations, ultimately resulting in a white model of the vehicle with a reconstruction accuracy of 0.5cm.

[0113] In step S530, the material matching stage.

[0114] In some implementations, the acquired images are first preprocessed, including grayscale conversion and denoising, and the target vehicle is converted into a 3D voxel mesh representation using depth information. Then, multi-level shape features are extracted from the voxel mesh, such as Hu geometric invariant moments, aspect ratio, contour curvature, and the number of wheel hubs. Finally, the extracted features are compared with template features in a vehicle material library using similarity calculations (e.g., Euclidean distance, cosine similarity), and the vehicle material with the highest similarity is selected as the matching result. The vehicle material library adopts a hierarchical storage architecture: the base layer stores general white models, the difference layer stores vehicle model feature parameters (wheelbase, window tilt angle, vehicle height, etc.), and the dynamic layer records the deformation sequence of the Gaussian kernel, achieving lightweight storage and fast retrieval.

[0115] In step S540, the dynamic deformation stage.

[0116] Here, the vehicle's white model is subjected to controllable deformation based on the vehicle's structural features and the parameters of the matching materials. Specifically, for ordinary vehicles, model personalization is achieved through global or local Gaussian kernel adjustments; for large engineering vehicles or articulated vehicles, semantic segmentation is used to partition and label the Gaussian kernel set in the white model, distinguishing the head Gaussian subset from the tail Gaussian subset; while maintaining the geometric integrity within the subsets, translational or separable deformations are applied to the head and tail respectively based on predefined deformation control parameters (separation direction, separation distance, smoothing coefficient, etc.), thereby achieving flexible reconstruction of the vehicle module. For example, based on the structural features of vehicles (such as tractors, trailers, etc.) (head: cab, engine hood area; tail: cargo box, rear chassis area), Gaussian kernel partitioning is performed. Independent deformation control parameters and driving rules are designed for the labeled head and tail Gaussian subsets to achieve separable dynamic deformation, while supporting adjustable separation distance and separation direction. This ensures the stability of the vehicle's local and overall geometry, while supporting dynamic visualization and simulation calls.

[0117] In step S550, the truth data output stage.

[0118] In dynamic simulation scenarios, when the derived vehicle model deforms or collides, the system performs physical collision detection based on bounding box parameters, generating corresponding collision response data. Simultaneously, it records the vehicle's deformation timeline and collision response sequence, forming a ground truth dataset that can be used for training and validation of autonomous driving systems. This dataset not only contains realistic feedback from geometric and physical interactions but is also bound to the dynamic deformation process, thus providing high-quality training samples for autonomous driving algorithms.

[0119] In some implementations, GPU Instancing technology can be further employed for batch rendering of vehicle models in the resource library, particularly in rendering and resource management. GPU Instancing loads the geometric and texture data of a vehicle model onto the graphics card at once, then reuses this data through instantiation, passing only the necessary transformation matrices and differential parameters (such as body color and material parameters) to each instance, thus avoiding repeated copying of the complete model data. Through this mechanism, even in city-level scenes requiring the simultaneous display of thousands of vehicles in different states, the video memory usage can be kept below 2GB, significantly reducing hardware load. Combined with the aforementioned hierarchical resource library design, this technology achieves lightweight scheduling and efficient rendering of vehicle models in large-scale scenes, providing a stable operating environment for dynamic simulation and interaction.

[0120] In this embodiment, a voxel matching algorithm is used to extract multiple structured vehicle DNA features, including aspect ratio, window tilt angle, and vehicle height, from the collected data, and these features are then parameterized and combined with a material library, supporting 10... 8 The generation of numerous derivative vehicle models significantly reduces scene repetition. An automated reconstruction process based on 3DGS reduces the time from single acquisition to model output to within 15 minutes, while maintaining a geometric accuracy of approximately 0.5cm. In terms of dynamic performance, Gaussian kernel semantic partitioning and separate deformation control ensure that the internal geometry of modules remains distorted and can achieve controllable separation and deformation according to preset parameters, thus balancing realism and simulation stability. At the system resource level, combining a hierarchical material library and GPU instantiation rendering enables the simultaneous display of thousands of vehicles on screen with a memory footprint of less than 2GB, thereby supporting large-scale, high-fidelity, and highly interactive simulation applications with lightweight overhead.

[0121] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0122] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform any of the above-described vehicle model construction methods.

[0123] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a vehicle model building method.

[0124] The apparatus described in the embodiments of this application can be used to execute the vehicle model construction method of the embodiments of this application, and accordingly achieve the technical effects achieved by the vehicle model construction method of the embodiments of this application, which will not be elaborated further here. In the embodiments of this application, the relevant functional modules can be implemented by a hardware processor.

[0125] Figure 6 This is a schematic diagram of the hardware structure of an electronic device for executing a vehicle model construction method according to another embodiment of this application, as shown below. Figure 6 As shown, the device includes: One or more processors 610 and memory 620, Figure 6 Take the 610 processor as an example.

[0126] The device for executing the vehicle model construction method may also include an input device 630 and an output device 640.

[0127] The processor 610, memory 620, input device 630, and output device 640 can be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.

[0128] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the vehicle model construction method in the embodiments of this application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 620, thereby implementing the vehicle model construction method of the above-described method embodiments.

[0129] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the device, etc. Furthermore, the memory 620 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include memory remotely located relative to the processor 610, and these remote memories may be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0130] Input device 630 can receive input digital or character information and generate signals related to user settings and function control of the device. Output device 640 may include display devices such as a display screen.

[0131] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, they execute the vehicle model construction method in any of the above method embodiments.

[0132] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0133] The electronic devices in this application embodiments exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include: smartphones (e.g., iPhones), multimedia phones, feature phones, and low-end phones, etc.

[0134] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as the iPad.

[0135] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players (such as iPods), handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0136] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0137] (5) Other electronic devices with data interaction functions.

[0138] In some embodiments, this application also provides a mobile platform on which the computer device described in any embodiment of this application is installed. The mobile platform includes, but is not limited to, vehicles, tracked robots, bipedal robots, quadrupedal robots, etc., wherein the vehicle can be a passenger car, pickup truck, truck, etc. It should be noted that the above are merely examples, and this application does not limit the specific form of the mobile platform.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for constructing a vehicle model, comprising: Based on the collected data of the target vehicle, a white model of the target vehicle is reconstructed and generated, and the bounding box parameters of the collision body of the white model are extracted. Extract the shape features of the target vehicle, and determine the corresponding matching vehicle material from the pre-built vehicle material library based on the extracted shape features; Extract the vehicle model feature parameters corresponding to the matched vehicle material, and deform the white vehicle model according to the vehicle model feature parameters to generate a derived vehicle model; Output a derived vehicle model carrying the bounding box parameters of the collider.

2. The method according to claim 1, wherein, The vehicle white model is a 3D Gaussian point cloud model; The step of reconstructing and generating a white model of the target vehicle based on the collected data of the target vehicle includes: The camera pose parameters of the acquired data are calculated using the structure-of-motion algorithm. A 3D Gaussian kernel set is initialized based on the camera pose parameters and the acquired data, wherein the 3D Gaussian kernel includes position parameters, covariance matrix parameters, transparency parameters, and color parameters; The parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted, and a white model of the target vehicle is generated based on the adjusted 3D Gaussian kernel set.

3. The method according to claim 2, wherein, The adjustment of the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set includes: A 3D Gaussian kernel set is projected onto the image plane to generate a rendered image, and the pixel-level error distribution between the rendered image and the acquired data is calculated. Based on the pixel-level error distribution, the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted in reverse.

4. The method according to claim 3, wherein, The step of adjusting the parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set in reverse according to the pixel-level error distribution includes: The parameter gradient of each 3D Gaussian kernel is calculated based on the pixel-level error distribution, wherein the parameter gradient includes the position parameter gradient, the covariance matrix parameter gradient, the transparency parameter gradient, and the color parameter gradient. The parameters of each 3D Gaussian kernel in the 3D Gaussian kernel set are adjusted according to the parameter gradient, including at least one of the following: The position parameters of the corresponding 3D Gaussian kernel are updated according to the gradient of the position parameters, and the covariance matrix parameters of the corresponding 3D Gaussian kernel are updated according to the gradient of the covariance matrix parameters. The transparency parameter of the corresponding 3D Gaussian kernel is updated according to the transparency parameter gradient, and when the transparency parameter of the 3D Gaussian kernel is lower than the visibility threshold, the corresponding 3D Gaussian kernel is deleted from the 3D Gaussian kernel set. The color parameters of the corresponding 3D Gaussian kernel are updated according to the color parameter gradient.

5. The method according to claim 1, wherein, The extraction of the collision bounding box parameters of the vehicle white model includes: Based on the spatial topology of the vehicle white model, a mesh model corresponding to the vehicle white model is generated; Calculate the axial bounding box parameters of the mesh model to determine the corresponding collider bounding box parameters.

6. The method according to claim 1, wherein, The step of extracting the shape features of the target vehicle and determining the corresponding matching vehicle material from a pre-built vehicle material library based on the extracted shape features includes: The collected data is preprocessed to generate a preprocessed image; Based on the depth information in the acquired data, the preprocessed image is converted into a three-dimensional voxel mesh representation; The shape features of the target vehicle are extracted from the three-dimensional voxel mesh representation, and the shape features include at least one of the following: Hu geometric invariant moment features, aspect ratio features, number of wheel hub features, number of contour concave and convex points features, and contour curvature features. Calculate the similarity between the shape feature and the vehicle template feature corresponding to each vehicle material in the vehicle material library, and determine the matching vehicle material based on the calculated similarity.

7. The method according to claim 1, wherein, The vehicle material library pre-stores multiple vehicle materials and corresponding structured vehicle model parameters; The step of extracting the vehicle model feature parameters corresponding to the matched vehicle material includes: The structured vehicle model parameters corresponding to the matching vehicle material are retrieved from the vehicle material library to determine the corresponding vehicle model feature parameters; the structured vehicle model parameters include at least one of the following: vehicle model category parameters, wheelbase feature parameters, track width feature parameters, window tilt angle feature parameters, and vehicle height feature parameters.

8. The method according to claim 2, wherein, The step of deforming the vehicle white model according to the vehicle model feature parameters to generate a derived vehicle model includes: When the matched vehicle material is detected to belong to the articulated vehicle category, the first Gaussian kernel subset and the second Gaussian kernel subset in the vehicle white model are labeled based on the semantic segmentation algorithm; the articulated vehicle category contains multiple vehicle modules, the first Gaussian kernel subset corresponds to the first vehicle module, and the second Gaussian kernel subset corresponds to the second vehicle module; Under the condition that the geometric shape of the vehicle module is not distorted, the first Gaussian kernel subset and the second Gaussian kernel subset are subjected to separate deformation processing according to predefined deformation control parameters to generate a derived vehicle model.

9. The method according to claim 8, wherein, The first vehicle module is a vehicle head module, and the second vehicle module is a vehicle rear module; Under the condition that the geometric shape of the constrained vehicle module is not distorted, the first Gaussian kernel subset and the second Gaussian kernel subset are subjected to deformation displacement processing according to predefined deformation control parameters to generate a derived vehicle model, including: Under the condition that the geometric shape of the vehicle module is not distorted, the midpoint of the vehicle's centerline of the white vehicle model is used as a reference. According to the predefined deformation control parameters, positive displacement vectors and negative displacement vectors are applied to the first Gaussian kernel subset and the second Gaussian kernel subset respectively in the direction of the vehicle's long axis to generate a derived vehicle model.

10. The method according to any one of claims 1-9, further comprising: During the dynamic simulation of the vehicle driving environment, the derived vehicle model is subjected to deformation processing based on events triggered by the simulation scenario. Physical collision detection is performed based on the bounding box parameters of the collider, and collision response data is generated; The deformation time series data and collision response data of the derived vehicle model are associated and recorded with the simulation scene trigger events to generate a ground truth dataset for training the autonomous driving system.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein, The processor executes the computer program to implement the steps of the method according to any one of claims 1-10.

12. A storage medium storing one or more programs including executable instructions, wherein, The execution instructions can be read and executed by an electronic device to perform the steps of the method according to any one of claims 1-10.

Images