A multi-scale traffic facility three-dimensional model efficient rendering method and system

Abstract

The application discloses a kind of multi-scale traffic facilities three-dimensional model high-efficiency rendering method and system, it is related to computer graphics and three-dimensional visualization technical field.For the problems that traffic three-dimensional scene static LOD rendering is prone to picture jump, rendering load fluctuation is violent, traffic key semantic simplification loss, the application is first based on camera motion parameter and constructs dynamic three-level sector partition;Then through lightweight CNN extraction semantic mask, fuse semantic weight constraint edge folding simplification, 100-300ms interpolation is realized to smooth LOD switching;Finally, construct quantization rendering cost model, rely on double-threshold closed-loop control dynamic adjustment model simplification degree.The application gives consideration to model light weight and traffic feature fidelity, stable rendering frame rate, and can be widely applied to urban traffic simulation, road digital twin, vehicle-mounted real scene visualization and the like.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and 3D visualization technology, specifically to an efficient rendering method and system for multi-scale 3D models of transportation facilities. Background Technology

[0002] With the rapid development of digital twin, intelligent transportation, and autonomous driving simulation technologies, real-time rendering of large-scale 3D traffic facility scenes has become a core technological requirement in the field of digital transportation. Traffic scenes such as urban roads, highways, and interchanges often contain massive geometric models, textures, and dynamic elements, placing extremely high demands on rendering performance at high resolutions such as 4K.

[0003] Currently, 3D rendering engines commonly use Level of Detail (LOD) technology to balance rendering quality and performance. However, traditional static LOD solutions have significant drawbacks: First, switching LOD at a fixed viewing distance can cause abrupt changes in model details at mid-range perspectives, resulting in a noticeable visual jump for users, severely impacting the visual experience and increasing the cognitive cost of the scene. Second, changes in both the number of models and the level of detail in dynamic scenes can cause significant fluctuations in rendering load, leading to unstable rendering frame rates and even stuttering.

[0004] In existing technologies, while continuous LOD generation schemes can alleviate visual jumps, they consume a large amount of CPU / GPU computing resources, making it difficult to meet real-time rendering requirements. Bounding box-based coarse-grained culling schemes can only solve the problem of clipping models outside the viewport, failing to address load abrupt changes within the mid-range. GPU-instantiated rendering schemes, while reducing drawing call overhead, are heavily reliant on hardware performance and cannot perform differentiated optimization for the semantic features of traffic facilities. Furthermore, existing schemes generally lack mechanisms to protect critical semantic information such as traffic signs, traffic lights, and lane lines, easily leading to the loss of important traffic facility features during model simplification, affecting the scene's recognizability and usability.

[0005] Therefore, achieving high-performance and stable rendering of large-scale traffic facility scenes while ensuring visual continuity and the integrity of key semantic features has become a core technical problem that urgently needs to be solved in the field of traffic 3D visualization. This field requires a technical solution that can dynamically adjust the spatial partitioning strategy according to the viewpoint's motion state, protect traffic semantic features during geometric simplification, and achieve elastic load balancing of rendering through a closed-loop feedback mechanism. Summary of the Invention

[0006] To address the technical challenges of static LOD visual jumps, large rendering load fluctuations, and easy loss of key semantic features in the rendering of large-scale 3D transportation facility scenes, this invention provides an efficient rendering method for multi-scale 3D transportation facility models, comprising:

[0007] S1. Real-time acquisition of camera pose parameters and motion state data, and based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output.

[0008] S2. Based on the three-level partitioned scene model set, perform semantic feature-sensitive progressive geometric simplification on each region model to generate a multi-resolution LOD model set with semantic protection constraints.

[0009] S3. Based on the multi-resolution LOD model set, monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame. Dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism and output the load-balanced rendering parameter set.

[0010] S4. Receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block-based streaming loading and differential rendering of dynamic and static objects, and output a visually continuous 3D rendering result.

[0011] Furthermore, S1 specifically includes:

[0012] S11. Real-time acquisition of camera viewpoint coordinates (x, y, z), pitch angle, yaw angle and velocity vector, combined with Kalman filtering to predict the view frustum space range of the next 3 frames;

[0013] S12. The near-scenic area is divided into a 0-50m range. An improved view cone culling algorithm is used to load a fully detailed model. The geometric subdivision of the curve area is automatically increased by calculating the road curvature change rate in real time. The formula for calculating the curvature change rate is:

[0014]

[0015] in, The rate of change of curvature, The tangent angle of the road. It is the arc length;

[0016] Further calculation of curvature enhancement coefficient:

[0017]

[0018] When curvature enhancement coefficient At that time, the subdivision level of the corresponding regional model will be increased to twice the base value;

[0019] S13. Divide the mid-range area into 50-200m zones, construct a deformable fan-shaped spatial index structure, and adjust the index based on the camera's movement speed. and yaw rate The dynamic adjustment of the sector angle is calculated using the following formula:

[0020]

[0021] in, Based on the angle, , These are the experimental calibration coefficients;

[0022] The boundary of the sector-shaped region is defined by parametric equations:

[0023]

[0024] in, The reference radius of the sector-shaped foundation. The boundary of the deformable sector-shaped spatial index structure. For the camera yaw angle, The deformation coefficient, It is the attenuation factor;

[0025] S14. Divide the distant area into regions beyond 200m, pre-calculate the road network topology G=(V,E), and generate Bézier surrogate models with continuous curvature constraints for non-critical nodes. The Bézier curve control points satisfy:

[0026] ,

[0027] in, On the Bézier curve The coordinates of the three-dimensional point corresponding to the parameter. For curve parameters, For control point numbers, For the coordinates of the control points, The order of the curve.

[0028] Furthermore, S2 specifically includes:

[0029] S21. Receive the set of three-level partition scene models output by S1, deploy a lightweight CNN network to extract semantic features from the models, identify the shape of traffic signs, the light-emitting area of ​​traffic lights, and the key areas of lane lines, and output a semantic feature protection mask M(x,y,z).

[0030] S22. Receive the semantic feature protection mask M(x,y,z), introduce a semantic weight factor on the basis of the traditional quadratic error metric, perform constrained edge folding optimization, and output the geometrically simplified multi-resolution model primitive.

[0031] The optimization formula for performing constrained edge folding is as follows:

[0032]

[0033] in, To measure the optimization error after introducing semantic constraints. It is a traditional quadratic error measure; This is a balance coefficient between geometric error and semantic features; Weights for semantic feature regions This represents the change in the normal of the triangular facet. A set of facets covered by a mask to protect semantic features;

[0034] S23. Receive the multi-resolution model primitives, establish a one-to-one mapping relationship between vertices between adjacent LOD levels, realize the gradient transition of geometric features through bilinear interpolation, and output a set of multi-resolution LOD models with semantic protection constraints.

[0035] The interpolation formula is:

[0036] ,

[0037] in, For the LOD switching transition time The corresponding interpolated 3D vertex coordinates, , These represent the vertex positions of two adjacent LOD levels. , Transition start and end times, transition time Keep it within the 100-300ms range.

[0038] Furthermore, S3 specifically includes:

[0039] S31. Receive the multi-resolution LOD model set, collect the number of drawing calls per frame, average number of vertices, average number of edges, and shader overhead indicators in real time through GPU performance counters, establish a rendering cost model, and output the quantized rendering cost value per frame.

[0040] The rendering cost model is as follows:

[0041]

[0042] in, For the first Total frame rendering cost , , The calibration parameters correspond to the overhead of drawing calls, vertex processing, and shader computation, respectively. To determine the number of calls to the drawing; , These represent the average number of vertices and the average number of edges in the model, respectively. Calculate the overhead metrics for the shaders;

[0043] S32. Based on the quantized rendering cost value per frame, set a high load threshold. and low load threshold Perform dual-threshold hysteresis adaptive control: when the rendering cost exceeds the high load threshold, increase the simplification rate of the mid-range area; when the rendering cost is lower than the low load threshold, activate the reverse refinement of the distant proxy model and output the dynamically adjusted model simplification parameters.

[0044] The adjustment formula for increasing the simplification rate of the mid-range scene when the rendering cost exceeds the high load threshold is as follows:

[0045]

[0046] in, To simplify the dynamic model of the mid-range scenic area, Let k = 0.05 be the initial simplification rate, and k = 0.05 be the adjustment coefficient. ;

[0047] The update formula for activating the reverse refinement of the distant proxy model when the rendering cost is lower than the low load threshold is:

[0048]

[0049] in, The total number of vertices in the distant Bézier proxy model after adjusting for low load. To ensure that the total number of existing vertices in the prospective proxy model before this refinement operation is ≤15%, the rendering load fluctuation range is determined.

[0050] S33. Receive the dynamic adjustment parameters output by S32 and set abnormal state monitoring: if the rendering cost of 5 consecutive frames exceeds 1.2 times the high threshold, then forcibly reduce the rendering resolution to 75% of the original resolution until the load returns to normal, and output the final load-balanced rendering parameter set.

[0051] Furthermore, S4 specifically includes:

[0052] S41. Receive the rendering parameter set and the adjusted multi-resolution LOD model, use block-based streaming loading technology to divide the scene data into 32m×32m spatial blocks, load high-priority area data 2-3 frames in advance by predicting the line of sight direction, and output the scene data stream after streaming loading.

[0053] S42. Receive the scene data stream and perform differentiated rendering processing on dynamic and static elements in the scene: use GPU instantiation rendering for moving vehicles, enable batch merging drawing for static facilities such as traffic signs, markings, and traffic lights, and output the rendering instruction stream after API call optimization.

[0054] S43. Based on the rendering instruction stream, integrate the rendering parameters to perform the final rasterization rendering and output a visually continuous 3D rendering result with a visual similarity SSIM value of not less than 0.9.

[0055] A high-efficiency rendering system for multi-scale 3D models of transportation facilities is also provided, based on the method described above, including...

[0056] The dynamic spatial partitioning module is used to collect camera pose parameters and motion state data in real time. Based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output.

[0057] The semantic feature-preserving LOD simplification module is used to perform semantic feature-sensitive progressive geometric simplification on each region model based on the three-level partitioned scene model set, and generate a multi-resolution LOD model set with semantic protection constraints.

[0058] The elastic load closed-loop control module is used to monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame based on the multi-resolution LOD model set, dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism, and output the rendering parameter set after load balancing.

[0059] The scene streaming differential rendering module is used to receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block streaming loading and differential rendering of dynamic and static objects, and output visually continuous 3D rendering results.

[0060] The beneficial effects of this invention are:

[0061] This invention overcomes the limitations of traditional static LOD fixed-distance switching by constructing a dynamic spatial partitioning engine for cone motion perception. It integrates camera kinematic parameters (velocity, angular velocity) into the spatial partitioning strategy, enabling the fan-shaped angle of the mid-range area to dynamically adjust with the movement of the viewpoint. This significantly reduces the visual abruptness phenomenon at mid-range perspectives and solves the visual jump problem in static LOD.

[0062] This invention establishes a closed-loop feedback control system for GPU rendering load, and achieves dynamic balancing of rendering load through a dual-threshold elastic adjustment mechanism. In addition, this invention significantly improves rendering performance while ensuring visual quality, and provides an efficient and reliable technical solution for real-time visualization of large-scale traffic digital twin scenarios. Attached Figure Description

[0063] Figure 1 This is a flowchart of the present invention;

[0064] Figure 2 This is a schematic diagram of a semantically preserved geometrically simplified pipeline.

[0065] Figure 3 This is a schematic diagram of the closed-loop regulation of a flexible load balancing control system.

[0066] Figure 4 This is a schematic diagram showing the dynamic spatial distribution of the near-field, mid-field, and far-field areas. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, process equipment or devices not specifically specified are all conventional equipment or devices in the art. Unless specifically specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.

[0068] Example 1, combined with Figure 1 This embodiment describes an efficient rendering method for multi-scale 3D models of transportation facilities, including:

[0069] S1. Real-time acquisition of camera pose parameters and motion state data, and based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output.

[0070] S2. Based on the three-level partitioned scene model set, perform semantic feature-sensitive progressive geometric simplification on each region model to generate a multi-resolution LOD model set with semantic protection constraints.

[0071] S3. Based on the multi-resolution LOD model set, monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame. Dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism and output the load-balanced rendering parameter set.

[0072] S4. Receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block-based streaming loading and differential rendering of dynamic and static objects, and output a visually continuous 3D rendering result.

[0073] Furthermore, S1 specifically includes:

[0074] S11. Real-time acquisition of camera viewpoint coordinates (x, y, z), pitch angle, yaw angle and velocity vector, combined with Kalman filtering to predict the view frustum space range of the next 3 frames;

[0075] S12. The near-scenic area is divided into a 0-50m range. An improved view cone culling algorithm is used to load a fully detailed model. The geometric subdivision of the curve area is automatically increased by calculating the road curvature change rate in real time. The formula for calculating the curvature change rate is:

[0076]

[0077] in, The rate of change of curvature, The tangent angle of the road. It is the arc length;

[0078] Further calculation of curvature enhancement coefficient:

[0079]

[0080] When curvature enhancement coefficient At that time, the subdivision level of the corresponding regional model will be increased to twice the base value;

[0081] Specifically, in combination Figure 2 As can be seen, the spatial regions are divided into near, middle, and far zones along the main viewing direction, with the camera as the vertex of the view cone. The near zone uses a high-precision model, the middle zone uses a simplified model, and the far zone uses a topological proxy model. The near zone, as the area most sensitive to the user's vision, is responsible for presenting the details of the road ahead from the driver's perspective. Traditional solutions use a uniform subdivision level for all near-view models, causing unnecessary performance waste on straight road sections and lane line distortion due to insufficient detail on curved road sections. This step detects curved areas in real time through the rate of curvature change. When the curvature enhancement coefficient exceeds a threshold, the mesh subdivision level is automatically increased, ensuring the geometric accuracy of traffic facilities such as lane lines and guardrails at curves while avoiding performance waste on straight road sections. The full-detail model set of the near zone output in this step will be directly fed into the semantic simplification pipeline of S2 for protection processing.

[0082] S13. Divide the mid-range area into 50-200m zones, construct a deformable fan-shaped spatial index structure, and adjust the index based on the camera's movement speed. and yaw rate The dynamic adjustment of the sector angle is calculated using the following formula:

[0083]

[0084] in, Based on the angle, , These are the experimental calibration coefficients;

[0085] The boundary of the sector-shaped region is defined by parametric equations:

[0086]

[0087] in, The reference radius of the sector-shaped foundation. The boundary of the deformable sector-shaped spatial index structure. For the camera yaw angle, The deformation coefficient, It is the attenuation factor;

[0088] Specifically, the mid-range area is where visual jumps and load fluctuations are most likely to occur. Traditional solutions use fixed circular or rectangular areas for mid-range processing. When the camera moves at high speed or turns rapidly, the model density in the main field of view is insufficient, while the side fields of view are loaded with too much unnecessary detail. The deformable fan-shaped index structure designed in this step can dynamically adjust the angle and shape according to the camera's motion: when the vehicle is moving at high speed, the fan angle is automatically increased to ensure the model density in the front field of view; when the vehicle turns, the fan orientation is adjusted according to the yaw rate to improve the rendering quality in the main field of view. The mid-range model set output in this step will serve as the main control object for S3 load adjustment.

[0089] S14. Divide the distant area into regions beyond 200m, pre-calculate the road network topology G=(V,E), and generate Bézier surrogate models with continuous curvature constraints for non-critical nodes. The Bézier curve control points satisfy:

[0090] ,

[0091] in, On the Bézier curve The coordinates of the three-dimensional point corresponding to the parameter. For curve parameters, For control point numbers, For the coordinates of the control points, The order of the curve.

[0092] Specifically, although the distant view model is far from the viewpoint, it plays a crucial role in maintaining the visual continuity of the road network. Traditional solutions often involve direct cropping or significant simplification of the distant view model, which can easily lead to visual breaks in the road network and severely impact the realism of the scene. This step, by pre-calculating the road network topology connectivity, generates Bézier curve proxy models for non-critical nodes such as straight road segments, ensuring the visual continuity of the distant road network. The distant view proxy model output from this step will undergo reverse refinement under low load conditions in S3, further improving the visual quality of the distant view.

[0093] Furthermore, S2 specifically includes:

[0094] S21. Receive the set of three-level partition scene models output by S1, deploy a lightweight CNN network to extract semantic features from the models, identify the shape of traffic signs, the light-emitting area of ​​traffic lights, and the key areas of lane lines, and output a semantic feature protection mask M(x,y,z).

[0095] S22. Receive the semantic feature protection mask M(x,y,z), introduce a semantic weight factor on the basis of the traditional quadratic error metric, perform constrained edge folding optimization, and output the geometrically simplified multi-resolution model primitive.

[0096] The optimization formula for performing constrained edge folding is as follows:

[0097]

[0098] in, To measure the optimization error after introducing semantic constraints. It is a traditional quadratic error measure; This is a balance coefficient between geometric error and semantic features; Weights for semantic feature regions This represents the change in the normal of the triangular facet. A set of facets covered by a mask to protect semantic features;

[0099] S23. Receive the multi-resolution model primitives, establish a one-to-one mapping relationship between vertices between adjacent LOD levels, realize the gradient transition of geometric features through bilinear interpolation, and output a set of multi-resolution LOD models with semantic protection constraints.

[0100] The interpolation formula is:

[0101] ,

[0102] in, For the LOD switching transition time The corresponding interpolated 3D vertex coordinates, , These represent the vertex positions of two adjacent LOD levels. , Transition start and end times, transition time Keep it within the 100-300ms range.

[0103] Specifically, such as Figure 3 As shown in the figure, this diagram illustrates the three-stage complete processing flow for generating semantically constrained multi-resolution LOD models according to the present invention. S2 first utilizes a lightweight CNN to extract key semantic regions of traffic facilities to generate a protective mask. Then, semantic weights are introduced based on the mask to correct the traditional secondary error metric, simplifying the constraint edges through folding. This reduces the number of model faces while protecting core semantic features such as traffic signs and traffic lights. Finally, linear vertex interpolation is used to construct a smooth transition effect between LOD models of different precisions. The entire process achieves automatic generation of multi-resolution LOD models that balance semantic integrity, topological invariance, and visual transitions without any abrupt changes, reducing the geometric overhead of subsequent scene rendering.

[0104] Furthermore, S3 specifically includes:

[0105] S31. Receive the multi-resolution LOD model set, collect the number of drawing calls per frame, average number of vertices, average number of edges, and shader overhead indicators in real time through GPU performance counters, establish a rendering cost model, and output the quantized rendering cost value per frame.

[0106] The rendering cost model is as follows:

[0107]

[0108] in, For the first Total frame rendering cost , , The calibration parameters correspond to the overhead of drawing calls, vertex processing, and shader computation, respectively. To determine the number of calls to the drawing; , These represent the average number of vertices and the average number of edges in the model, respectively. Calculate the overhead metrics for the shaders;

[0109] S32. Based on the quantized rendering cost value per frame, set a high load threshold. and low load threshold Perform dual-threshold hysteresis adaptive control: when the rendering cost exceeds the high load threshold, increase the simplification rate of the mid-range area; when the rendering cost is lower than the low load threshold, activate the reverse refinement of the distant proxy model and output the dynamically adjusted model simplification parameters.

[0110] The adjustment formula for increasing the simplification rate of the mid-range scene when the rendering cost exceeds the high load threshold is as follows:

[0111]

[0112] in, To simplify the dynamic model of the mid-range scenic area, Let k = 0.05 be the initial simplification rate, and k = 0.05 be the adjustment coefficient. ;

[0113] The update formula for activating the reverse refinement of the distant proxy model when the rendering cost is lower than the low load threshold is:

[0114]

[0115] in, The total number of vertices in the distant Bézier proxy model after adjusting for low load. To ensure that the total number of existing vertices in the prospective proxy model before this refinement operation is ≤15%, the rendering load fluctuation range is determined.

[0116] S33. Receive the dynamic adjustment parameters output by S32 and set abnormal state monitoring: if the rendering cost of 5 consecutive frames exceeds 1.2 times the high threshold, then forcibly reduce the rendering resolution to 75% of the original resolution until the load returns to normal, and output the final load-balanced rendering parameter set.

[0117] Specifically, Figure 4 The flowchart of the rendering load closed-loop control process of this invention is shown. This step collects various rendering overhead indicators in real time through GPU performance counters and constructs a quantitative rendering cost model to obtain the rendering cost value of a single frame. Based on high and low load thresholds, hysteresis adaptive control is carried out. When the load is too high, the simplification rate of the mid-range scene model is increased. When the load is too low, the vertex scale of the distant Bezier proxy model is refined to control the load fluctuation to no more than 15%. At the same time, an overload anomaly monitoring mechanism is added. When multiple consecutive frames are severely overloaded, the rendering resolution is automatically compressed. The entire closed-loop control logic dynamically adjusts the model simplification parameters of each region in real time and finally outputs balanced rendering parameters. This stabilizes the GPU rendering load, eliminates drastic frame rate fluctuations, and ensures continuous and stable rendering of large-scale traffic scenes.

[0118] Furthermore, S4 specifically includes:

[0119] S41. Receive the rendering parameter set and the adjusted multi-resolution LOD model, use block-based streaming loading technology to divide the scene data into 32m×32m spatial blocks, load high-priority area data 2-3 frames in advance by predicting the line of sight direction, and output the scene data stream after streaming loading.

[0120] S42. Receive the scene data stream and perform differentiated rendering processing on dynamic and static elements in the scene: use GPU instantiation rendering for moving vehicles, enable batch merging drawing for static facilities such as traffic signs, markings, and traffic lights, and output the rendering instruction stream after API call optimization.

[0121] S43. Based on the rendering instruction stream, integrate the rendering parameters to perform the final rasterization rendering and output a visually continuous 3D rendering result with a visual similarity SSIM value of not less than 0.9.

[0122] Specifically, large-scale scenes cannot be loaded into memory all at once. This step uses standard 32m×32m blocks for data management. Based on the camera motion trajectory predicted by Kalman filtering, the block data in front of the viewpoint is preloaded 2-3 frames in advance, achieving "streaming" data loading and avoiding stuttering during block switching. Differentiated rendering strategies are adopted for the different characteristics of dynamic and static elements in the scene: moving vehicles are numerous but have the same geometry, suitable for GPU instantiation rendering; static traffic facilities have fixed positions, suitable for batch merging rendering. Through dynamic and static divide-and-conquer processing, the API call overhead from CPU to GPU is significantly reduced.

[0123] A high-efficiency rendering system for multi-scale 3D models of transportation facilities is also provided, based on the method described above, including...

[0124] The dynamic spatial partitioning module is used to collect camera pose parameters and motion state data in real time. Based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output.

[0125] The semantic feature-preserving LOD simplification module is used to perform semantic feature-sensitive progressive geometric simplification on each region model based on the three-level partitioned scene model set, and generate a multi-resolution LOD model set with semantic protection constraints.

[0126] The elastic load closed-loop control module is used to monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame based on the multi-resolution LOD model set, dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism, and output the rendering parameter set after load balancing.

[0127] The scene streaming differential rendering module is used to receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block streaming loading and differential rendering of dynamic and static objects, and output visually continuous 3D rendering results.

Claims

1. A method for efficiently rendering multi-scale 3D models of transportation facilities, characterized in that, include: S1. Real-time acquisition of camera pose parameters and motion state data, and based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output. S2. Based on the three-level partitioned scene model set, perform semantic feature-sensitive progressive geometric simplification on each region model to generate a multi-resolution LOD model set with semantic protection constraints. S3. Based on the multi-resolution LOD model set, monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame. Dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism and output the load-balanced rendering parameter set. S4. Receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block-based streaming loading and differential rendering of dynamic and static objects, and output a visually continuous 3D rendering result. S1 specifically includes: S11. Real-time acquisition of camera viewpoint coordinates (x, y, z), pitch angle, yaw angle and velocity vector, combined with Kalman filtering to predict the view frustum space range of the next 3 frames; S12. The near-scenic area is divided into a 0-50m range. An improved view cone culling algorithm is used to load a fully detailed model. The geometric subdivision of the curve area is automatically increased by calculating the road curvature change rate in real time. The formula for calculating the curvature change rate is: ； in, The rate of change of curvature, The tangent angle of the road. It is the arc length; Further calculation of curvature enhancement coefficient: ； When curvature enhancement coefficient At that time, the subdivision level of the corresponding regional model will be increased to twice the base value; S13. Divide the mid-range area into 50-200m zones, construct a deformable fan-shaped spatial index structure, and adjust the index based on the camera's movement speed. and yaw rate The dynamic adjustment of the sector angle is calculated using the following formula: ； in, Based on the angle, , These are the experimental calibration coefficients; The boundary of the sector-shaped region is defined by parametric equations: ； in, The reference radius of the sector-shaped foundation. The boundary of the deformable sector-shaped spatial index structure. For the camera yaw angle, The deformation coefficient, It is the attenuation factor; S14. Divide the distant area into regions beyond 200m, pre-calculate the road network topology G=(V,E), and generate Bézier surrogate models with continuous curvature constraints for non-critical nodes. The Bézier curve control points satisfy: , ； in, On the Bézier curve The coordinates of the three-dimensional point corresponding to the parameter. For curve parameters, For control point numbers, For the coordinates of the control points, The order of the curve; S2 specifically includes: S21. Receive the set of three-level partition scene models output by S1, deploy a lightweight CNN network to extract semantic features from the models, identify the shape of traffic signs, the light-emitting area of ​​traffic lights, and the key areas of lane lines, and output a semantic feature protection mask M(x,y,z). S22. Receive the semantic feature protection mask M(x,y,z), introduce a semantic weight factor on the basis of the traditional quadratic error metric, perform constrained edge folding optimization, and output the geometrically simplified multi-resolution model primitive. The optimization formula for performing constrained edge folding is as follows: ； in, To measure the optimization error after introducing semantic constraints. It is a traditional quadratic error measure; This is a balance coefficient between geometric error and semantic features; Weights for semantic feature regions This represents the change in the normal of the triangular facet. A set of facets covered by a mask to protect semantic features; S23. Receive the multi-resolution model primitives, establish a one-to-one mapping relationship between vertices between adjacent LOD levels, realize the gradient transition of geometric features through bilinear interpolation, and output a set of multi-resolution LOD models with semantic protection constraints. The interpolation formula is: ， ； in, For the LOD switching transition time The corresponding interpolated 3D vertex coordinates, , These represent the vertex positions of two adjacent LOD levels. , Transition start and end times, transition time Keep it within the 100-300ms range.

2. The efficient rendering method for multi-scale traffic facility 3D models according to claim 1, characterized in that, S3 specifically includes: S31. Receive the multi-resolution LOD model set, collect the number of drawing calls per frame, average number of vertices, average number of edges, and shader overhead indicators in real time through GPU performance counters, establish a rendering cost model, and output the quantized rendering cost value per frame. The rendering cost model is as follows: ； in, For the first Total frame rendering cost , , The calibration parameters correspond to the overhead of drawing calls, vertex processing, and shader computation, respectively. To determine the number of calls to the drawing; , These represent the average number of vertices and the average number of edges in the model, respectively. Calculate the overhead metrics for the shaders; S32. Based on the quantized rendering cost value per frame, set a high load threshold. and low load threshold Perform dual-threshold hysteresis adaptive control: when the rendering cost exceeds the high load threshold, increase the simplification rate of the mid-range area; when the rendering cost is lower than the low load threshold, activate the reverse refinement of the distant proxy model and output the dynamically adjusted model simplification parameters. The adjustment formula for increasing the simplification rate of the mid-range scene when the rendering cost exceeds the high load threshold is as follows: ； in, To simplify the dynamic model of the mid-range scenic area, Let k = 0.05 be the initial simplification rate, and k = 0.05 be the adjustment coefficient. ; The update formula for activating the reverse refinement of the distant proxy model when the rendering cost is lower than the low load threshold is: ； in, The total number of vertices in the distant Bézier proxy model after adjusting for low load. To ensure that the total number of vertices in the prospective proxy model before this refinement operation is ≤15%, the rendering load fluctuation range is guaranteed. S33. Receive the dynamic adjustment parameters output by S32 and set abnormal state monitoring: if the rendering cost of 5 consecutive frames exceeds 1.2 times the high threshold, then forcibly reduce the rendering resolution to 75% of the original resolution until the load returns to normal, and output the final load-balanced rendering parameter set.

3. The efficient rendering method for multi-scale traffic facility 3D models according to claim 2, characterized in that, S4 specifically includes: S41. Receive the rendering parameter set and the adjusted multi-resolution LOD model, use block-based streaming loading technology to divide the scene data into 32m×32m spatial blocks, load high-priority area data 2-3 frames in advance by predicting the line of sight direction, and output the scene data stream after streaming loading. S42. Receive the scene data stream and perform differentiated rendering processing on dynamic and static elements in the scene: use GPU instantiation rendering for moving vehicles, enable batch merging drawing for static facilities such as traffic signs, markings, and traffic lights, and output the rendering instruction stream after API call optimization. S43. Based on the rendering instruction stream, integrate the rendering parameters to perform the final rasterization rendering and output a visually continuous 3D rendering result with a visual similarity SSIM value of not less than 0.

9.

4. A high-efficiency rendering system for multi-scale traffic facility 3D models, implemented based on the methods described in claims 1-3, characterized in that, include The dynamic spatial partitioning module is used to collect camera pose parameters and motion state data in real time. Based on the view frustum motion perception model, the rendering scene is divided into three spatial regions: near-field, middle-field, and far-field, and the scene model set after the three-level partitioning is output. The semantic feature-preserving LOD simplification module is used to perform semantic feature-sensitive progressive geometric simplification on each region model based on the three-level partitioned scene model set, and generate a multi-resolution LOD model set with semantic protection constraints. The elastic load closed-loop control module is used to monitor the GPU rendering pipeline status in real time and calculate the rendering cost per frame based on the multi-resolution LOD model set, dynamically adjust the model simplification rate of each region based on the dual threshold feedback mechanism, and output the rendering parameter set after load balancing. The scene streaming differential rendering module is used to receive the rendering parameter set and the adjusted multi-resolution LOD model, perform block streaming loading and differential rendering of dynamic and static objects, and output visually continuous 3D rendering results.

Images