A cartographic method for real-time map rendering optimization

Abstract

This invention provides a cartographic method for optimizing real-time map rendering, belonging to the field of cartographic technology for real-time map rendering optimization. This invention pre-calculates the screen space projection area at the central processing unit (CPU) and repackages it into batches with varying loads, then asynchronously submits the batches to the graphics processor (GPU) via a drawing indirect instruction mechanism. It uses 3D Hilbert curves to map tile coordinates as linear indices, combining quadtree divide-and-conquer recursion with user motion prediction to drive tile prefetching. It reconstructs low-resolution tile textures using a bi-branch map texture super-resolution fusion model and adaptively adjusts the high-frequency gating weight inference scaling factor through a comprehensive quality adjustment value. It compresses the occlusion query overhead per frame through a bi-phase occlusion culling pipeline combined with temporal reprojection. This invention solves the technical problems of unbalanced geometric subdivision load, low tile scheduling cache hit rate, and difficulty in adaptively balancing texture super-resolution reconstruction quality and inference latency during real-time map rendering.

Description

Technical Field

[0001] This invention belongs to the field of cartographic technology for real-time map rendering optimization, and more specifically, relates to a cartographic method for real-time map rendering optimization. Background Technology

[0002] Real-time map rendering is a core technology of Geographic Information Systems (GIS) and navigation software. Traditional methods convert vector geometric data into screen pixels through a rasterization pipeline of the graphics processor (GPU) and organize map texture data in a tile pyramid structure. In the geometric subdivision stage, traditional methods divide geometric blocks into batches based on a fixed subdivision level, without considering the differences in the screen space projection area of ​​each block from the current viewpoint. This leads to a severe imbalance in the computational load on the GPU between different batches, resulting in significant idle time and queuing of shader processor units. In the tile scheduling stage, traditional methods typically employ a rectangular scanning strategy based on the viewport range. Spatially adjacent tiles are discretely distributed in memory, resulting in low cache hit rates and insufficient bandwidth utilization. In the texture quality management stage, traditional super-resolution methods lack a joint dynamic control mechanism for inference latency and reconstruction quality, and cannot adaptively adjust the model inference strategy within the frame time budget. In existing technologies, due to the lack of a collaborative optimization mechanism among the aforementioned rendering subsystems, real-time map rendering struggles to simultaneously guarantee frame rate stability and visual quality in complex viewpoint motion scenes. In other words, existing technologies suffer from technical problems such as unbalanced geometric subdivision load, low tile scheduling cache hit rate, and difficulty in adaptively balancing texture super-resolution reconstruction quality and inference latency during real-time map rendering. Summary of the Invention

[0003] In view of this, the present invention provides a cartographic method for optimizing real-time map rendering, which can solve the technical problems in the prior art such as unbalanced geometric subdivision load, low tile scheduling cache hit rate, and difficulty in adaptively balancing texture super-resolution reconstruction quality and inference latency during real-time map rendering.

[0004] This invention is implemented as follows: This invention provides a cartographic method for real-time map rendering optimization, comprising the following steps:

[0005] The screen space projection area of ​​each geometric block is pre-calculated at the central processing unit. After sorting by screen space projection area, the subdivision tasks are repackaged into equal-load subdivision batches and submitted to the graphics processor for execution asynchronously through the drawing indirect instruction mechanism.

[0006] Map tile coordinates within the viewport are mapped to Hilbert linear indices using 3D Hilbert curves. Priority queues are established based on the Hilbert linear indices. Quadtree divide-and-conquer recursive scheduling is performed with the viewport center as the root node. The next batch of Hilbert indices is predicted based on the user velocity vector and user acceleration, and prefetching is triggered. When network bandwidth is insufficient, it degenerates to a lower-level Hilbert recursive hierarchy.

[0007] A map texture super-resolution fusion model is used to reconstruct low-resolution tile textures to obtain super-resolution textures; a comprehensive quality adjustment value is calculated, and the high-frequency gated weight inference scaling coefficient of the map texture super-resolution fusion model is adjusted according to the range of the comprehensive quality adjustment value through a high-frequency gated weight adaptive adjustment function;

[0008] Based on the signed distance field value from the pixel to the level detail boundary, the texture of adjacent level detail layers is blended with transparency. The multi-level progressive texture bias of the two level detail layers is smoothly transitioned through a sliding interpolation function. Temporal anti-aliasing is combined with historical frame feedback to smooth residual flicker.

[0009] Kalman filtering is used to estimate the future frustum range and asynchronous input / output requests are triggered in advance. Super-resolution textures and streaming textures are backfilled in a hierarchical caching system, and each level of cache is managed using the least recently used Kth access strategy. High-frequency hotspot tiles are pinned to the first-level graphics processor memory cache.

[0010] The first phase uses the hierarchical occlusion map of the previous frame to perform conservative occlusion culling on building instances, resulting in a candidate instance set; the second phase uses a computational shader to perform precise occlusion query on the candidate instance set; combined with temporal reprojection to compensate for the visibility results of the previous frame with motion vectors, only instances whose visibility state has changed are retested.

[0011] Specifically, the screen space projection area refers to the pixel area occupied by the three-dimensional geometric object after it is projected onto the two-dimensional pixel plane of the screen, which is calculated frame by frame by the central processing unit before the subdivision task is submitted.

[0012] Specifically, the equal-load subdivision batch refers to regrouping the vector geometric subdivision tasks according to the pre-calculated screen space projection area, so that the computational load of the graphics processor in each batch is approximately equal.

[0013] Specifically, the drawing indirect instruction mechanism refers to pre-storing parameters of equal-load subdivision batches into the graphics processor buffer through the graphics application programming interface, which is then read and executed autonomously by the graphics processor, thereby achieving asynchronous decoupling between the central processing unit and the graphics processor.

[0014] Specifically, the Hilbert linear index refers to a one-dimensional integer index obtained by mapping the horizontal and vertical coordinates of a map tile to a scaling level triplet through three-dimensional Hilbert curve encoding.

[0015] Specifically, the user velocity vector and user acceleration are calculated frame-by-frame from the viewpoint motion trajectory and are used to drive the prefetch window to slide and predict the next batch of Hilbert sequence number intervals.

[0016] The map texture super-resolution fusion model specifically includes an explicit branch and an implicit branch; the explicit branch is a convolutional upsampling network containing residual convolutional blocks, which outputs a low-frequency feature map; the implicit branch is a multilayer perceptron using a sinusoidal activation function, which outputs a high-frequency feature map; the low-frequency feature map and the high-frequency feature map are dynamically fused through a frequency separation gating network to output a super-resolution texture.

[0017] Among them, the comprehensive quality adjustment value The calculation uses a weighted combination of the normalized values ​​of the structural similarity index, the normalized values ​​of the peak signal-to-noise ratio, and the normalized values ​​of the graphics processor inference latency. The weight coefficients are determined by performing a grid search on the validation set.

[0018] Specifically, the high-frequency gated weight adaptive adjustment function is based on the comprehensive quality adjustment value. Range of high-frequency gated weight inference scaling factor Segmented adjustments are made, and when the overall quality adjustment value is lower than the minimum quality threshold, the system switches to the low-frequency feature map separate inference mode.

[0019] Specifically, the signed distance field value refers to the signed nearest distance from a pixel to the boundary of the level detail, with positive values ​​taken inside the boundary and negative values ​​taken outside, and is calculated from the screen space distance field.

[0020] Specifically, the multi-level progressive texture bias refers to the texture sampling bias parameters configured for detail layers at different resolution levels, which are used to control the sampling frequency of anisotropic filtering.

[0021] The hierarchical caching system consists of four levels: graphics processor memory, system memory, non-volatile memory, high-speed interface solid-state drive, and network content delivery network. When a cache hit occurs at any level, the data is returned directly. When a cache miss occurs, the data is requested from the next level and filled in step by step.

[0022] The "least recently used Kth access strategy" specifically refers to using the timestamp of the Kth access to the data as the basis for cache eviction priority, so as to avoid occasional low-frequency access data polluting the cache.

[0023] The hierarchical occlusion map is specifically a depth pyramid structure on the graphics processor side, which is constructed by downsampling the depth buffer level by level, and is used to quickly and conservatively determine whether the bounding box of a building instance is occluded.

[0024] Among them, the comprehensive quality adjustment value The high-quality threshold ranges from 0.90 to 0.98, with an optimal value of 0.95; the medium-quality threshold ranges from 0.75 to 0.85, with an optimal value of 0.80; the low-quality threshold ranges from 0.55 to 0.65, with an optimal value of 0.60; and the high-frequency gating weight inference scaling factor... The incrementing step size ranges from 0.03 to 0.08, with an optimal value of 0.05; the decrementing step size ranges from 0.08 to 0.15, with an optimal value of 0.10; the transition zone width ranges from 32 pixels to 128 pixels, with the optimal value determined through a subjective visual evaluation experiment using an 8-pixel step size.

[0025] This invention addresses the technical challenges of uneven geometric subdivision load, low tile scheduling cache hit rate, and difficulty in adaptively balancing texture super-resolution reconstruction quality and inference latency during real-time map rendering by pre-calculating the screen space projection area at the central processing unit (CPU), employing 3D Hilbert curve mapping and quadtree divide-and-conquer recursive tile scheduling, and introducing a high-frequency gating weight adaptive adjustment mechanism driven by a comprehensive quality adjustment value. The screen space projection area quantifies the actual computational contribution of each geometric block to the current viewpoint, providing a physical basis for equal-load batch partitioning and eliminating the root cause of shader processor unit idleness. The spatial filling characteristics of the Hilbert curve ensure the continuity of spatially adjacent tiles in a one-dimensional sequence. Quadtree divide-and-conquer combined with user motion prediction ensures that the scheduling order closely matches the viewpoint movement, thereby improving cache hit rate and reducing bandwidth waste caused by invalid prefetching. The comprehensive quality adjustment value comprehensively considers three indicators: structural similarity index, peak signal-to-noise ratio, and inference latency. The high-frequency gating weight adaptive adjustment function dynamically switches the inference mode based on this value, ensuring that the model consistently outputs the optimal quality super-resolution texture within the frame time budget. In summary, this invention solves the technical problems mentioned in the background art, such as insufficient collaborative optimization of multiple subsystems in real-time map rendering and difficulty in simultaneously ensuring rendering frame rate and visual quality. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a diagram showing the adaptive switching relationship between the distribution of comprehensive quality adjustment values ​​and the high-frequency gating weight inference scaling coefficient. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shown is a flowchart of a real-time map rendering optimization cartographic method provided by the present invention. This method includes the following steps:

[0030] S01. Pre-calculate the screen space projection area of ​​each geometric block at the central processing unit, sort them according to the screen space projection area, repackage the subdivision tasks into equal-load subdivision batches, and submit them to the graphics processor for execution asynchronously through the drawing indirect instruction mechanism.

[0031] S02. Map the map tile coordinates within the viewport area to Hilbert linear numbers using a 3D Hilbert curve. Establish a priority queue based on the Hilbert linear numbers. Perform quadtree divide-and-conquer recursive scheduling with the viewport center as the root node. Predict the next batch of Hilbert number intervals based on the user velocity vector and user acceleration and trigger prefetching. When network bandwidth is insufficient, degrade to a lower-level Hilbert recursive hierarchy.

[0032] S03. Use a map texture super-resolution fusion model to perform super-resolution reconstruction of low-resolution tile textures to obtain super-resolution textures; calculate the comprehensive quality adjustment value, and adjust the high-frequency gated weight inference scaling coefficient of the map texture super-resolution fusion model according to the range of the comprehensive quality adjustment value through the high-frequency gated weight adaptive adjustment function.

[0033] S04. Based on the signed distance field value from the pixel to the boundary of the layer detail, perform transparency blending on the texture of adjacent layer detail layers, and use a sliding interpolation function to smoothly transition the multi-level progressive texture offset of the two layer detail layers, combined with time-based anti-aliasing historical frame feedback to smooth residual flicker.

[0034] S05. Use Kalman filtering to estimate the future frustum range and trigger asynchronous input / output requests in advance; fill the super-resolution texture and streaming texture back level by level according to the hierarchical caching system, and manage each level of cache with the least recently used Kth access strategy, pinning high-frequency hotspot tiles to the first level of graphics processor memory cache;

[0035] S06. The first phase uses the hierarchical occlusion map of the previous frame to perform conservative occlusion removal on building instances to obtain a candidate instance set; the second phase uses a computation shader to perform precise occlusion query on the candidate instance set; combined with temporal reprojection to compensate for the visibility results of the previous frame with motion vectors, only instances whose visibility state has changed are retested.

[0036] Among them, the screen space projection area refers to the pixel area occupied by the three-dimensional geometry after it is projected onto the two-dimensional pixel plane of the screen. It is used to measure the visual contribution of the geometric block to the current viewpoint and is calculated frame by frame by the central processing unit before the subdivision task is submitted. The equal load subdivision batch refers to regrouping the vector geometry subdivision task according to the pre-calculated screen space projection area, so that the amount of computation of the graphics processor in each batch is approximately equal, thereby eliminating the idle running and uneven queuing of shader processor units. The drawing indirect instruction mechanism refers to the pre-storing of the parameters of the equal load subdivision batch into the graphics processor buffer through the graphics application programming interface, which is then read and executed by the graphics processor, realizing the asynchronous decoupling of the central processing unit and the graphics processor.

[0037] Among them, the Hilbert linear index refers to the one-dimensional integer index obtained by mapping the horizontal and vertical coordinates of map tiles to the scaling level triplet through three-dimensional Hilbert curve encoding, so that spatially adjacent tiles are arranged continuously in memory to maximize cache hit rate; the user velocity vector and user acceleration are calculated frame by frame difference from the viewpoint motion trajectory, and are used to drive the prefetch window to slide and predict the next batch of Hilbert index intervals; the quadtree divide-and-conquer recursion refers to recursively dividing the viewport range into four sub-regions with the viewport center as the root node, and scheduling the tile batches in the sub-regions in sequence according to the continuous segments of the Hilbert linear index; The adaptive Hilbert curve divide-and-conquer tile scheduling algorithm achieves the following technical advantages: Based on the cache locality characteristic of the spatial filling curve, it compresses the three-dimensional coordinates of map tiles into a spatially continuous one-dimensional sequence, enabling the GPU and CPU caches to hit the same cache line in batches during reads, reducing cache miss penalties; the quadtree divide-and-conquer recursion combined with a prefetch window driven by user velocity vectors and user acceleration ensures that the scheduling order closely matches the view movement direction, reducing bandwidth waste caused by invalid prefetches; the mechanism of dynamically degenerating to a lower-level Hilbert recursion level guarantees the robustness of the system under network bandwidth constraints. The overall algorithm complexity is O(n log n). .

[0038] The specific structure of the map texture super-resolution fusion model is as follows: the input of the map texture super-resolution fusion model is a normalized coordinate triple, wherein the normalized coordinate triple is composed of horizontal coordinates, vertical coordinates, and scaling level, which are respectively normalized to [value missing]. The intervals are then concatenated; the normalized coordinate triples first pass through a Fourier feature location encoding layer, mapping to a 256-dimensional Fourier feature vector; the Fourier feature vectors are then fed into the explicit and implicit branches respectively; the explicit branch is a standard convolutional upsampling network containing four residual convolutional blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and a linear rectified activation function, outputting a low-frequency feature map; the implicit branch is a multilayer perceptron containing six fully connected layers, each with a width of 512, using a sinusoidal activation function, outputting a high-frequency feature map; the low-frequency and high-frequency feature maps are dynamically fused through a frequency separation gating network; the frequency separation gating network takes the local gradient magnitude at the corresponding position of the normalized coordinate triples as input and outputs pixel-wise high-frequency gating weights, the sum of which is 1; the high-frequency and low-frequency feature maps are multiplied by their respective gating weights and then added, then processed by a 1×1 convolution to output a super-resolution texture. The steps for establishing the training dataset for the map texture super-resolution fusion model specifically include: collecting high-resolution map tiles with a resolution of no less than 4096×4096 pixels, covering four types of land features: roads, buildings, green spaces, and water bodies; downsampling the high-resolution map tiles with a downsampling factor of uniform sampling between 2 and 8 to obtain low-resolution map tiles, forming paired training samples; applying random rotation, random cropping, and brightness perturbation data augmentation operations to the paired training samples; dividing the training set, validation set, and test set into an 8:1:1 ratio; the training steps for the map texture super-resolution fusion model specifically include: using the weighted sum of the pixel-wise mean squared error loss and the perceptual loss as the total loss function; determining the perceptual loss weight by performing a grid search on the validation set within the range of 0.01 to 0.1, and taking the weight value that is optimal when the peak signal-to-noise ratio and structural similarity index of the validation set are combined; and using an adaptive moment estimation optimizer with an initial learning rate of... The model parameters are decayed by 0.5 times every 50 training rounds; the batch size is 16; training stops when the validation set loss no longer decreases for 10 consecutive rounds; after verifying that the peak signal-to-noise ratio and structural similarity index are qualified on the test set, the model parameters are solidified for inference.

[0039] The technical effects of the map texture super-resolution fusion model are as follows: the implicit neural representation approximates the high-frequency spatial distribution of map texture with a continuous function, overcoming the loss of high-frequency details caused by discretization in traditional convolutional networks; Fourier feature location encoding expands the input feature space through multi-frequency trigonometric functions, enabling the multilayer perceptron to fit high-frequency components; frequency separation gating adaptively allocates the fusion weights of low-frequency feature maps and high-frequency feature maps based on the local gradient magnitude, suppressing high-frequency noise in flat color block areas and enhancing high-frequency details in edge areas; the dual-branch fusion architecture enables the model to restore high-frequency texture edges while maintaining the accuracy of low-frequency colors, supporting continuous super-resolution output with arbitrary scaling factors.

[0040] Among them, the comprehensive quality adjustment value The calculation formula is expressed as follows: ;in The average structural similarity index for the current batch of reasoning. The average peak signal-to-noise ratio of the current batch of inference. For the current frame's graphics processor inference latency, , , These are the benchmark structural similarity index, benchmark peak signal-to-noise ratio, and benchmark inference delay, respectively, as determined by the validation set. , , The weighting coefficients are determined by performing a grid search on the validation set for different weight combinations, selecting the weight combination that maximizes the correlation coefficient between the overall quality adjustment value and the actual perceived quality; the high-frequency gated weight adaptive adjustment function is based on the overall quality adjustment value. Range of high-frequency gated weight inference scaling factor Adjustments will be made when hour, Keep the current value unchanged; when hour, Increment by 0.05 steps; when hour, Decrease in increments of 0.10; when hour, Set the minimum value and switch to the low-frequency feature map independent inference mode; the above thresholds of 0.95, 0.80, and 0.60 are obtained in the following way: collect 500 sets of inference samples at different viewpoint speeds and zoom frequencies in the test scenario, perform piecewise linear regression analysis on the comprehensive quality adjustment value and user subjective rating, and take the segment points where the slope of the regression curve changes significantly as threshold candidates. After verification by 3 rounds of iterative experiments, the thresholds are determined.

[0041] Among them, the signed distance field value refers to the signed nearest distance from a pixel to the boundary of the level detail layer, with positive values ​​inside the boundary and negative values ​​outside, calculated from the screen space distance field, and used to drive the transparency blending weights; multi-level progressive texture bias refers to the texture sampling bias parameters configured for different resolution level detail layers, used to control the sampling frequency of anisotropic filtering and prevent abrupt changes in sampling direction at the boundary of the transition area between adjacent level detail layers; temporal anti-aliasing history frame feedback refers to mixing the rendering results of several previous frames with the current frame after motion vector compensation, suppressing residual flicker in a temporal accumulation manner; the technical effect of the multi-level progressive texture bias blending gradient algorithm based on the screen space distance field is: the signed distance field value drives the transparency blending weights. The method eliminates pixel-level aliasing caused by discontinuous texture frequencies in adjacent detail layers; the smooth transition of multi-level progressive texture bias by the sliding interpolation function ensures continuous anisotropic filtering within the transition zone, eliminating abrupt changes in sampling direction; the historical frame feedback of temporal anti-aliasing further suppresses residual flicker that cannot be eliminated within a single frame, improving visual continuity; the transition zone width setting is obtained by collecting subjective visual evaluation data at different screen resolutions and scaling speeds, using the disappearance of transition tearing and the absence of obvious blurring as dual standards, conducting experiments with a step size of 8 pixels in the range of 32 pixels to 128 pixels, recording the average subjective score at each step size, and taking the step size value with the best score and acceptable computational cost as the final transition zone width.

[0042] The hierarchical caching system consists of four levels: graphics processor memory, system memory, non-volatile memory (SSD) with high-speed interface, and network content delivery network. When a cache hit occurs at any level, the data is returned directly; when a cache miss occurs, the data is requested from the next level and filled in step by step. The Least Recently Used (LRU) Kth Access strategy uses the timestamp of the Kth access to the data as the cache eviction priority to avoid occasional low-frequency accesses polluting the cache. The Kalman filter takes the user velocity vector and user acceleration obtained by frame-by-frame difference of the viewpoint motion trajectory as input, and performs prediction and update steps alternately to output an estimate of the future view frustum range, which is used to trigger asynchronous input / output requests in advance. The technical effects of the proactive page prefetching system are as follows: Kalman filtering dynamically models the user's perspective motion state, ensuring that asynchronous input / output requests are completed before page missing occurs, minimizing the impact of page missing on rendering frame time; the hierarchical backfilling mechanism of the tiered caching system keeps high-frequency hotspot tiles residing in the high-speed cache layer for a long time, reducing the number of cross-layer data transfers and lowering the probability of input / output queue depth saturation; the page missing rate threshold is determined by collecting page missing rate data at different perspective speeds, with the rendering frame time variance not exceeding a preset value, and using the point where the correlation coefficient between page missing rate and frame time variance shows a significant jump in multiple experiments as the threshold.

[0043] Among them, the hierarchical occlusion map is a depth pyramid structure on the graphics processor side, constructed by downsampling the depth buffer level by level, used to quickly and conservatively determine whether the bounding box of a building instance is occluded; the candidate instance set refers to the set of building instances that have not yet been confirmed as invisible after the first phase of conservative occlusion culling, which serves as the input for the second phase of precise occlusion query; temporal reprojection refers to mapping the visibility test results of the previous frame to the coordinate system of the current frame through motion vectors, as the initial visibility estimate for occlusion culling in the current frame; the technical effect of the two-phase culling pipeline is: first Conservative occlusion removal utilizes existing depth information to quickly compress the size of the candidate instance set, reducing the computational cost of precise occlusion queries in the second phase. Temporal reprojection leverages the temporal continuity of inter-frame visibility, ensuring that each frame only needs to retest instances whose visibility state has changed, keeping the per-frame overhead of occlusion removal at a stable low level. The per-frame removal overhead threshold is determined by sampling the actual frame time of the two-phase removal pipeline under different building density scenarios, with the constraint that the proportion of the occlusion removal stage in the frame time budget does not exceed 5%, and taking the 95th percentile of the time consumption in multiple experiments as the threshold.

[0044] Optionally, the present invention also provides a computer-based method for forming a real-time map rendering optimization cartographic system, wherein the computer is provided with a readable storage medium storing program instructions, which are used to execute the above-described method when the computer is run.

[0045] The specific implementation of step S01 is as follows: At the beginning of each frame, the CPU performs perspective projection calculations on all geometric blocks within the viewport range, projecting the vertices of the 3D bounding box of each geometric block onto the 2D pixel plane of the screen. The area of ​​the axis-aligned bounding rectangle of the projected polygon is taken as the estimated screen space projection area of ​​that geometric block. The screen space projection area reflects the actual contribution of each geometric block to the graphics processor's shading calculations at the current viewpoint. The larger the area, the more triangular fragments are generated after subdivision, and the greater the computational load required. After sorting all subdivision tasks in descending order based on the screen space projection area, a greedy bin packing algorithm is used to allocate the subdivision tasks one by one to the batch with the smallest current cumulative computational load until all tasks are allocated, thereby making the estimated graphics processor computational load of each batch approximately equal, forming an equal-load subdivision batch. The parameters of the equal-load subdivision batch, including the instance index, subdivision level, and vertex buffer offset, are pre-written into the graphics processor buffer. Through the drawing indirect instruction mechanism, the graphics processor reads and schedules the execution autonomously, realizing asynchronous decoupling between the CPU and the graphics processor, and eliminating pipeline pauses where the CPU waits for the graphics processor to complete frame by frame.

[0046] The specific implementation of step S02 is as follows: For each map tile within the viewport range, its horizontal coordinates, vertical coordinates, and scaling level triplet are input into a 3D Hilbert curve encoding function to obtain the corresponding Hilbert linear index. The spatial filling property of the 3D Hilbert curve ensures that adjacent tiles in the 3D coordinate space are also adjacent in terms of Hilbert linear index, thus ensuring that tiles stored continuously in memory are also spatially adjacent, maximizing the hit rate of the graphics processor and central processing unit cache. After all tiles are inserted into the priority queue according to their Hilbert linear index, the viewport range is recursively divided into four equal parts using the Hilbert linear index corresponding to the viewport center as the root node. Continuous Hilbert index segments within each sub-region are scheduled sequentially, forming a quadtree divide-and-conquer recursive scheduling structure. The user velocity vector and user acceleration are calculated frame-by-frame from the viewpoint motion trajectory. The velocity vector and acceleration of the current frame are substituted into the uniform acceleration extrapolation model to predict the Hilbert linear index interval of the next batch of viewport centers. A tile prefetching request is initiated to the network layer in advance, ensuring that the tile data is loaded before the viewport arrives. When the network bandwidth is detected to be lower than the preset bandwidth threshold, the scheduler automatically degrades to a lower Hilbert recursion level and schedules only the tile batches with lower resolution to ensure the robustness of the rendering frame rate.

[0047] The specific implementation of step S03 is as follows: The input to the map texture super-resolution fusion model is a normalized coordinate triplet, which is mapped to a 256-dimensional Fourier feature vector after passing through a Fourier feature location encoding layer, and then simultaneously fed into the explicit branch and the implicit branch. The explicit branch consists of four residual convolutional blocks, each containing two 3×3 convolutional layers, a batch normalization layer, and a linear rectified activation function, responsible for extracting low-frequency color and structural features and outputting a low-frequency feature map. The implicit branch consists of six fully connected layers, each with a width of 512, using a sinusoidal activation function, and approximating the high-frequency spatial distribution with a continuous function using implicit neural representation, outputting a high-frequency feature map. The frequency separation gating network takes the local gradient magnitude at the corresponding position of the normalized coordinate triplet as input and outputs pixel-wise high-frequency gating weights. The sum of the low-frequency gating weights and the high-frequency gating weights is constrained to 1. After weighted fusion of the two feature maps, a 1×1 convolution is performed to output the super-resolution texture. A comprehensive quality adjustment value is then applied. According to the formula Calculate frame by frame, when High-frequency gated weighted inference scaling factor Remain unchanged; when hour Increment by 0.05 increments to improve high-frequency detail quality; when hour Decrease in increments of 0.10 to reduce inference overhead; when Switch to low-frequency feature map separate inference mode to ensure that the frame time budget is not exceeded.

[0048] The specific implementation of step S04 is as follows: For each pixel on the screen, calculate its signed distance field value to the nearest level detail boundary, taking positive values ​​inside the boundary and negative values ​​outside. Based on the signed distance field value, linearly interpolate the transparency blending weights of adjacent level detail layer textures within a preset transition zone width using a sliding interpolation function, so that pixel-level aliasing gradually disappears rather than abruptly changing when switching level detail layers. Simultaneously, multi-level progressive texture bias is continuously interpolated within the transition zone based on the signed distance field value, controlling the sampling frequency of anisotropic filtering to smoothly transition and eliminate visual tearing caused by abrupt changes in sampling direction. Temporal anti-aliasing history frame feedback weights and blends the rendering results of previous frames with the current frame after motion vector compensation, suppressing residual flicker that cannot be completely eliminated within a single frame through temporal accumulation. The reference value for the transition zone width is 64 pixels, which can be adjusted from 32 pixels to 128 pixels depending on the screen resolution and scaling speed.

[0049] The specific implementation of step S05 is as follows: Kalman filtering uses the user velocity vector and user acceleration as input state variables, performs prediction and update steps in each frame, and outputs the optimal estimate of the future frustum range. The future frustum range estimation result is used to initiate asynchronous input / output requests to the storage layer in advance, ensuring that data transmission is completed before page misses occur. Super-resolution textures and streaming textures are backfilled in a four-level order: graphics processor memory, system memory, non-volatile memory high-speed interface solid-state drive, and network content delivery network. When a missing texture is found, a request is made to the next level. Each level of cache uses the least recently used Kth access strategy for eviction management, with the Kth access timestamp as the eviction priority, to avoid occasional low-frequency accesses polluting the high-speed cache layer. High-frequency hotspot tiles are identified through access frequency statistics and are forcibly pinned to the graphics processor memory cache, excluding them from the regular eviction process, ensuring zero-latency hits for high-frequency tiles.

[0050] The specific implementation of step S06 is as follows: At the beginning of each frame, the dual-phase culling pipeline first uses the depth pyramid constructed after the previous frame's rendering as a hierarchical occlusion map. A conservative occlusion test is performed on the axis-aligned bounding boxes of all building instances. Instances whose bounding boxes are completely located beyond the maximum depth of the corresponding level in the depth pyramid are determined to be invisible and directly culled. The remaining instances form a candidate instance set. Secondly, for each instance in the candidate instance set, a precise occlusion query is submitted in the compute shader. The depth of the instance's actual geometry is compared pixel-by-pixel with the depth buffer to obtain a precise visibility result. Temporal reprojection maps the visibility results of each instance in the previous frame to the current frame's coordinate system through motion vectors, serving as an initial visibility estimate. Occlusion tests are only re-executed for instances whose visibility state has changed, thereby controlling the actual computational load of occlusion culling in each frame at a stable low level. The frame time proportion of the occlusion culling stage is referenced to 5% of the total frame time budget.

[0051] It should be noted that the key technologies of this invention include: equal-load subdivision batch partitioning technology, adaptive Hilbert curve divide-and-conquer tile scheduling technology, and super-resolution adaptive inference technology driven by comprehensive quality adjustment value. The equal-load subdivision batch partitioning technology uses the screen space projection area as an objective metric, ensuring that batch partitioning directly corresponds to the actual computational load of the graphics processor, fundamentally eliminating shader processor unit idleness; compared to traditional fixed-level subdivision, its improved load balancing is guaranteed by information theory-level optimality. The adaptive Hilbert curve divide-and-conquer tile scheduling technology utilizes the cache locality of the space-filling curve to compress the three-dimensional tile coordinates into a spatially continuous one-dimensional sequence, making batch hits of hardware cache lines inevitable rather than accidental; quadtree divide-and-conquer combined with motion prediction further ensures that the scheduling order highly matches the viewpoint movement direction, reducing invalid prefetches. The super-resolution adaptive inference technology driven by comprehensive quality adjustment value incorporates perceptual quality and inference latency into the same scalar metric, achieving continuous adaptive switching of inference modes through a piecewise linear strategy, avoiding quality collapse caused by frame time fluctuations in fixed inference strategies. When the three key technologies work together, the frame time ratio of each of the three subsystems—geometry, texture, and scheduling—is precisely controlled, enabling the overall rendering pipeline to maintain a stable frame rate and visual quality in complex perspective motion scenes.

[0052] It should be noted that this invention also solves the following technical problem: In real-time map rendering, the overhead of occlusion culling for building instances increases linearly with scene density. Traditional frame-by-frame full occlusion query methods lead to uncontrolled frame time proportions in the occlusion culling stage in high-density urban scenes, squeezing the time budget of the geometry and texture rendering stages. This invention solves this problem through a two-phase culling pipeline and a temporal reprojection mechanism: The first-phase conservative occlusion culling utilizes the hierarchical structure of the depth pyramid to quickly compress the size of the candidate instance set with extremely low computational overhead, significantly reducing the input size of the second-phase precise occlusion query; Temporal reprojection utilizes the temporal continuity of inter-frame visibility to limit the number of instances that need to be retested in each frame to a small number of instances whose visibility state has changed, decoupling the frame-by-frame computational load of occlusion culling from the absolute scene density and instead relating it to the rate of visibility change. Thus, even in high-density urban scenes, the frame time proportion of the occlusion culling stage can be stably controlled within 5% of the total frame time budget, ensuring the predictability of frame time allocation in each stage of the rendering pipeline.

[0053] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that each subsystem is based on strict physical quantities or mathematical models to drive decisions, rather than relying on fixed parameters based on experience. The screen space projection area is uniquely determined by the perspective projection geometry, and using it as a batch load metric ensures that the equal load partitioning has theoretical optimality, guaranteeing the balanced utilization of the parallel computing units of the graphics processor. The three-dimensional Hilbert curve maps the three-dimensional tile coordinates to one-dimensional integer indices, and its space-filling property ensures the continuity of adjacent coordinates on the curve indices, thus making the memory access pattern naturally match the hardware cache line structure, and the improvement of cache hit rate has information theory-level guarantees. The comprehensive quality adjustment value unifies perceived quality and computational overhead into the same scalar index, and the piecewise adjustment strategy of the high-frequency gating weight inference scaling coefficient is based on piecewise linear regression to determine the threshold, so that the switching of the model inference strategy has a data-driven statistical basis. The subsystems collaborate through shared frame-level temporal information. Kalman filtering prediction, temporal reprojection, and temporal anti-aliasing are all linked by motion vectors, enabling prefetching, occlusion culling, and anti-aliasing to form a closed-loop feedback in the temporal domain, thereby ensuring rendering quality in the three dimensions of geometry, texture, and visibility within the frame time budget.

[0054] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0055] The specific implementation of step S01 is as follows: The central processing unit calculates the screen space projection area of ​​each geometric block within the viewport frame by frame. The calculation formula is expressed as follows:

[0056] ;

[0057] In the formula, For the first The world space surface area of ​​a geometric block. Let be the angle between the normal vector of the geometric block and the direction of the line of sight. Let be the distance from the center of the geometric block to the camera. The focal length parameter is the projection focal length, derived from the field of view angle. With screen pixel width according to Calculated. (According to...) After sorting all geometric blocks in descending order, a greedy bin packing algorithm is used to repackage the subdivision task into equal-load subdivision batches, so that each batch... Approximately equal, the batch parameters are then written to the graphics processor buffer through the drawing indirect instruction mechanism, thereby achieving asynchronous decoupling between the central processing unit and the graphics processor.

[0058] The specific implementation of step S02 is as follows: The horizontal coordinates of the map tiles within the viewport... Vertical coordinates With scaling levels After normalization, they are mapped to one-dimensional integer indices using three-dimensional Hilbert curve encoding. The calculation formula is expressed as follows:

[0059] ;

[0060] In the formula, For encoding functions of 3D Hilbert curves, the 3D coordinate triplet is used. The mapping is a one-dimensional integer that preserves spatial locality. User velocity vector. With acceleration It is calculated frame-by-frame from the perspective motion trajectory, and the formula is as follows:

[0061] , ;

[0062] In the formula, For the first Frame view position vector This represents the inter-frame time interval. The Hilbert sequence number range of the prefetch window is... Prediction number Calculated by the following formula:

[0063] ;

[0064] In the formula, The estimated prefetch time lead is 2 to 4 frame intervals. To pre-determine the window radius, based on the current network bandwidth Data volume per tile according to The estimate is determined. This is a floor function. When network bandwidth is insufficient, it degenerates to a lower Hilbert recursion level. Subtract 1 and re-encode.

[0065] The specific implementation of step S03 is as follows: the map texture super-resolution fusion model uses normalized coordinate triples. For input, where , , Normalize the horizontal coordinates, vertical coordinates, and scaling levels respectively. The dimensionless value after that. Fourier feature location encoding will... Mapped to 256-dimensional feature vectors The formula is expressed as follows:

[0066] ;

[0067] In the formula, for A random frequency matrix, where each element has a mean of 0 and a standard deviation of . Sampling from a Gaussian distribution, The empirical value is 10. The implicit branching multilayer perceptron uses a sinusoidal activation function, and the outputs of each layer are:

[0068] ;

[0069] In the formula, For the first Layer weight matrix, For the first Layer bias vector, Each layer is 512 pixels wide. This is the input for the implicit branch. The frequency separation gating network uses the local gradient magnitude. For input, From texture image in pixels horizontal gradient at with vertical gradient according to The calculated output pixel-by-pixel high-frequency gated weights are obtained. The low-frequency gating weight is The super-resolution texture output is:

[0070] ;

[0071] In the formula, This is the high-frequency feature map of the implicit branch output. This is the low-frequency feature map of the explicit branch output. for Convolutional output layer. Total loss function during model training. The calculation formula is expressed as follows:

[0072] ;

[0073] In the formula, For pixel-by-pixel mean square error loss, In order to perceive loss, To determine the perceived loss weight, a grid search is performed on the validation set within the range of 0.01 to 0.1, and the weight value that best combines the peak signal-to-noise ratio (PSNR) and structural similarity index on the validation set is used. (Comprehensive quality adjustment value) The calculation formula is expressed as follows:

[0074] ;

[0075] In the formula, The average structural similarity index for the current batch of reasoning. The average peak signal-to-noise ratio of the current batch of inference. For the current frame's graphics processor inference latency, , , These are the benchmark structural similarity index, benchmark peak signal-to-noise ratio, and benchmark inference delay, respectively, as determined by the validation set. , , The weighting coefficients are obtained by performing a grid search on the validation set for different weight combinations. The weight combination is determined when the correlation coefficient with the actual perceived quality is maximized. High-frequency gated weight inference scaling factor. in accordance with Range adjustment: when hour, Keep the current value unchanged; when hour, Increment by 0.05 steps; when hour, Decrease in increments of 0.10; when hour, Set it to the minimum value and switch to the low-frequency feature map separate inference mode.

[0076] The specific implementation of step S04 is as follows: Let the pixel be... The signed distance field to the level detail boundary is Positive values ​​are taken inside the boundary, and negative values ​​are taken outside. Adjacent level detail layer. and The weighting of transparency blending between elements is determined by a sliding interpolation function, expressed by the following formula:

[0077] , ;

[0078] In the formula, This is the interpolation steepness control parameter; an empirical value of 6 is used. To determine the width of the transition zone, subjective visual evaluation data was collected at different screen resolutions and scaling speeds. Experiments were conducted within the range of 32 pixels to 128 pixels, using 8-pixel steps. The step size that yielded the best score and had acceptable computational cost was selected for determination. This represents the distance ratio after dimensionless processing. Multi-level progressive texture offset of layers Follow Smooth transition, the formula is expressed as follows:

[0079] ;

[0080] In the formula, and The first Layer and First The base texture sampling bias parameters of the layer are all dimensionless values. Current frame blending color. It is obtained by two-layer weighted mixed calculation, and the formula is expressed as follows:

[0081] ;

[0082] In the formula, For the first Layer of detail at the pixel level Texture sampling color value at that location, For the first Layer of detail at the pixel level The texture sample color value at that location. The final output pixel color. Further smoothing is achieved through feedback from historical frames used in temporal anti-aliasing, as expressed in the following formula:

[0083] ;

[0084] In the formula, For historical frame colors, For pixels The motion vector at that location is used to compensate historical frame coordinates to the current frame coordinate system. This is the mixing coefficient for the current frame, with an empirical value of 0.1 to 0.2.

[0085] The specific implementation of step S05 is as follows: Kalman filtering uses the user velocity vector With acceleration For input, state vector By alternating between prediction and update steps, the estimated future frustum range is output, triggering asynchronous input / output requests in advance. Super-resolution textures and streaming textures are sequentially backfilled using a four-level caching system: GPU memory, system memory, non-volatile memory, high-speed interface solid-state drives, and network content delivery network. Each caching level uses the least recently used cache. Access policy management based on data access priority The timestamp of the access is used as the basis for eviction priority, and high-frequency hotspot tiles are pinned to the L1 graphics processor memory cache. The experience value is 2.

[0086] The specific implementation of step S06 is as follows: The first phase uses the hierarchical occlusion map of the previous frame to perform conservative occlusion culling on building instances. The hierarchical occlusion map is constructed by downsampling the depth buffer step by step, and the minimum normalized depth value of the bounding box of each instance is used. Maximum normalized depth value corresponding to the layer of the hierarchical occlusion map When comparing, An instance is determined to be occluded and removed, resulting in a candidate instance set. A second relative candidate instance set is used to perform an exact occlusion query using a computational shader, combined with temporal reprojection and motion vectors. Compensate for the visibility result of the previous frame, retest only for instances where the visibility state has changed, and keep the per-frame overhead of occlusion culling within the frame time budget to no more than 5% of the occlusion culling stage.

[0087] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians set up a test environment and deployed the full-process method of the invention on a standard map rendering test platform to conduct continuous rendering tests on a map scene covering high-density urban areas, rivers, large areas of green space and major road networks. During the test, users were simulated to perform translation, scaling and rotation operations with different speed vectors and accelerations, and the key performance indicators of each subsystem were recorded throughout the process.

[0088] The test platform was configured as follows: the graphics processor had a video memory capacity of 8192 kilobytes per second. The system memory capacity is 32 The sequential read speed of the non-volatile memory high-speed interface solid-state drive is 3500. The network bandwidth is set to 50. The test scene has a total of 4096 map tiles, with tile resolutions ranging from 128×128 to 4096×4096 pixels, and a total of 12800 building instances. The scene depth complexity is moderate to high.

[0089] In the equal load subdivision batch division stage of step S01, the technicians performed screen space projection area pre-calculation on all geometric blocks in the viewport. The calculation results are shown in Table 1, which records the screen space projection area distribution and load variance between batches under different viewing angle speed levels, and verifies the equal load division balance effect.

[0090] Table 1 Load Distribution of Different Load Sub-batches at Different Viewpoints and Speeds

[0091]

[0092] As shown in Table 1, the load variance between batches remained at a low level under each speed setting, verifying the effectiveness of the equal load subdivision batch segmentation strategy.

[0093] In the tile scheduling stage of step S02, the technicians recorded the comparison of cache hit rate and prefetch hit rate data before and after Hilbert curve mapping, as shown in Table 2.

[0094] Table 2 Comparison of Cache Hit Rates for Tile Scheduling Strategies

[0095]

[0096] As shown in Table 2, the Hilbert curve mapping significantly improves the hit rate of each level of cache. Combined with the user motion prediction-driven prefetching mechanism, the prefetch hit rate reaches 0.883, effectively reducing tile loading latency.

[0097] In the super-resolution inference stage of step S03, such as Figure 2 As shown, the technicians adjusted the overall quality value. The distribution under different combinations of viewpoint velocity and zoom frequency was statistically analyzed, verifying the effectiveness of the high-frequency gated weight adaptive adjustment function in various applications. The test investigated whether the switching of inference modes within the value range was highly correlated with subjective visual ratings. The results showed that the piecewise threshold setting of the high-frequency gating weight adaptive adjustment function resulted in a high degree of agreement between the inference mode switching point and the abrupt change in the slope of the subjective rating curve, verifying the rationality of the threshold calibration method.

[0098] In the layer detail gradient blending stage of step S04, with a transition zone width of 64 pixels, the technicians performed pixel-by-pixel transparency blending weight verification at the layer detail layer switching boundary. This confirmed that the blending function driven by the signed distance field value was monotonically continuous throughout the entire transition zone, with no abrupt pixel changes. The number of blending frames for the temporal anti-aliasing history frame feedback was set to 8 frames, and it was verified that the residual flicker amplitude was reduced to a level imperceptible to the human eye.

[0099] In the proactive prefetching and hierarchical caching stage of step S05, the technicians recorded the convergence process of the Kalman filter prediction error with the number of frames, as well as the refill frequency of each level of cache, as shown in Table 3.

[0100] Table 3. Statistics on the frequency of hierarchical cache refilling

[0101]

[0102] As shown in Table 3, the vast majority of tile requests are satisfied at the graphics processor memory or system memory layer, with very few cross-layer data transfers. The prefetching mechanism driven by Kalman filter effectively reduces the frequency of network layer backfilling and keeps the average backfilling latency within the rendering frame time budget.

[0103] In the dual-phase occlusion removal stage of step S06, the technicians calculated the size of the candidate instance set after the first phase conservative removal and the time taken for the second phase precise query for 12,800 building instances. The frame time ratio of the occlusion removal stage was consistently less than 5%, which verified the effectiveness of the dual-phase removal pipeline and the temporal reprojection collaborative mechanism.

[0104] Compared to traditional methods, the advancements of this invention are reflected in the following aspects. In the geometric subdivision stage, traditional methods divide batches based on fixed levels, resulting in significant differences in computational load between different batches. This leads to some threads in the parallel units of the graphics processor running idle for extended periods. In contrast, this invention dynamically divides load batches based on the screen space projection area, making the actual utilization rate of the parallel computing units closer to the theoretical peak. In the tile scheduling stage, traditional rectangular scanning strategies ignore the cache locality of the space-filling curve. This invention utilizes Hilbert curves to map spatially adjacent tiles into a continuous sequence in memory, making the batch hits of cache lines mathematically inevitable rather than relying on the randomness of system prefetching. In the super-resolution inference stage, traditional methods employ fixed inference strategies, failing to adaptively adjust model inference overhead when frame time fluctuates. This invention, by comprehensively adjusting the quality adjustment value, models perceptual quality and inference latency in a unified manner, achieving continuous adaptive switching of the inference strategy, ensuring that texture quality remains optimal within the frame time budget. These three technological advancements synergistically make the frame time allocation at each stage of the rendering pipeline more predictable, systematically improving overall rendering quality and frame rate stability in complex perspective motion scenes.

[0105] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0106] Table 4. Variable Explanation Table (Part 1)

[0107]

[0108] Table 5. Variable Explanation Table (Part Two)

[0109]

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A cartographic method for real-time map rendering optimization, characterized in that, Includes the following steps: The screen space projection area of ​​each geometric block is pre-calculated at the central processing unit. After sorting by screen space projection area, the subdivision tasks are repackaged into equal-load subdivision batches and submitted to the graphics processor for execution asynchronously through the drawing indirect instruction mechanism. Map tile coordinates within the viewport are mapped to Hilbert linear indices using 3D Hilbert curves. Priority queues are established based on the Hilbert linear indices. Quadtree divide-and-conquer recursive scheduling is performed with the viewport center as the root node. The next batch of Hilbert indices is predicted based on the user velocity vector and user acceleration, and prefetching is triggered. When network bandwidth is insufficient, it degenerates to a lower-level Hilbert recursive hierarchy. A map texture super-resolution fusion model is used to reconstruct low-resolution tile textures using super-resolution methods, resulting in super-resolution textures. Calculate the overall quality adjustment value, and adjust the high-frequency gated weight inference scaling factor of the map texture super-resolution fusion model according to the range of the overall quality adjustment value through the high-frequency gated weight adaptive adjustment function; Based on the signed distance field value from the pixel to the level detail boundary, the texture of adjacent level detail layers is blended with transparency. The multi-level progressive texture bias of the two level detail layers is smoothly transitioned through a sliding interpolation function. Temporal anti-aliasing is combined with historical frame feedback to smooth residual flicker. Kalman filtering is used to estimate the future frustum range and asynchronous input / output requests are triggered in advance. Super-resolution textures and streaming textures are backfilled in a hierarchical caching system, and each level of cache is managed using the least recently used Kth access strategy. High-frequency hotspot tiles are pinned to the first-level graphics processor memory cache. The first phase uses the hierarchical occlusion map of the previous frame to perform conservative occlusion culling on building instances, resulting in a candidate instance set; the second phase uses a computational shader to perform precise occlusion query on the candidate instance set; combined with temporal reprojection to compensate for the visibility results of the previous frame with motion vectors, only instances whose visibility state has changed are retested.

2. The cartographic method for real-time map rendering optimization according to claim 1, characterized in that, The screen space projection area specifically refers to the pixel area occupied by the three-dimensional geometric object after it is projected onto the two-dimensional pixel plane of the screen, which is calculated frame by frame by the central processing unit before the subdivision task is submitted.

3. The cartographic method for real-time map rendering optimization according to claim 2, characterized in that, The equal-load subdivision batch refers to regrouping vector geometric subdivision tasks according to the pre-calculated screen space projection area, so that the computational load of the graphics processor in each batch is approximately equal.

4. The cartographic method for real-time map rendering optimization according to claim 3, characterized in that, The aforementioned indirect drawing instruction mechanism specifically refers to pre-storing parameters in equal-load subdivision batches into the graphics processor buffer through the graphics application programming interface, which is then read and executed autonomously by the graphics processor, thereby achieving asynchronous decoupling between the central processing unit and the graphics processor.

5. The cartographic method for real-time map rendering optimization according to claim 4, characterized in that, The Hilbert linear index specifically refers to the one-dimensional integer index obtained by mapping the horizontal and vertical coordinates of a map tile to the scaling level triplet through three-dimensional Hilbert curve encoding.

6. The cartographic method for real-time map rendering optimization according to claim 5, characterized in that, The user velocity vector and user acceleration are specifically calculated frame-by-frame from the viewpoint motion trajectory and are used to drive the prefetch window to slide and predict the next batch of Hilbert sequence intervals.

7. The cartographic method for real-time map rendering optimization according to claim 6, characterized in that, The map texture super-resolution fusion model specifically includes an explicit branch and an implicit branch; the explicit branch is a convolutional upsampling network containing residual convolutional blocks, which outputs low-frequency feature maps; the implicit branch is a multilayer perceptron using a sinusoidal activation function, which outputs high-frequency feature maps; the low-frequency feature maps and high-frequency feature maps are dynamically fused through a frequency separation gating network to output super-resolution textures.

8. The cartographic method for real-time map rendering optimization according to claim 7, characterized in that, The comprehensive quality adjustment value The calculation uses a weighted combination of the normalized values ​​of the structural similarity index, the normalized values ​​of the peak signal-to-noise ratio, and the normalized values ​​of the graphics processor inference latency. The weight coefficients are determined by performing a grid search on the validation set.

9. The cartographic method for real-time map rendering optimization according to claim 8, characterized in that, The high-frequency gated weight adaptive adjustment function is specifically based on the comprehensive quality adjustment value. Range of high-frequency gated weight inference scaling factor Segmented adjustments are made, and when the overall quality adjustment value is lower than the minimum quality threshold, the system switches to the low-frequency feature map separate inference mode.

10. The cartographic method for real-time map rendering optimization according to claim 9, characterized in that, The signed distance field value specifically refers to the signed nearest distance from a pixel to the boundary of the level detail, with positive values ​​taken inside the boundary and negative values ​​taken outside, and is calculated from the screen space distance field.

Images