A three-dimensional image rendering system for urban reality creation

Abstract

The present application relates to the technical field of image data processing, and discloses a three-dimensional image drawing system for urban real scene establishment, comprising a geometric analysis module, a sampling feature evaluation module, a texture scheduling module and a rendering pipeline, visible polygon patches are extracted from a viewpoint pose parameter, and the projection gradient of the polygon patches in the screen space is calculated to determine a sampling state, when the projection gradient exceeds a preset grazing angle threshold, the sampling state is marked as a texture sampling bypass mode, and the rendering pipeline colors corresponding pixels with low-frequency baked color values pre-stored in the geometric topology data layer in the sampling state, the present application drives the decoupling scheduling of the geometric layer and the texture layer through the projection gradient, eliminates the redundant bandwidth overhead of high-frequency texture sampling under extreme grazing angles, effectively suppresses the distortion interference of side elevations, balances the visual fidelity and system load in real scene roaming, and guarantees the stability of large-scale drawing.

Description

Technical Field

[0001] This invention relates to a three-dimensional image rendering system for urban real-world scene reconstruction, belonging to the field of image data processing technology. Background Technology

[0002] Current mainstream technical solutions typically employ a slice-based detailed hierarchical scheduling architecture to manage massive spatial data. This architecture organizes spatial data into a pyramid index tree based on the viewpoint pose, enabling synchronous mapping and rendering of geometric models and surface texture data. In conventional view frustum clipping and depth buffer testing processes, the system primarily determines the necessity of data loading and processing based on the visibility of geometric boundaries. However, during dynamic roaming of ultra-large-scale urban scenes, the physical characteristics of perspective projection cause severe geometric deformation of the building facades on both sides of the street depth. Although such building primitives can be determined to be visible through depth testing, their surface normals and view vectors are often in an extreme swoop angle state. Due to the limitations of screen space projection rules, the actual pixel density occupied by such patches in the display area is low, causing the rendering pipeline to continuously call high-frequency texture data for pixel clusters with low visual feature contribution. This neglect of the necessity of sampling results in hidden memory bandwidth waste and memory resource turnover pressure when the system processes facade details.

[0003] To address this technical bottleneck, industry attempts have attempted compensation through methods such as encrypted geometric detail layering or increased anisotropic filtering magnification. Analysis reveals that these linear improvement methods only partially alleviate image aliasing and cannot fundamentally resolve the mismatch between texture addressing frequency and projected pixel density. Encrypted layering introduces an exponentially increasing number of rendering batch calls, while increasing the anisotropic filtering magnification significantly increases the computational load on the shader, and neither effectively blocks the flow of redundant texture data in areas of sample degradation. Traditional spatial layering scheduling frameworks suffer from limitations in their underlying physical structure, and current software control methods in rendering pipelines also have shortcomings. For example, Chinese invention patent CN117541744B discloses a rendering method and apparatus for city-level real-scene 3D images, which, during the culling stage, relies on visible triangles... The surface is mapped to the pixel size in screen space. Different areas of surface are marked separately for hardware or software rasterization. The aim is to optimize the rasterization computation load by distinguishing the absolute pixel area of ​​the surface. However, the judgment benchmark is only at the macroscopic scalar level of two-dimensional projected area, which is completely detached from the normal gradient changes generated by three-dimensional spatial perspective projection. In actual large-scale urban roaming scenarios, the side facades of buildings are presented as narrow strips due to extreme perspective compression. The overall pixel area is still easy to exceed the preset area threshold, which continues to trigger full high-frequency texture addressing. The judgment boundary is established under the premise of ideal proportional scaling, which is fundamentally misaligned with the objective physical constraints of severe unidirectional distortion generated by real grazing angle projection. This causes the system to be unable to accurately intercept redundant memory access instructions in the high-frequency sampling deterioration interval, fails to substantially reduce the consumption of video memory bandwidth, and cannot suppress aliasing from the source.

[0004] Therefore, how to accurately determine the necessity of texture sampling by utilizing the physical laws of spatial projection transformation, and dynamically block the redundant flow of high-frequency texture data according to the projection gradient, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A three-dimensional image rendering system for urban real-scene reconstruction, comprising: The geometric analysis module is used to obtain viewpoint pose parameters, which include the three-dimensional position coordinates of the viewpoint in the world coordinate system and the rotation quaternion representing the viewing direction. Based on the viewpoint pose parameters, the module performs intersection retrieval within a spatial index tree built using a hierarchical bounding box structure to extract visible polygon patches, and then constructs a geometric topology data layer containing vertex topology data and vertex color information. The sampling feature evaluation module is used to establish a linear mapping relationship between texture mapping coordinates and screen space coordinates for visible polygon patches, calculate the first derivative of texture mapping coordinates with respect to screen space coordinates in the horizontal and vertical directions to determine the projection gradient, and determine the sampling state of the corresponding pixel based on the deviation of the projection gradient from the preset swoop angle threshold. The texture scheduling module is used to maintain a high-frequency texture detail layer in memory space that is independent of the geometric topology data layer, and to establish a texture data loading index for the corresponding visible polygon patches based on the sampling state determined by the sampling feature evaluation module. The rendering pipeline receives the geometric topology data layer and retrieves the high-frequency texture detail layer from the texture scheduling module based on the texture data loading index. The sampling feature evaluation module marks the sampling state as texture sampling bypass mode when it determines that the projection gradient represented by the first derivative exceeds a preset swoop angle threshold. When the sampling state is in texture sampling bypass mode, the rendering pipeline blocks the sampling addressing path for the high-frequency texture detail layer and retrieves the pre-stored low-frequency baked color values ​​from the vertex color information within the geometric topology data layer to perform diffuse shading and filling on the pixels.

[0006] Preferably, when calculating the projection gradient, the sampling feature evaluation module obtains the normal vector of the visible polygon patch and the viewing vector of the current observation space, and calculates the absolute value of the dot product of the normal vector and the viewing vector; the sampling feature evaluation module also determines the pixel projection compression rate of the pixel in the screen space based on the absolute value of the dot product, and lowers the loading priority corresponding to the texture data loading index when the pixel projection compression rate is greater than the preset compression threshold.

[0007] Preferably, the system further includes: a memory consistency constraint module, used to maintain the cache hit rate sequence within the memory paging area; the memory consistency constraint module combines the movement vector in the viewpoint pose parameters to evaluate the residency value of the memory paging area, and when the resource occupancy rate of the memory paging area exceeds 85%, it releases cache blocks located in the opposite direction of the movement vector and with a loading priority lower than a preset level.

[0008] Preferably, the geometry parsing module includes: a view frustum clipping submodule, used to determine the view frustum bounding box of the viewpoint in three-dimensional space; and a topology extraction submodule, used to traverse the spatial index tree to obtain nodes that intersect with the view frustum bounding box, and to load the geometric index data within the nodes into the geometric topology data layer.

[0009] Preferably, the rendering pipeline includes: a depth pretest submodule for occlusion culling of the geometric topology data layer; and a material mask verification submodule for extracting the transparency channel parameters in the high-frequency texture detail layer and correcting the depth buffer value of the depth pretest submodule based on the transparency channel parameters to preserve the geometric features of the glass curtain wall and the hollow structure.

[0010] Preferably, the texture scheduling module includes an anisotropic sampler, used to determine the texture filtering ratio based on the projection gradient when the sampling state is in non-texture sampling bypass mode, so as to suppress sampling distortion of the building side facade under the grazing angle projection.

[0011] Preferably, the low-frequency baked color values ​​pre-stored in the geometric topology data layer are obtained by pre-rendering the original real-world 3D model through the geometric analysis module; when the high-frequency texture detail layer is missing, the low-frequency baked color values ​​provide the average diffuse reflection color characteristics of the corresponding visible polygon facets.

[0012] Preferably, the system further includes: a dynamic threshold adjuster, used to monitor the frame rate stability parameters of the rendering pipeline and dynamically adjust the preset swoop threshold in the sampling feature evaluation module according to the frame rate stability parameters; when the frame rate stability parameters are lower than 60fps, the dynamic threshold adjuster reduces the preset swoop threshold to expand the effective range of the texture sampling bypass mode.

[0013] Preferably, the system is deployed in a distributed rendering cluster; the geometry analysis module and the texture scheduling module are deployed on different computing nodes; the computing nodes synchronize texture data loading indexes through a high-speed bus.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In 3D image rendering, by introducing a projection gradient interception mechanism based on the dot product of normal vector and view vector, the system can identify extreme grazing angle faces at the geometric stage and actively block high-frequency texture addressing and memory access requests for such faces. This leap from spatial visibility to sampling necessity removes the traditional logic constraint of rendering as visible. Without affecting the expression of visual features, it reduces the invalid video memory bandwidth occupied by perspective projection compression and improves the effective throughput performance of the rendering pipeline.

[0015] 2. An asynchronous decoupled architecture between the geometric topology contour layer and the independent high-frequency texture detail layer, combined with parallel comparison of low-resolution layered depth buffers, constructs a data flow model. The system pre-parses visible polygonal patches based on the current viewpoint pose and generates texture injection instructions only for valid pixel clusters that pass the projection gradient test. This cross-dimensional collaborative scheduling mechanism eliminates the strong binding constraint between geometric loading and texture sampling, ensuring that the rendering bus load is always kept within a safe threshold during rapid viewpoint movement, avoiding screen tearing caused by explosive data surges.

[0016] 3. Based on the projection gradient, the underlying addressing blocking logic smoothly backs up fragments in the high-frequency sampling deterioration region to the low-frequency primary color of the geometry layer for filling. This removes noise sources that cause aliasing interference from the data source. This mechanism can effectively suppress moiré patterns generated by distant building facades under swoop projection without relying on additional post-processing anti-aliasing algorithms, achieving a dynamic balance between visual fidelity and computational overhead, and ensuring the visual smoothness of image edges during ultra-large-scale real-scene roaming. Attached Figure Description

[0017] Figure 1A data processing flowchart of the 3D image rendering system for urban real-scene creation in this invention; Figure 2 The diagram shows the architecture and hardware interaction of the 3D image rendering system for urban real-scene creation in this invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A 3D image rendering system for urban real-world scene reconstruction includes: The geometric analysis module is used to obtain viewpoint pose parameters, which include the three-dimensional position coordinates of the viewpoint in the world coordinate system and the rotation quaternion representing the viewing direction. Based on the viewpoint pose parameters, the module performs intersection retrieval within a spatial index tree built using a hierarchical bounding box structure to extract visible polygon patches, and then constructs a geometric topology data layer containing vertex topology data and vertex color information. The sampling feature evaluation module is used to establish a linear mapping relationship between texture mapping coordinates and screen space coordinates for visible polygon patches, calculate the first derivative of texture mapping coordinates with respect to screen space coordinates in the horizontal and vertical directions to determine the projection gradient, and determine the sampling state of the corresponding pixel based on the deviation of the projection gradient from the preset swoop angle threshold. The texture scheduling module is used to maintain a high-frequency texture detail layer in memory space that is independent of the geometric topology data layer, and to establish a texture data loading index for the corresponding visible polygon patches based on the sampling state determined by the sampling feature evaluation module. The rendering pipeline receives the geometric topology data layer and retrieves the high-frequency texture detail layer from the texture scheduling module based on the texture data loading index. The sampling feature evaluation module marks the sampling state as texture sampling bypass mode when it determines that the projection gradient represented by the first derivative exceeds a preset swoop angle threshold. When the sampling state is in texture sampling bypass mode, the rendering pipeline blocks the sampling addressing path for the high-frequency texture detail layer and retrieves the pre-stored low-frequency baked color values ​​from the vertex color information within the geometric topology data layer to perform diffuse shading and filling on the pixels.

[0021] Preferably, when calculating the projection gradient, the sampling feature evaluation module obtains the normal vector of the visible polygon patch and the viewing vector of the current observation space, and calculates the absolute value of the dot product of the normal vector and the viewing vector; the sampling feature evaluation module also determines the pixel projection compression rate of the pixel in the screen space based on the absolute value of the dot product, and lowers the loading priority corresponding to the texture data loading index when the pixel projection compression rate is greater than the preset compression threshold.

[0022] Preferably, the system further includes: a memory consistency constraint module, used to maintain the cache hit rate sequence within the memory paging area; the memory consistency constraint module combines the movement vector in the viewpoint pose parameters to evaluate the residency value of the memory paging area, and when the resource occupancy rate of the memory paging area exceeds 85%, it releases cache blocks located in the opposite direction of the movement vector and with a loading priority lower than a preset level.

[0023] Preferably, the geometry parsing module includes: a view frustum clipping submodule, used to determine the view frustum bounding box of the viewpoint in three-dimensional space; and a topology extraction submodule, used to traverse the spatial index tree to obtain nodes that intersect with the view frustum bounding box, and to load the geometric index data within the nodes into the geometric topology data layer.

[0024] Preferably, the rendering pipeline includes: a depth pretest submodule for occlusion culling of the geometric topology data layer; and a material mask verification submodule for extracting the transparency channel parameters in the high-frequency texture detail layer and correcting the depth buffer value of the depth pretest submodule based on the transparency channel parameters to preserve the geometric features of the glass curtain wall and the hollow structure.

[0025] Preferably, the memory consistency constraint module follows the following quantization rules when calculating the residency value weight of cache blocks: Where W is the dwell value weight; R is the resolution level corresponding to the high-frequency texture detail layer; D is the Euclidean distance between the viewpoint pose parameter and the center point of the visible polygon patch; and θ is the angle between the surface normal of the visible polygon patch and the viewing direction, calculated by the sampling feature evaluation module.

[0026] Preferably, the texture scheduling module includes an anisotropic sampler, used to determine the texture filtering ratio based on the projection gradient when the sampling state is in non-texture sampling bypass mode, so as to suppress sampling distortion of the building side facade under the grazing angle projection.

[0027] Preferably, the low-frequency baked color values ​​pre-stored in the geometric topology data layer are obtained by pre-rendering the original real-world 3D model through the geometric analysis module; when the high-frequency texture detail layer is missing, the low-frequency baked color values ​​provide the average diffuse reflection color characteristics of the corresponding visible polygon facets.

[0028] Preferably, the system further includes: a dynamic threshold adjuster, used to monitor the frame rate stability parameters of the rendering pipeline and dynamically adjust the preset swoop threshold in the sampling feature evaluation module according to the frame rate stability parameters; when the frame rate stability parameters are lower than 60fps, the dynamic threshold adjuster reduces the preset swoop threshold to expand the effective range of the texture sampling bypass mode.

[0029] Preferably, the system is deployed in a distributed rendering cluster; the geometry analysis module and the texture scheduling module are deployed on different computing nodes; the computing nodes synchronize texture data loading indexes through a high-speed bus.

[0030] Example 1: In a large-scale urban digital twin system's real-world roaming scenario, the viewpoint moves deeper along a street of densely built-up buildings. The side facades of the buildings on both sides of the street have no foreground occlusion in physical space. According to the physical laws of perspective projection, the surface normal of the side facade forms a swoop angle with the viewing direction. This geometric relationship leads to a reduction in the actual projected pixel area in screen space. Traditional synchronous loading and rendering architectures trigger full-process shading under unobstructed spatial conditions. The rendering pipeline continuously initiates high-frequency anisotropic memory access requests for texture details for swoop angle pixel clusters with low projected pixel density. This redundant data memory access continuously occupies bus bandwidth and increases the video memory load. Due to insufficient pixel sampling rate, high-frequency signal aliasing of moiré patterns occurs. The geometric analysis module obtains the viewpoint pose parameters, which include the three-dimensional position coordinates of the viewpoint in the world coordinate system and the rotation quaternion representing the viewing direction. Based on the viewpoint pose parameters, the geometric analysis module performs intersection search within the spatial index tree established using a hierarchical bounding box structure to extract visible polygon patches. The geometric analysis module constructs a geometric topology data layer containing vertex topology data and vertex color information. This geometric topology data layer pre-stores low-frequency baked color values, which are obtained from the system's pre-rendered original real-world 3D model.

[0031] The sampling feature evaluation module establishes a linear mapping relationship between the texture mapping coordinates of visible polygonal patches and their screen space coordinates. It obtains the normal vector of the visible polygonal patch and the view vector in the current viewing space, and calculates the absolute value of their dot product. Based on this dot product absolute value, the module calculates the first derivatives of the texture mapping coordinates in the horizontal and vertical directions relative to the screen space coordinates, thereby determining the projection gradient. Based on the fundamental physical mapping principle of light projection in spatial geometry, the calculation of the above first derivative depends on the sampling feature evaluation module extracting the model-view projection matrix in the current viewing space and constructing a local Jacobian operator by combining the dot product absolute value parameter. This module calculates the partial derivative values ​​of the texture coordinates relative to the screen pixel space by performing Jacobian matrix differentiation operations, and outputs the combined norm matrix as a quantized projection gradient, which is then processed in the underlying shader. In this method, the calculation process is specifically set as a 2x2 pixel sliding sampling window. By comparing the texture coordinate differences of adjacent pixels in the horizontal and vertical axes within the window, and using the center difference algorithm to obtain the coordinate change rate under unit pixel displacement, the preset swoop angle threshold is set to 4.5, whose physical dimension corresponds to the number of texture units contained in a single pixel span. When the absolute value of the calculated coordinate change rate reaches 4.5 or more, it is determined that the texture sampling signal is aliased, and the bypass mechanism is activated. The sampling feature evaluation module determines the sampling state of the corresponding pixel based on the deviation of the projection gradient from the preset swoop angle threshold. When it is determined that the projection gradient represented by the first derivative exceeds the preset swoop angle threshold, the sampling feature evaluation module marks the sampling state as texture sampling bypass mode. This quantization deviation feature indicates that the pixel is in the high-frequency sampling restricted range, and the sampling feature evaluation module outputs the underlying addressing control command accordingly.

[0032] The texture scheduling module maintains a high-frequency texture detail layer independent of the geometry topology data layer within memory space. Based on the determined sampling state, the texture scheduling module establishes a texture data loading index for the corresponding visible polygon facet. The rendering pipeline receives the geometry topology data layer. When the sampling state is in texture sampling bypass mode, the rendering pipeline blocks the sampling addressing path for the high-frequency texture detail layer through hardware logic. In the programmable pipeline stage of the graphics processor, the rendering pipeline sends control instructions to the fragment shader front end to reset the texture read level bias parameter of the corresponding pixel to the underlying hardware constant. This overwrite operation skips the register activation cycle of the corresponding texture mapping unit and replaces the physical read stream of external video memory with a determined internal pipeline state flag. The rendering pipeline retrieves the pre-stored low-frequency baked color value from the vertex color information in the geometry topology data layer to complete the diffuse shading fill of the pixel. The system's built-in dynamic threshold adjuster continuously monitors the frame rate stability parameters of the rendering pipeline. When the frame rate stability parameter is below 60fps, the dynamic threshold regulator reduces the preset swipe angle threshold, expanding the effective range of the texture sampling bypass mode. The sampling step size of the dynamic threshold regulator is set to 16.6ms, that is, the load data is updated once per frame. When the system detects that the frame rate is below 60fps for 5 consecutive frames and is in the action trigger zone below 58fps, the regulator reduces the preset swipe angle threshold in a fixed increment of 0.1 until the frame rate recovers and stabilizes in the recovery zone above 62fps. If the frame rate exceeds 65fps due to scene simplification and lasts for more than 1 second, the regulator slowly increases the threshold in a step of 0.05 to balance image detail and throughput performance. The system relies on the asynchronous scheduling logic of geometric topology data and independent texture detail layers to actively intercept high-frequency texture data addressing requests under specific viewpoints, eliminate high-frequency sampling actions that cause image signal aliasing, and maintain the stability of the rendering bus load in the real-world roaming state.

[0033] Example 2: In the verification phase of the 3D image rendering system built for urban real-world scenes, a graphics rendering test platform was established. This test platform is configured with a maximum video memory bandwidth of 256GB / s and is equipped with a graphics processor that supports programmable rendering pipelines. The test data uses a publicly available large-scale urban building complex real-world dataset, which includes dense geometric topology data of building facades and high-frequency texture detail layers with a resolution of 4K. To simulate the mechanical vibration interference generated during real-world vehicle roaming, a Gaussian random jitter signal with a signal-to-noise ratio of 20dB is actively superimposed on the input viewpoint pose parameters. This jitter acts on the rotation quaternion representing the viewing direction, causing high-frequency fluctuations in the absolute value of the dot product between the view vector and the polygon patch normal vector. The preset swoop angle threshold is determined... Based on a quantitative trade-off between texture feature fidelity and memory bandwidth utilization, when the pixel sampling rate corresponding to a polygonal patch in screen space is lower than the Nyquist frequency limit of the high-frequency texture detail layer, the preset swoop angle threshold is set to the lower limit, causing the rendering pipeline to trigger the texture sampling bypass mode in advance. The test platform initially calibrates the preset swoop angle threshold to 4.5 as the starting point for balancing anti-aliasing requirements and memory overhead. Three independent verification models are set in the test phase. The control group adopts the native synchronous loading and rendering architecture without bypass interception logic. The partially missing control group is equipped with a sampling feature evaluation module and a preset swoop angle threshold with a fixed value of 4.5, and the dynamic threshold adjuster is removed. The sample group of this invention deploys a complete dynamic scheduling technology solution.

[0034] As the input viewpoint moves along the street at a speed of 15 m / s, Gaussian random jitter signals interfere with the viewpoint pose parameters, causing continuous fluctuations in the original viewpoint pose parameters. The sampling feature evaluation module extracts the normal vector and the view vector of the polygonal facets, calculating that the absolute value of the dot product oscillates between 0.05 and 0.15. Based on this absolute value of the dot product, the sampling feature evaluation module calculates the horizontal and vertical first derivatives of the texture mapping coordinates relative to the screen space coordinates and outputs the projection gradient value. In the control group, the rendering pipeline continuously initiates full high-frequency texture addressing. The system monitoring indicated that the video memory bandwidth utilization rate reached the 240GB / s limit, and the image structure similarity index dropped to 0.72. The sample group of this invention monitored that the calculated projection gradient value was greater than 4.5. The sampling feature evaluation module output the sampling status of the texture sampling bypass mode. The rendering pipeline blocked the high-frequency texture addressing path according to the sampling status and retrieved the low-frequency baked color value in the geometric topology data layer to complete the pixel diffuse shading and filling. The test data showed that the video memory bandwidth utilization rate of the sample group of this invention decreased to 110GB / s, and the image structure similarity index increased to 0.88.

[0035] The preset swoop angle threshold of the sample group of this invention was independently adjusted to verify the boundary effect and nonlinear evolution law of the parameter setting. When the threshold was lowered to 2.0, the texture sampling bypass mode took effect prematurely, the rendering pipeline used low-frequency baked color values ​​to replace a large amount of effective texture data, the image structure similarity index decreased sharply to 0.65, and the image lost basic visual semantics. When the threshold was raised to 6.0, the bypass interception mechanism failed in the swoop angle region, the memory bandwidth utilization rebounded exponentially to 215GB / s, and aliasing noise degraded the image quality again. This test data confirms that 4.0 to 5.0 constitutes the optimal working window for balancing rendering throughput and visual fidelity. In the long-sequence continuous roaming test, as the frequency of the Gaussian random jitter signal increased, the rendering pipeline of the partial missing control group showed Due to the response lag caused by the fixed threshold determination, the frame rate dropped to 42fps. The dynamic threshold adjuster in the sample group of this invention detected that the frame rate stability parameter was approaching the critical lower limit of 60fps, and actively lowered the preset swoop angle threshold from 4.5 to 3.8. This adjustment increased the effective pixel ratio of texture sampling bypass mode from 35% to 58%. The final test data showed that the rendering frame rate of the sample group of this invention was stably maintained at 62fps, and the peak memory bandwidth usage was locked at 135GB / s. This quantitative data proved that the decoupling scheduling mechanism of geometric topology contour and independent texture detail layer combined with the dynamic threshold adjuster effectively suppressed the memory bandwidth consumption of high-frequency texture sampling under high-frequency vibration conditions of viewpoint, and maintained the throughput stability and output reliability of the 3D image rendering system.

[0036] Example 3: In a large-scale urban landscape roaming scenario, the viewpoint pose parameters undergo continuous coordinate shifts over time. The geometry analysis module continuously feeds a sequence of visible polygonal faces into the rendering pipeline, while the texture scheduling module imports the corresponding high-frequency texture detail layers from external storage devices to the graphics processor's memory paging area. As the roaming path progresses unidirectionally, historical cache blocks accumulated in the memory paging area gradually build up. When the memory resource utilization rate continues to rise and approaches the hardware capacity limit, the memory data eviction mechanism is activated. The sampling feature evaluation module extracts the normal vector of the visible polygonal face and the view vector of the current observation space, calculates the absolute value of their dot product, and uses this absolute value to calculate the visible polygonal face's position on the screen. The topological contraction feature within the space outputs a pixel projection compression rate. A preset compression threshold is set to 0.75. When the calculated pixel projection compression rate is greater than 0.75, the sampling feature evaluation module sends a state update command to the texture scheduling module, lowering the loading priority corresponding to the texture data loading index from the initial level 1 to level 3. The memory consistency constraint module periodically reads the resource status of the memory paging area, extracts the 3D position coordinates of adjacent rendering frames from the viewpoint pose parameters, calculates the position difference to construct a movement vector representing the current viewpoint movement trajectory, and for each cached high-frequency texture detail layer within the memory paging area, the memory consistency constraint module extracts the corresponding world coordinate system center point and calculates the position of this center point relative to the current viewpoint. The cosine of the angle between the direction vector and the movement vector is used to output the dwell value assessment result. Based on the general inverse depth buffer physics law in the graphics rendering pipeline, the 3D image rendering system uses a floating-point inverted image format to store spatial variables in the underlying data channel. The Euclidean distance parameter set by this architecture is presented as a strictly decreasing dimensionless value in the hardware computing module. Its value boundary is restricted to between 0 and 1, and the closer the spatial physical viewing distance is, the closer the value is to 1. In specific implementation, the system uses a pre-calibrated maximum visible radius of 1000m as a benchmark, subtracts the physical distance value from the current viewpoint to the center point of the patch from 1000m, and divides the difference by 1000m, thereby linearly mapping the physical length to 0.0 to 1. The system extracts the standardized distance value between 0 and 1. If the physical distance exceeds 1000m, the value is forcibly reduced to zero. The system then multiplies this standardized distance value with the resolution level value corresponding to the high-frequency texture detail layer and the cosine of the angle between the surface normal and the viewing direction. Finally, it obtains the dwell value weight in the range of 0 to 1. This weight directly determines the retention probability of the memory block when resources are scarce. This underlying data representation structure establishes the dimensionless dwell value weight calculated according to the limited rules, which forms an accurate inverse attenuation mapping with the actual physical viewing distance, maintaining the effectiveness of the multi-resolution resource elimination strategy. A negative cosine value indicates that the corresponding high-frequency texture detail layer is located in the physical region opposite to the viewpoint movement vector.

[0037] The hardware bus continuously monitors the video memory throughput data. When the resource utilization rate of the video memory paging area exceeds 85%, the video memory consistency constraint module initiates the cache release logic. The video memory consistency constraint module traverses the texture data loading index in the video memory paging area, locks the cache block target with a negative resident value assessment result and a loading priority reduced to level 3, and issues an address space release instruction to the graphics processor's memory controller. After confirming that the cache block in a specific physical address space has been cleared, the video memory consistency constraint module synchronously rewrites the spatial index status identifier of the corresponding polygon face, forcing the subsequent rendering sequence to roll back the drawing path for that face to the original geometry layer. The rendering pipeline continuously extracts the low-frequency baked primary color bound in the geometry layer to fill the display pixels, eliminating the black block breaks in the screen caused by asynchronous video memory page faults. The 3D image rendering system releases the historical texture cache that deviates from the roaming path and has a limited loading priority based on the spatial mapping relationship between the viewpoint movement vector and the pixel projection compression rate, maintaining the turnover rate of texture resources in the video memory paging area and ensuring the data throughput stability of the rendering pipeline.

[0038] When the system faces a situation where the viewpoint continuously traverses different large-scale urban blocks, causing frequent texture data page-scrambling, the view frustum clipping submodule within the geometry analysis module extracts the viewpoint's pose parameters in 3D space to construct a view frustum bounding box defined by intersecting planes. The topology extraction submodule traverses a spatial index tree built using a hierarchical bounding box structure and extracts visible polygonal faces within the intersection range based on the view frustum bounding boxes. The memory consistency constraint module intercepts texture addressing commands initiated by the texture scheduling module within the sampling period of the rendering pipeline's main clock and compares the target address pointed to by the texture addressing command with the current memory allocation. Within the page area, when the target address is determined to be within the resident memory address range, a scalar 1 is added to the hardware register data stack; otherwise, a scalar 0 is added. The memory consistency constraint module summarizes the set of newly added scalar states in the most recent 128 consecutive sampling periods and establishes it as the cache hit rate sequence. The memory consistency constraint module parses the cache hit rate sequence and calculates the arithmetic mean to quantify and output the current memory load ramp-up gradient. The system merges the memory load ramp-up gradient with the movement vector in the viewpoint pose parameters and inputs it into the optimization judgment logic to calculate the resident value evaluation results of each historical cache block in the memory page area.

[0039] Example 4: In the initial deployment of the urban real-scene 3D image rendering system, in order to adapt to the hardware processing limits of the graphics processor and establish a rendering benchmark, the system initiates an offline baseline calibration process before real-scene roaming. The calibration module reads the current hardware's memory bandwidth limit and loads a standard urban street test sequence. Under the preset viewpoint trajectory, the preset swoop angle threshold and preset compression threshold are incremented by a fixed step size gradient. At the same time, the hardware bus synchronously collects the real-time memory throughput, rendering frame rate, and image structure similarity index under the corresponding threshold combination. Based on this, the calibration module constructs a 3D surface representing the relationship between the threshold combination and rendering performance. By eliminating invalid test points that cause the memory throughput to exceed 85% of the memory bandwidth limit or the rendering frame rate to be lower than 60fps, the system locks the candidate parameter domain that meets the hardware capability limits and outputs the initial calibration data.

[0040] For discrete test points within the candidate parameter domain, the numerical optimization engine extracts the image structure similarity index synchronously acquired under each set of parameters, and calculates the partial derivative of the image structure similarity index with respect to the preset swoop angle threshold and the preset compression threshold. Based on the partial derivative value, the system identifies the nonlinear inflection point that causes a step decay, and writes the numerical coordinates of the front end of the inflection point into the initialization configuration file to set as the preset swoop angle threshold and the preset compression threshold. The rendering pipeline starts the asynchronous scheduling mechanism for visible polygon patches and high-frequency texture detail layers based on the calibrated hardware adaptation parameters. The 3D image rendering system maintains continuous data throughput without exceeding the physical limits of the video memory.

[0041] Example 5: When the system faces boundary roaming scenarios involving viewpoint dwell or in-situ rotation, the spatial coordinate displacement between adjacent rendering frames approaches 0, causing the memory consistency constraint module to generate a zero-denominator anomaly when calculating the cosine of the angle between the movement vector and the direction vector. To address this computational anomaly, the calibration module presets a global jitter tolerance scalar during the initial system deployment phase. The geometric analysis module continuously extracts the viewpoint pose parameters of adjacent rendering frames and calculates the Euclidean distance displacement of the spatial coordinates. The memory consistency constraint module compares this Euclidean distance displacement with the global jitter tolerance scalar. When it determines that the Euclidean distance displacement is less than the global jitter tolerance scalar, the memory consistency constraint module suspends the direction determination logic for calculating the cosine of the angle and simultaneously activates the time decay counters for each corresponding high-frequency texture cache block. These time decay counters accumulate frame by frame with the main clock cycle of the rendering pipeline, thereby constructing an objective quantitative indicator of the cache lifecycle in a static viewpoint state.

[0042] Under normal operating conditions, the sampling feature evaluation module extracts the texture space resolution of the polygon facet in the world coordinate system and the current screen viewport resolution. This module multiplies the absolute value of the dot product of the normal vector and the view vector of the polygon facet with the texture space resolution and divides the product by the screen viewport resolution to quantify the output pixel projection compression rate. During continuous system operation, when the resource occupancy rate of the memory paging area exceeds 85% and the Euclidean distance displacement is less than the global jitter tolerance scalar, the memory consistency constraint module retrieves historical cache blocks whose time decay counter value is greater than the preset lifespan and issues an address space release instruction to the memory controller of the graphics processor. The 3D image rendering system uses the above state transition mechanism to clear high-frequency texture detail layers that have left the viewport during the alternation of viewpoint translation and static dwell, maintaining the resource throughput turnover of the memory paging area.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A three-dimensional image mapping system for city reality creation, characterized by comprising: include: The geometric analysis module is used to obtain viewpoint pose parameters, which include the three-dimensional position coordinates of the viewpoint in the world coordinate system and the rotation quaternion representing the viewing direction. Based on the viewpoint pose parameters, the module performs intersection retrieval within a spatial index tree built using a hierarchical bounding box structure to extract visible polygon patches, and then constructs a geometric topology data layer containing vertex topology data and vertex color information. The sampling feature evaluation module is used to establish a linear mapping relationship between texture mapping coordinates and screen space coordinates for visible polygon patches, calculate the first derivative of texture mapping coordinates with respect to screen space coordinates in the horizontal and vertical directions to determine the projection gradient, and determine the sampling state of the corresponding pixel based on the deviation of the projection gradient from the preset swoop angle threshold. The texture scheduling module is used to maintain a high-frequency texture detail layer in memory space that is independent of the geometric topology data layer, and to establish a texture data loading index for the corresponding visible polygon patches based on the sampling state determined by the sampling feature evaluation module. The rendering pipeline is used to receive the geometric topology data layer and retrieve the high-frequency texture detail layer from the texture scheduling module according to the texture data loading index; among them, the sampling feature evaluation module marks the sampling state as texture sampling bypass mode when it determines that the projection gradient represented by the first derivative exceeds the preset swoop angle threshold. When the rendering pipeline is in texture sampling bypass mode, it blocks the sampling addressing path for high-frequency texture detail layers and retrieves the low-frequency baked color values ​​pre-stored in the vertex color information in the geometry topology data layer to perform diffuse shading and filling on the pixels.

2. The three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, When calculating the projection gradient, the sampling feature evaluation module obtains the normal vector of the visible polygon patch and the viewing vector of the current observation space, and calculates the absolute value of the dot product of the normal vector and the viewing vector. The sampling feature evaluation module also determines the pixel projection compression rate of the pixel in the screen space based on the absolute value of the dot product, and lowers the loading priority corresponding to the texture data loading index when the pixel projection compression rate is greater than the preset compression threshold.

3. The three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The system also includes: a memory consistency constraint module, which is used to maintain the cache hit rate sequence within the memory paging area; the memory consistency constraint module combines the movement vector in the viewpoint pose parameters to evaluate the residency value of the memory paging area, and when the resource utilization rate of the memory paging area exceeds 85%, it releases cache blocks located in the opposite direction of the movement vector and with a loading priority lower than the preset level.

4. The three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The geometry parsing module includes: a view frustum clipping submodule, used to determine the view frustum bounding box of the viewpoint in 3D space; and a topology extraction submodule, used to traverse the spatial index tree to obtain nodes that intersect with the view frustum bounding box, and to load the geometric index data within the nodes into the geometric topology data layer.

5. A three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The rendering pipeline includes: a depth pretesting submodule, used to perform occlusion culling on the geometric topology data layer; and a material mask verification submodule, used to extract the transparency channel parameters in the high-frequency texture detail layer and correct the depth buffer value of the depth pretesting submodule based on the transparency channel parameters in order to preserve the geometric features of the glass curtain wall and the hollow structure.

6. The three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The texture scheduling module includes an anisotropic sampler, which determines the texture filtering ratio based on the projection gradient when the sampling state is in non-texture sampling bypass mode, so as to suppress sampling aliasing of the building side facade under the grazing angle projection.

7. A three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The low-frequency baked color values ​​pre-stored in the geometric topology data layer are obtained by pre-rendering the original real-world 3D model through the geometric analysis module; when the high-frequency texture detail layer is missing, the low-frequency baked color values ​​provide the average diffuse color characteristics of the corresponding visible polygon facets.

8. A three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The system also includes a dynamic threshold adjuster, which monitors the frame rate stability parameters of the rendering pipeline and dynamically adjusts the preset swoop threshold in the sampling feature evaluation module according to the frame rate stability parameters. When the frame rate stability parameters are below 60fps, the dynamic threshold adjuster reduces the preset swoop threshold to expand the effective range of the texture sampling bypass mode.

9. A three-dimensional image rendering system for urban real-scene reconstruction according to claim 1, characterized in that, The system is deployed in a distributed rendering cluster; the geometry analysis module and the texture scheduling module are deployed on different computing nodes; the computing nodes synchronize texture data loading indexes through a high-speed bus.

Images