A method for automatically calculating and checking quality of road width of road network data

Abstract

The application discloses a kind of road width automatic calculation and quality checking method of road network data, it is related to road network data processing technical field, to solve the problem of low efficiency of multi-source data fusion in existing method;The application is through collecting multi-source road network data and pre-standardization processing, unified coordinate system mapping and metadata annotation generate labeled original dataset;Perform noise adaptive filtering and occlusion detection, utilize U-Net semantic segmentation to identify interference area, and apply GAN local completion output purification dataset;Construct the multi-modal network calculation dynamic fusion weight of attention mechanism, and parallel integration forms fusion feature vector;Through Canny optimization edge extraction and DBSCAN clustering calculation width parameter, embedding topological structure;Use multi-scale consistency index to check result, if threshold backtracking optimization;Output result and iteratively update parameter, form closed loop mechanism.The method improves fusion efficiency and calculation accuracy, applicable to traffic navigation and urban planning, supports real-time dynamic application.

Description

Technical Field

[0001] This invention relates to the field of road network data processing technology, specifically to a method for automatic calculation and quality verification of road width in road network data. Background Technology

[0002] Automatic calculation and quality verification of road width in road network data is one of the core technologies supporting high-precision road network modeling in fields such as traffic navigation, urban planning, and autonomous driving. Existing technologies typically rely on the fusion processing of multi-source data (such as remote sensing imagery, vehicle trajectory point clouds, and GIS vector data), using techniques like edge extraction, cluster analysis, or image registration to achieve quantitative evaluation and quality verification of road width. For example, CN111126427A discloses a method for evaluating the quality of OpenStreetMap (OSM) road network data based on remote sensing imagery. This method simplifies the OSM road network and registers high-resolution imagery, using edge features to calculate road width to assess data quality. CN105975913B proposes a road network extraction method based on adaptive clustering learning, calculating road width through trajectory point connection and cluster analysis, and performing preliminary verification. These technologies improve the efficiency of automated processing of road network data to a certain extent, providing fundamental support for data modeling and optimization in related application areas.

[0003] However, existing methods for automatic road width calculation and quality verification of road network data still have significant shortcomings in the multi-source data fusion stage, resulting in low overall computational efficiency and insufficient reliability, making it difficult to meet the requirements of high precision and real-time performance. On the one hand, the heterogeneity of multi-source data makes it difficult to efficiently align key parameters such as coordinate systems and resolution, leading to high computational complexity in the fusion process and high latency in width calculation in real-time applications. On the other hand, the fusion algorithm lacks a dynamic optimization mechanism, relying solely on fixed thresholds or simple stitching strategies, which cannot adapt to the exponential growth of data scale, resulting in a non-linear surge in fusion time and resource consumption. Furthermore, in scenarios with noise interference or occlusion, the lack of effective integration of multi-temporal data or auxiliary information for adaptive completion results in missing road segment widths or calculation errors. These shortcomings directly reduce the overall reliability of road network data, easily leading to navigation path errors and wasted planning resources, severely limiting its application value in high-precision real-time road network modeling scenarios.

[0004] Therefore, it is necessary to provide an improved method for automatic calculation and quality verification of road width in road network data to overcome the problem of low efficiency in multi-source data fusion in existing technologies. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method for automatically calculating and verifying the quality of road width in road network data, thereby solving the problem of low efficiency in multi-source data fusion in existing technologies.

[0007] (II) Technical Solution

[0008] To achieve the goal of improving the efficiency of multi-source data fusion mentioned in the background section, the present invention provides the following technical solution:

[0009] A method for automatically calculating and verifying the quality of road width in road network data, comprising:

[0010] S1: Collect multi-source road network data and perform pre-standardization processing. Through unified coordinate system mapping and metadata annotation, convert remote sensing images, vehicle trajectory point clouds and GIS vector data into a compatible format to generate a labeled raw dataset.

[0011] S2: Perform noise adaptive filtering and occlusion detection, use a deep learning model to identify noise interference and occlusion areas, and apply a generative adversarial network for local completion to output a cleaned dataset;

[0012] S3: Construct an efficient multi-source fusion framework, use a multimodal network with attention mechanism to calculate dynamic fusion weights, input the purified dataset into the parallel computing module to form a fusion feature vector;

[0013] S4: Calculate road width parameters, process the fused feature vector through edge extraction algorithm and cluster analysis, quantify the road segment width value, and embed it into the road network topology;

[0014] S5: Perform quality verification and evaluation, and use multi-scale consistency indicators to verify the calculation results. If the deviation exceeds the preset threshold, trigger backtracking optimization of the fusion weight.

[0015] S6: Outputs optimization results and updates iteratively, adjusts fusion parameters based on verification feedback to form a closed-loop mechanism, and supports the dynamic application of road network data.

[0016] In a preferred embodiment, multi-source road network data is collected and pre-standardized. Through unified coordinate system mapping and metadata annotation, remote sensing imagery, vehicle trajectory point clouds, and GIS vector data are converted into a compatible format to generate a labeled raw dataset, including:

[0017] A unified coordinate datum is used to map data from different sources to the same spatial reference system and align scale, resolution, and timestamps.

[0018] Metadata annotation includes source device or service identifier, collection time, spatial coverage, resolution description, and data quality identifier;

[0019] Before data is written to disk, quality gating is performed. The quality gating is limited to: automatic detection of image sharpness and contrast indicators, investigation of trajectory integrity and point density anomalies on the point cloud side, and topological validity and field consistency checks on the vector side. If any step fails to meet the standard, re-sampling or marking it as "grayscale input" that needs to be reviewed is triggered.

[0020] The division of responsibilities is carried out between edge nodes and the cloud: edge nodes are responsible for local data collection and format standardization, while the cloud is responsible for batch verification and master data registration.

[0021] In a preferred embodiment, noise adaptive filtering and occlusion detection are performed, a deep learning model is used to identify noise interference and occlusion regions, and a generative adversarial network is applied for local completion, outputting a cleaned dataset, including:

[0022] Establish a cleanup process of "problem area annotation - local reconstruction - quality write-back", output cleanup dataset that simultaneously carries the spatial range, source type and confidence level of the problem area, and provides binary or multi-value masks in the same resolution / coordinate as the original data;

[0023] Problem area annotations are used to identify typical interferences such as clouds, shadows, vegetation, building obstruction, GPS drift, and abnormal sparseness;

[0024] Local reconstruction is used to complete missing information while preserving the continuity of boundary geometry and texture / structure;

[0025] Quality writeback is used to write the spatial extent, confidence level, and quality score of the reconstructed region into the structured log of the cleaned dataset and establish a one-to-one correspondence with the corresponding source data;

[0026] When the purification quality score is lower than the set threshold or the proportion of problematic areas exceeds the limit, a backtracking identification or resampling prompt is triggered.

[0027] In a preferred embodiment, an efficient multi-source fusion framework is constructed, employing a multimodal network with an attention mechanism to compute dynamic fusion weights. The purified dataset is input into a parallel computing module to form a fusion feature vector, including:

[0028] Parallel processing pathways for image branches, point cloud branches, and vector topology branches are established using modal splitting and adaptive weight allocation. In the fusion layer, the features of each branch are weighted, stitched together, and uniformly represented according to dynamic weights.

[0029] Image branching is used to extract spatial edges and texture semantics;

[0030] Point cloud branches are used to characterize local geometric and density variations;

[0031] Vector topology branches are used to encode the connectivity and hierarchical attributes of road segments and intersections;

[0032] The fusion output is organized as a "fusion feature vector carrying road segment identification", which serves as the sole input for subsequent width calculation and includes a source mapping.

[0033] The convergence framework supports coordinated scheduling of batch processing and parallel computing in resource-constrained scenarios, allowing parameterized configuration of batches, caches, and parallelism based on throughput and memory constraints.

[0034] In a preferred embodiment, road width parameters are calculated, and the fused feature vector is processed using an edge extraction algorithm and cluster analysis to quantify the road segment width value and embed it into the road network topology, including:

[0035] Edge extraction is performed on the fused features to obtain a set of road boundary points. The set of boundary points is clustered into pairs of curb clusters according to density and geometric consistency. The spatial proximity relationship and the consistency of the normal angle between the curb pairs are used as constraints for width calculation to obtain the width sequence along the road segment.

[0036] Based on spatial indexing, the width attribute is accurately matched with road segments and written as edge attribute in the loop network diagram. At the same time, it completes: structural repair of isolated nodes, hanging edges and obviously unclosed boundaries, outlier labeling and recording of abnormally fluctuating widths, and local geometric fidelity strategy for curves / complex intersections to ensure the consistency between width parameters and topology.

[0037] The output topology enhancement graph includes the representative width of each edge, width fluctuation statistics, and a list of outliers.

[0038] In a preferred embodiment, a quality verification and evaluation is performed, and the calculation results are verified using a multi-scale consistency index. If the deviation exceeds a preset threshold, a backtracking optimization of the fusion weights is triggered, including:

[0039] Establish a multi-scale consistency index system that combines macro and micro perspectives, supplemented by topological integrity and boundary closure indicators, to support the simultaneous identification of both network-wide and single-segment problems;

[0040] Macro-level consistency is used to calculate the average deviation across the entire network and to provide an overall risk threshold.

[0041] Microscopic consistency is used to check whether deviations exceed limits segment by segment and to locate local anomalies such as curves.

[0042] Topological integrity is used to measure whether connected components and connectivity satisfy engineering constraints.

[0043] Boundary closure is used to measure whether the spatial separation between the first and last points of a boundary polygon is acceptable.

[0044] When any indicator exceeds the limit, the backtracking process is triggered in the order of "local first, then global, and then backtracking", and the road segment exceeding the limit, the problem type and the trigger threshold are recorded in the structured feedback to support subsequent parameter optimization.

[0045] The indicator system allows for parallel computation and maintains manageable computational complexity as the scale expands.

[0046] In a preferred embodiment, the optimization results are output and iteratively updated, and the fusion parameters are adjusted based on verification feedback to form a closed-loop mechanism, supporting the dynamic application of road network data, including:

[0047] The verification feedback is carried out using structured data in a closed loop of "initialization - parameter fine-tuning - local recalculation - result verification";

[0048] The initialization phase is used to load the previous round's state and adjustable threshold of the fusion parameters;

[0049] Parameter fine-tuning is used to adjust the processing thresholds associated with cross-modal weights between the edge and the cloud according to the principle of minimum necessity;

[0050] Local recalculation is used to regenerate the fusion features and width results within the problem road segment;

[0051] The result review is used to determine whether the convergence threshold has been met and to decide whether to publish;

[0052] If there are still road sections that exceed the limits after optimization, rollback to the backup parameter set is allowed to ensure task continuity, and the rollback reason, scope of impact and recovery strategy are included in the audit log.

[0053] The output is provided both in a GIS-friendly vector data format and in a structured report format with version, threshold, and difference descriptions.

[0054] In a preferred embodiment, it includes:

[0055] Logically fragment edge nodes according to geographical region, factory or road level and maintain routing and health status;

[0056] When adding or removing nodes, seamless scaling is achieved through task reallocation and state migration, maintaining mapping context and data consistency during migration.

[0057] When a single point of failure or network isolation occurs, adjacent nodes temporarily take over the data collection and calculation tasks in the affected area according to a preset strategy. After recovery, two-way reconciliation and difference merging are performed.

[0058] When collaborating across regions, master data synchronization and version orchestration ensure the consistency of attributes of the same road element in different domains, and the scope of impact and order of effectiveness of cross-domain corrections are managed in a controlled manner.

[0059] In a preferred embodiment, it includes:

[0060] Record the time, subject, and related data summaries of key events such as width calculation, quality assessment, parameter adjustment, link switching, and rollback at the edge and cloud respectively, and aggregate them using a unified log model;

[0061] To meet higher levels of compliance or multi-entity collaboration requirements, the log model can be connected to immutable storage with a trusted time source to solidify key change records and support cross-domain verification.

[0062] In the event of a dispute or security incident, the event sequence can be quickly reconstructed based on the log chain, the version evolution can be replayed, and responsibility can be identified.

[0063] Compared with existing technologies, this invention provides a method for automatic calculation and quality verification of road width in road network data, which has the following advantages:

[0064] 1. This invention first collects remote sensing image data using satellite remote sensing equipment, vehicle trajectory point cloud data using a vehicle-mounted LiDAR system, and extracts vector data from a GIS database. After collection, a unified coordinate system mapping algorithm is used, and labels are added through metadata annotation to generate a labeled raw dataset. This pre-standardization reduces heterogeneous conflicts and improves compatibility. Furthermore, a multimodal network with an attention mechanism is constructed to calculate dynamic fusion weights, and the purified dataset is input into a parallel computing module to form a fused feature vector. Through modal processing and dynamic weight adaptation, the fusion efficiency is significantly improved, real-time latency is reduced, dynamic road network applications are supported, and the fusion complexity is significantly reduced, thus solving the problem of low efficiency in multi-source data fusion in traditional methods.

[0065] 2. This invention inputs the original dataset into a deep learning model to identify noise interference and occlusion areas; then, a generative adversarial network is applied for local completion, with completion accuracy controlled within 0.1m; noise is classified into dynamic and static categories, with dynamic noise removed through time-series analysis and static noise adaptively adjusted using a threshold; after outputting the cleaned dataset, an integrity check is performed, and if the coverage is <95%, re-completion is triggered; through adaptive networks and adversarial completion, the proportion of noise is reduced, the risk of noise amplification error >10% is also reduced, the accuracy of quality verification is improved, and width deviation is reduced in high-noise environments, ensuring the stability and security of road network data, supporting reliable autonomous driving and urban planning applications, and reducing the potential accident rate.

[0066] 3. This invention quantifies road segment width values ​​and embeds them into the topology by applying edge extraction algorithms and cluster analysis to the fused feature vectors; the processing includes curvature compensation, outputting the average width and variation range with centimeter-level accuracy; and the optimization algorithm ensures accuracy >95%; it reduces the width error in complex environments from >2m to the centimeter level, improves calculation accuracy, reduces deviations in curve intersection scenarios, ensures the reliability of quality verification, supports accurate autonomous driving path optimization, and reduces traffic accident hazards. Attached Figure Description

[0067] Figure 1 This is a flowchart of a method for automatically calculating and verifying the quality of road width in road network data according to the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example: Figure 1 This invention provides a method for automatic calculation and quality verification of road width in road network data, comprising:

[0070] S1: Collect multi-source road network data and perform pre-standardization processing. Through unified coordinate system mapping and metadata annotation, convert remote sensing images, vehicle trajectory point clouds and GIS vector data into a compatible format to generate a labeled raw dataset.

[0071] S2: Perform noise adaptive filtering and occlusion detection, use a deep learning model to identify noise interference and occlusion areas, and apply a generative adversarial network for local completion to output a cleaned dataset;

[0072] S3: Construct an efficient multi-source fusion framework, use a multimodal network with attention mechanism to calculate dynamic fusion weights, input the purified dataset into the parallel computing module to form a fusion feature vector;

[0073] S4: Calculate road width parameters, process the fused feature vector through edge extraction algorithm and cluster analysis, quantify the road segment width value, and embed it into the road network topology;

[0074] S5: Perform quality verification and evaluation, and use multi-scale consistency indicators to verify the calculation results. If the deviation exceeds the preset threshold, trigger backtracking optimization of the fusion weight.

[0075] S6: Outputs optimization results and updates iteratively, adjusts fusion parameters based on verification feedback to form a closed-loop mechanism, and supports the dynamic application of road network data.

[0076] S1: Collect multi-source road network data and perform pre-standardization processing. Through unified coordinate system mapping and metadata annotation, convert remote sensing imagery, vehicle trajectory point clouds, and GIS vector data into a compatible format to generate labeled raw datasets. The specific implementation is as follows:

[0077] In the multi-dimensional data visualization method for road networks, multi-source road network data is first collected and pre-standardized to achieve comprehensive perception and data unification of the urban or regional road network environment, ensuring the accuracy and consistency of subsequent fusion and rendering. This multi-source road network data includes satellite remote sensing imagery data from high-resolution optical sensors to capture the macroscopic structure of the road network; vehicle-mounted LiDAR point cloud data from laser scanning systems to acquire three-dimensional details of the road network; and GIS vector data from geographic information databases to provide road network topology information. The selection of data sources is based on the complexity of the road network and application requirements. For example, in densely populated urban areas, remote sensing imagery can reflect the overall layout, while LiDAR point cloud data can reflect the overall structure of the road network. DAR point clouds supplement height and edge details, while GIS vectors provide logical connections. Based on the statistical distribution of road network change rates, the acquisition frequency is uniformly set to once per task cycle to balance real-time performance and resource consumption, ensuring the capture of dynamic changes without generating excessive data. The entire acquisition and pre-standardization process is executed on edge computing nodes or local servers, configured with high-performance embedded processors. Operational latency is controlled within 50ms based on end-to-end system response time requirements. Through coordination with satellite equipment, vehicle systems, and database interfaces, an automated chain from acquisition to standardization is achieved. This not only reduces heterogeneous conflicts but also provides high-quality input for downstream multi-layer fusion.

[0078] In the multi-source road network data acquisition of this invention, two-dimensional images are first acquired using satellite remote sensing equipment: a high-resolution optical image sensor is deployed on a low-orbit satellite at an altitude of approximately 500 km to image the target road network and extend it to a 1.5-fold buffer. Clear weather conditions are selected, and the images are stored in GeoTIFF format along with metadata such as satellite orbit parameters and imaging angles. After acquisition, a preliminary inspection is performed according to a contrast formula for quality gating. The contrast formula is:

[0079]

[0080] A contrast ratio greater than 0.05 is considered acceptable; otherwise, resampling is triggered, with a maximum of two resampling attempts. Secondly, the vehicle-mounted LiDAR system, installed on the roof of a dedicated survey vehicle, uses GPS-assisted positioning to collect 3D point cloud data along the road network path, including road surface, boundaries, and obstacle heights. This data is then calculated using the point cloud density formula.

[0081]

[0082] When the speed is greater than 30 km / h, a density of less than 800 points / m² results in a loss of more than 15% of details. To ensure a point cloud density of greater than 1000 points / m², the acquisition speed is controlled at 20 km / h. The data is stored in LAS format and vehicle trajectory logs are recorded. During the point cloud cleaning stage, motion-blurred points are removed (speed threshold < 5 km / h). Finally, vector features within the same range as the remote sensing buffer are extracted from the GIS database to obtain the road centerline and boundary polygons, as well as their road grade and width attributes. The topological integrity is checked using QGIS tools. Through multi-source device parameter optimization and quality control, comprehensive coverage and high accuracy of the road network data are ensured, laying the foundation for subsequent pre-standardization processing.

[0083] After data acquisition, pre-standardization processing is performed. First, the coordinate system of the acquired data is unified. Since the source data may use different coordinate systems, a unified coordinate system mapping algorithm based on the Proj4 / PROJ library is used for projection transformation to convert all data to the same coordinate system. A projection transformation algorithm based on the Proj4 library is then used to convert all data to the CGCS2000 coordinate system. This process includes resampling the image and reprojecting the point cloud using a matrix multiplication formula. The matrix multiplication formula is as follows:

[0084]

[0085] in, The transformation matrix is ​​4x4. After the transformation, coordinate consistency is verified by randomly sampling 100 points, with a required deviation of <0.1m. Next, labels are added through metadata annotation. The metadata includes data source, collection time, resolution information, and device ID. The annotation uses JSON to attach and embed the data file header to ensure traceability >99%. After annotation, a labeled raw dataset is generated. The dataset structure includes columns (timestamp, type, value, and label JSON) to ensure heterogeneous data compatibility and reduce subsequent conflicts.

[0086] To generate a labeled raw dataset, the data is first transformed at a mid-level to integrate multi-source features (dimension > 20, including image pixels, point cloud coordinates, and vector attributes). Min-Max scaling is then performed on numerical fields to unify the values. The scaling formula for the interval is:

[0087]

[0088] The "Source Tag" is retained as a column; then, datasets are generated by batching and stored, organized by hierarchical directory "Region ID / Date / Source", and written to columnar files according to naming rules; to ensure integrity, a checksum is calculated for each segment and recorded in a side vehicle file with the same name; after generation, sampling verification is performed, and compatibility is measured based on the parsing results of the system and downstream tools, with a target compatibility rate of >98%; in urban road network scenarios, 1,000 remote sensing images, tens of thousands of trajectory point clouds and GIS vector layers are uniformly formatted, and coordinate systems and field dictionaries are used, which significantly reduces subsequent heterogeneous conflicts, ensures data compatibility and replayability, and provides a high-quality, homogeneous input base for S2 noise filtering.

[0089] S2: Perform adaptive noise filtering and occlusion detection, use a deep learning model to identify noise interference and occlusion regions, and apply a generative adversarial network for local completion, outputting a cleaned dataset. The specific implementation is as follows:

[0090] The noise detection model was configured and initialized to achieve comprehensive noise perception and occlusion analysis of the original dataset. The model used was a U-Net-based semantic segmentation network with an encoder-decoder symmetric topology. The encoder extracts features through multiple convolutions in the downsampling path, while the decoder recovers spatial details step-by-step through transposed convolutions in the upsampling path, fusing semantic and texture information at different scales via skip connections. The model input consisted of multi-source heterogeneous data, including pixel matrices from remote sensing images and coordinate arrays from point clouds. The output was a segmentation mask, providing pixel-level labels on the image side and point-level labels on the point cloud side to identify noise or occlusion regions. For training and deployment, small-sample fine-tuning was performed on the COCO dataset after transfer learning, with a learning rate of 0.001, an Adam optimizer, a batch size of 16, and 100 training epochs to ensure convergence and stability. Generalization was defined; to facilitate field engineering applications, the boundaries for determining noise and occlusion were agreed upon during the preparation phase: image pixels covered by clouds and fog were identified as low-grayscale clusters with pixel intensities below 50, based on grayscale histograms; occlusion of road edges by buildings was identified as abrupt regions with a density increase of more than 30% compared to the normal road edge density in the neighborhood, based on point cloud distribution statistics; data interface adaptation was completed, an end-to-end inference pipeline was built based on PyTorch and deployed to GPU nodes, and the processing latency of a single inference was controlled within 100 milliseconds to meet the real-time processing requirement that the total system link latency does not exceed 200 milliseconds; with the above configuration, the model can effectively distinguish between image noise caused by clouds and fog and real road surface texture when facing complex scenes such as dense urban areas, and accurately label the road edge point cloud anomalies caused by building occlusion, providing high-quality noise-free input for subsequent map updates and road network element extraction;

[0091] In the noise and occlusion identification stage, the original dataset is fed into a U-Net-based semantic segmentation model. The encoding layer employs four convolutional layers (each layer uses a 3x3 convolutional kernel and downsampling with a stride of 2 until the original resolution is compressed to one-sixteenth). The decoding layer recovers the input scale by upsampling and establishing skip connections with the corresponding encoding layers. Finally, a probability map of each pixel or point belonging to "noise," "occlusion," or "normal" is output using softmax. For noise interference identification, for remote sensing imagery, suspected cloud and fog areas are first screened out using multispectral contrast. Cloud and fog areas are marked when the blue band reflectance is higher than 0.6 and the "cloud and fog index" (i.e., the normalized difference obtained by dividing the difference between blue and red by their sum) exceeds 0.3. For point cloud data, abnormally sparse clusters with local point densities below 500 points per square meter are identified as GPS data. Noise caused by drift; In the identification of occluded areas, for abrupt elevation changes in the edge point cloud caused by building occlusion, if the "height gradient" (i.e., the absolute value of the elevation difference between two points divided by the distance between the two points) between adjacent points is greater than 1 and the height jump exceeds 2m, it is marked as building occlusion; For the image edge occluded by trees, when the reflectivity of the green band is higher than 0.5, it is marked as canopy occlusion; After the identification is completed, a binary mask with the same resolution as the input is generated (1 represents the problem area and 0 represents the normal area), and an adaptive threshold update is performed based on the mean and standard deviation of historical data (the threshold is "mean plus twice the standard deviation", so as to be strict when the standard deviation is small in sunny weather and moderately relaxed when the standard deviation is large in rainy weather).

[0092] After identifying the target and noise regions, an engineered process using generative adversarial networks (GANs) is employed to adaptively reconstruct images or point clouds affected by cloud noise and occlusion. The generator uses a variant based on U-Net, taking as input the conditional context formed by the mask region and its surrounding normal data, and outputting as synthetic data of the region to be completed. The discriminator is a convolutional network used to distinguish between real and synthetic data and is trained adversarially against the generator. During image completion, a conditional Pix2Pix scheme is used, enabling the generator to generate missing pixels under the premise of "given neighborhood pixels and mask conditions" (i.e., the generator receives a noise vector and local context, and outputs synthetic pixels that match the context). During point cloud completion, a generator based on PointNet is used to generate missing points with the same density as the neighborhood by "concatenating local geometric features and noise signals and then inputting them into a multilayer perceptron". During the training phase, the joint adversarial loss and L1 reconstruction loss are optimized (i.e., the discriminator strives to increase the probability of correctly identifying real samples and decrease the probability of correctly identifying generated samples, while the generator strives to deceive the discriminator while minimizing the pixel / point difference from the ground truth). The batch size is 32, and the training is conducted for 200 rounds to ensure that the reconstruction error is less than 5%. After the completion is completed, the synthesized result is merged with the original data in a weighted fusion manner (the weight is "synthetic 0.7 + original 0.3"). The generator depth is adaptively adjusted according to the scale of the missing region (shallow layer for small regions and deep layer for large regions) to ensure that the structural continuity across the boundary is not less than 98%. The above network selection and training paradigm are consistent with the technical consensus in the literature on U-Net structure, conditional adversarial learning (Pix2Pix), and direct point cloud modeling (PointNet), which can ensure stable and reproducible completion quality and timeliness under engineering constraints.

[0093] After completion, the system outputs a "cleaned dataset," which is hierarchically organized and persistently stored in an HDF5 container. This facilitates efficient subset access and chunked reading of large volumes of multimodal features (HDF5 uses a group / dataset model and Hyperslab selection, combined with chunked layout and compression to improve local I / O performance, suitable for unified encapsulation of imagery, point clouds, and vector features). The data content includes: cleaned remote sensing imagery, point cloud data, and vector layers, as well as a completion log (recording the spatial extent and corresponding confidence level of the repaired area). During generation, a "global quality score" is calculated simultaneously, which is "1 minus the proportion of repaired pixels to total pixels." When this score is below 0.9, backtracking identification is automatically triggered for verification and re-completion. After output, this cleaned dataset is directly used as high-quality input for the S3 fusion stage, ensuring a stable and traceable data foundation for subsequent multi-source feature registration and consistency assessment.

[0094] S3: Construct an efficient multi-source fusion framework, employing a multimodal network with an attention mechanism to compute dynamic fusion weights. Input the purified dataset into the parallel computing module to form a fusion feature vector. The specific implementation is as follows:

[0095] The preparation phase involves the input configuration and framework initialization of the multi-source clean dataset, aiming to ensure high-quality and standardized input data for the fusion process. First, the clean dataset is obtained from step S2. This dataset has a unified coordinate system and format, and simultaneously includes the pixel matrix of remote sensing imagery, the 3D coordinate sequence of vehicle-mounted LiDAR point clouds, and the topological attribute table of GIS vectors. During preparation, the scale is set according to road network area coverage, based on GPU memory optimization tests. When the batch size is >128, the overflow rate is >5%. Therefore, the input batch size is set to 64 to achieve a balance between throughput and stability. Framework initialization uses TensorFlow. The 2.x library configures a multimodal network environment and preloads the Transformer module, while allocating GPU resources. At the data path level, a splitting logic is established: remote sensing images are assigned to the spatial branch for feature extraction by convolutional layers, point clouds are assigned to the temporal branch for sequence processing by RNN layers, and vectors are assigned to the topological branch for connection modeling by graph neural networks (GNNs). Through input standardization, batch control, and pre-configuration of computing resources, the stability of the fusion framework is ensured. For example, in highway networks, the initialization time is less than 1 second, avoiding the startup delay of traditional serial fusion.

[0096] In the multimodal network construction phase, a Transformer variant based on an attention mechanism is adopted, forming a three-branch structure of "remote sensing imagery - vehicle-mounted LiDAR point cloud - GIS vector topology": The imagery branch uses a ResNet-50 variant to extract spatial texture and edge semantics, outputting a feature map of size B×H / 16×W / 16×C (C=512) to provide global scene perception; the point cloud branch uses PointNet++'s hierarchical sampling and local aggregation to perform robust local geometry and density modeling on temporal point sets, outputting a B×N / 4×D (D=256) point cloud embedding; the vector branch uses GraphSAGE's two-layer neighborhood sampling and mean aggregation to model topological relationships such as road segments and intersections, outputting a B×M×E (E=128) node embedding; the three branches interact at the fusion layer through "cross-modal attention"; the features of different branches are first mapped to queries, keys, and values ​​(weights) through a learnable weight matrix. The system initializes with a normal distribution (mean 0, variance 0.02), then generates cross-modal attention scores and fused representations by multiplying each query by all keys, scaling by the square root of the key dimension, normalizing by softmax to obtain weights, and then using these weights to perform a weighted summation of the corresponding values. Training employs a contrastive learning objective to enhance intermodal correspondences. For each positive sample pair (the same target from different modalities), the system maximizes the normalized probability of its similarity under scaling with a temperature parameter τ=0.07, while simultaneously reducing the similarity between this sample and all negative sample pairs, thereby improving the discriminative power and cross-modal alignment of the representation. The implementation framework is based on PyTorch and supports DataParallel distributed training for horizontal scaling in multi-GPU environments.

[0097] After constructing the network, dynamic fusion weights are calculated. For the cleaned dataset, a multimodal network with an attention mechanism is used to assign weights to each modality branch. Specifically, the remote sensing image branch is assigned a high spatial prior weight based on the spatial resolution test results (initial value 0.4–0.6; when the image contribution exceeds 50%, it can bring about a 10% accuracy improvement). The specific method involves first performing global average pooling on the image features, then feeding them into a fully connected layer with 256 hidden units to obtain a spatial score, followed by softmax normalization to obtain the "spatial attention" weight. The trajectory point cloud branch sets a temporal prior weight based on the sequence length test (initial value 0.3–0.5; when the sequence length exceeds 1000 points, ...). The temporal contribution typically exceeds 40%. Specifically, the point cloud feature sequence is input into an LSTM network with a hidden dimension of 128. The hidden state vector at the last time step is taken and the "temporal attention" weight is obtained through softmax. The GIS vector branch sets topological prior weights based on the graph degree distribution (initial value is 0.2–0.4; when the node degree is greater than 5, the topological contribution typically exceeds 30%). Specifically, a 4-head graph attention network is used to perform weighted aggregation of the neighborhood on the topological graph of the road vector and its edges, and the "topological attention" weight is obtained through softmax. Then, the three types of attention weights are concatenated by channel, multiplied by a 3×3 learnable projection matrix on the left, and softmax is performed again to obtain the final cross-modal weight vector. The calculation is iterated in 2-3 rounds. When the weight change rate between two adjacent rounds is less than 0.01, convergence is determined, so that the weight can adapt to the differences in road network complexity and data quality. In actual resource allocation, the weight can be automatically adjusted according to the scenario. For example, in the curved road network area, the temporal weight of the point cloud is automatically increased to more than 0.5 to better capture trajectory turning information. At the same time, the redundant fusion calculation overhead is reduced in long straight road sections, which improves the overall fusion efficiency and steadily improves accuracy and robustness.

[0098] After calculating the weights of each modality feature, the cleaned dataset is fed into the parallel computing module. Based on the GPU-accelerated TensorFlow framework, synchronous distributed training is performed on 2–4 GPUs using MirroredStrategy. AllReduce aggregation is used to ensure gradient consistency across replicas. Data is input in batches of 64 and shuffled with a buffer of 1024 to ensure throughput and randomness. In the feature integration stage, each modality feature is first scaled according to its corresponding weight and concatenated into a unified representation before entering the multi-head attention fusion layer. Specifically, eight parallel attention heads perform "self-attention" matching on the same weighted feature sequence in a 64-dimensional subspace to obtain a cross-modal weighted fusion representation. The fusion result is then summed with its corresponding input, and layer normalization is applied to stabilize training and convergence. To further improve efficiency, mixed precision is enabled during training (float16 is used for the main calculations, and key statistics are retained for higher precision), and XLA graph-level optimization is enabled by wrapping the core calculations in the JIT compilation context of tf.function.

[0099] After parallel integration, the fused output features are first subjected to "global average pooling," which means averaging each channel across the spatial dimension to obtain channel-level semantics and suppress overfitting; then, they are fed into a fully connected layer that maps from 2048 dimensions to 1024 dimensions and activated using ReLU, resulting in a fused feature vector of dimension 1024; "firstly..." Perform global average pooling, then pass it through a fully connected layer (ReLU) from 2048 to 1024 to obtain the vector. During the generation process, the vector is simultaneously embedded with road network semantic annotations (such as road segment IDs and type labels) to ensure consistency between tracing and subsequent comparisons. Since the vector aligns the spatial boundary information of the image side with the topological connection relationship of the vector road network in the same representation space, it can more completely depict the multimodal nature, especially in complex scenarios such as intersections, thereby improving the accuracy and robustness of subsequent width calculation and quality verification.

[0100] S4: Calculate road width parameters, process the fused feature vector through edge extraction algorithms and cluster analysis, quantify the road segment width value, and embed it into the road network topology. The specific implementation is as follows:

[0101] The process of calculating road width parameters first involves preparing the fused feature vector and initializing the environment. The fused feature vector obtained from step S3 integrates spatial, temporal, and topological elements, with a vector dimension of 1024. To balance throughput and memory safety, the input batch size is set to 128. Based on GPU memory stress test results, the memory overflow rate exceeds 5% when the batch size exceeds 256, so the algorithm is run below this threshold to achieve a steady-state load. Algorithm initialization uses OpenCV and Scikit-learn to configure the image and clustering environments. Edge extraction uses an improved Canny process, specifically using a 5x5 height... The kernel performs smoothing and uses an adaptive threshold to determine the gradient threshold; density clustering uses DBSCAN with a radius parameter set to 0.5m and a minimum sample size set to 5 to match the minimum resolvable scale of road geometry and the sampling noise level; on the topology side, a GIS vectorized road network map is preloaded and modeled using NetworkX, with a node scale exceeding one thousand and an edge scale exceeding two thousand. Node attributes include structured fields such as coordinates and type to support subsequent connectivity verification and path constraints; to ensure cross-component compatibility, vectors are uniformly stored in NumPy float32 array format, and GPU resources are pre-allocated to the NVIDIA RTX 3090 at the hardware level to immediately enter the computation state after data loading; providing a stable and scalable operating foundation for subsequent computation and real-time verification of large-scale road network scenarios such as highways;

[0102] In the edge extraction algorithm application stage, an improved Canny framework is adopted for the fused road network feature vectors: first, Gaussian smoothing is applied to the image to suppress noise; then, gradients are calculated and non-maximum suppression is performed along the gradient direction; finally, double-threshold hysteresis connections are used to ensure boundary continuity. A multi-scale filtering scale-space representation is introduced, processed step-by-step through a Gaussian pyramid of three to five scale layers, where the smoothing intensity (standard deviation) increases from 1.0 to 4.0, determined based on road width distribution tests. When the standard deviation is less than 1.0, the loss of detail in narrow roads exceeds one-tenth, while when it is greater than 4.0, more than 50% blurring occurs in wide roads. At each scale layer, a 3×3 Sobel operator is used to estimate the gradients in the horizontal and vertical directions respectively (which can be understood as using a 3×3 operator that emphasizes "left minus right, center row double weight"). The weighted template calculates the horizontal change, and its transpose template calculates the vertical change. The gradient intensity of each pixel is obtained by squaring, summing, and then taking the square root. The gradient direction is given by the arctangent function (two-parameter form) with the vertical and horizontal components as independent variables, used for subsequent suppression by comparing neighborhood extrema along this direction. Multi-scale fusion employs a weighting strategy of "double decay around the optimal scale." The weight decreases rapidly with a square relationship as the current scale moves further from the optimal scale, reaching its maximum at the optimal scale. The optimal scale is determined by road width estimation, such as in typical urban roads. The second layer is taken as the empirical optimum; the dual thresholds are adaptively set: the lower threshold is 40% of the average gradient intensity of the entire image, and the higher threshold is 1.6 times the average value (experiments show that a fixed threshold will lead to a positioning error of more than five meters, while this adaptive setting can converge the error to within two meters); the hysteresis connection stage uses eight-neighborhood to track edges to ensure that there are no breaks or holes in the boundaries; the algorithm outputs a two-dimensional edge map with a height-to-width ratio, and the positioning accuracy of the boundary point set reaches the centimeter level; in curved road sections, the system will automatically bias towards smaller scale layers to preserve curvature details, thereby avoiding the arc breaks and details missed by standard Canny; in engineering implementation, combined with OpenCV GPU acceleration (such as CUDA backend) for parallel processing, the processing time of a 1024×1024 feature map can be controlled within one hundred milliseconds, taking into account multi-scale robustness and online processing efficiency, and meeting the needs of practical application scenarios with significant differences in road levels and frequent curvature changes;

[0103] After edge extraction is completed, the "density-based adaptive clustering - road width quantization" sub-step is started simultaneously: DBSCAN is used to perform density clustering on the boundary point set in the fused feature vector, with the core parameters being the neighborhood radius eps and the minimum number of samples. The initial setting was 0.5 meters for EPS (based on the average width of urban roads being approximately 5 meters; actual measurements showed that when EPS was less than 0.3 meters, over 20% excessive fragmentation occurred, hence the EPS was set to 0.5 meters). Based on noise comparison tests, when the minimum sample size was less than 5, the false positive rate exceeded 15%, therefore... Set to 10; to adapt to different road shapes, an adaptive mechanism is introduced: first calculate... For order The distance sequence is sorted in ascending order, and the derivative of this sequence is calculated. The "knee" corresponding to the peak of the derivative is used as the updated optimal EPS. Based on this, in narrow road scenarios (less than 3 meters), the EPS is automatically shrunk to about 0.3 meters, and in wide road scenarios (greater than 10 meters), the EPS is further reduced. The width is widened to approximately 1.0 meter. During clustering, Euclidean distance is used as the metric. Adjacent clusters are identified as left and right curb pairs. For any point in the left cluster, a corresponding point in the right cluster whose line of connection is approximately perpendicular is found (using the cosine of the angle as a filtering condition). The Euclidean distance between the two points is calculated as the instantaneous width at that point. The average of all matching pairs is then used to obtain the representative width of the road segment. To suppress abnormal deformation, the standard deviation of points in each cluster is calculated. If the standard deviation exceeds 0.2 meters, it is marked as a suspected pothole or construction disturbance point, and a local outlier factor (LOF, neighborhood size 20) is used for secondary detection to reduce false alarms. The output is a "width sequence containing m sampling points per road segment," along with both the mean and standard deviation statistics. In internal validation, the width estimation achieved centimeter-level accuracy (test error less than 0.05). For curved road sections, the system uses curvature as an adjustment factor: dynamically scaling eps based on the reciprocal of curvature (i.e., radius of curvature) to make the point density threshold at the curve closer to the real geometry. This allows the system to automatically complete distance and width statistics while marking the unique variation points of the curve. The overall accuracy of the curved scene is improved by about 15% compared to the straight line, and the end-to-end processing time is controlled within 1 second, which can support the rapid quantification and quality screening of a large-scale road network.

[0104] In the "clustering quantization-topology embedding" stage, the road network is modeled as an undirected graph. Nodes represent intersections and carry coordinate and attribute information, while edges represent road segments with their weights set to the sum of the segment's length and width. The width parameter obtained from clustering is then integrated as an edge attribute, with attributes including average width, standard deviation of width, and a list of mutation points for recording local anomalies. To achieve accurate correspondence between clusters and road segments, a spatial index-based matching mechanism is adopted. The index structure is an R-tree with a leaf capacity of 10, retrieving candidate road segments within a neighborhood radius of 0.1m and writing or updating their attributes. When updating the topology, corresponding node pairs are added or updated. The edge is mapped, and the width attribute is bound to the edge. Simultaneously, a connectivity check is performed. For isolated nodes with a degree of zero, connections are automatically repaired based on geometric constraints and road connectivity rules to ensure topological consistency after embedding. For easy and intuitive verification, the attribute graph can be displayed through a visualization component, where node colors are mapped according to the variance of the width. The output of this stage is an enhanced road network map, saved as a persistent binary file. Later, the average width and other attribute values ​​of road segments between any two nodes can be directly read through the edge data access interface. Furthermore, structured storage and compression strategies reduce the overall storage size by more than 50%, providing efficient and searchable basic data for subsequent quality verification stages.

[0105] After generating the width parameters, the spatially ordered "segment width value sequence" and the "embedded representation of the road network topology" (i.e., the vectorized description obtained after encoding nodes, edges and their connectivity, loop features, etc.) are passed as input to the S5 multi-scale verification module for consistency verification at both macro and micro levels, and are compared item by item with the preset deviation threshold. The Gaussian convolution kernel for edge detection is set to a "7x7" window size to enhance robustness to fuzzy boundaries. For future upgrades, a graph convolutional network (GCN) can be introduced to perform end-to-end learning of the topology embedding, further improving the joint representation capability of "geometric width - structural connectivity - attribute semantics". In terms of engineering deployment, it can be connected to cloud computing resources to complete the output of width and topology of a batch of road segments through distributed parallel scheduling, with the processing time of a single session controlled within "less than two seconds", thus providing S5's multi-scale verification with input that is both timely and accurate.

[0106] S5: Perform quality verification and evaluation, using multi-scale consistency indicators to verify the calculation results. If the deviation exceeds a preset threshold, trigger backtracking optimization of the fusion weights. The specific implementation is as follows:

[0107] The quality verification and evaluation process involves receiving a quantized road width parameter dataset from S4. This dataset includes: the width value of each road segment (in floating-point format in meters), the embedded road network topology (including node connection matrices and boundary coordinate sequences), and labels associated with the samples (such as road segment ID, calculation source, and other metadata). The data size is set according to a single road network region. The input is in the form of a structured table with clearly defined column fields. "As a unique identifier," "Record the road segment width value," "Save a list of nodes at both ends or adjacent nodes of this road segment," "Save the polygon coordinate sequence of the road segment or its boundary; for stable operation of subsequent index calculations and topology analysis, the preparation stage involves completing the environment readiness on a high-performance computing node, running Python 3.x, and loading NetworkX for graph structure and connectivity analysis, and Shapely for geometric objects and distance metrics; then initialize two types of verification thresholds: the first type is the width deviation threshold, set to ±0.5 meters; its engineering basis is the road network accuracy standard test, that is, when the threshold is below 0.5 meters, it is easy to have more than 20% false alarms, while when the threshold is above 1 meter, the false alarm rate may exceed 15%; the second type is the topology integrity threshold, requiring the overall connectivity to be higher than 0.95 and the boundary closure error to be lower than 0.1 meters; when using it, the deviation of the road segment is calculated by the width obtained in step S4 and the value from the GIS." The absolute difference between the benchmark data or the measured width with an accuracy better than 0.1 meters; the macro-scale "average deviation" is: summing the deviations of all road segments and then dividing by the total number of road segments to obtain the average deviation of the entire network; "connectivity" is: adding the degrees of all nodes in the entire network and then dividing the sum by twice the number of edges to obtain the theoretical connectivity index, which is then compared with the expected connectivity relationship calculated based on the degree distribution; "closure error" is: for each boundary polygon, calculating the distance between its first and last coordinates, and taking the maximum value of this distance among all boundaries as the closure error; to meet the timeliness of the project, the parallel acceleration in the preparation stage adopts a multi-process scheme, and the overall processing time can be calculated according to the number of road segments. Its logarithmic complexity is estimated. With 1000 road segments, initialization and data loading can usually be completed within 1 second. At the same time, in typical scenarios such as urban road networks, the startup wait time can be compressed to within 500 milliseconds, thereby avoiding the startup delay of traditional manual verification and providing a stable and traceable input and operating environment for subsequent multi-scale index calculations.

[0108] In the multi-scale indicator construction phase, a set of verification indicator systems is established based on the multi-level characteristics of the road network. A "macro-micro" dual-scale structure is adopted for consistency measurement. The macro-scale measures the entire road network, statistically analyzing the average width consistency; the micro-scale focuses on the local deviations of individual road segments. In the specific construction process, the width deviation is first calculated for each road segment, where the "single-segment deviation" is calculated as: the calculated width of the road segment output from the S4 stage and the width from the GIS. The reference widths, whether from benchmarks or actual field measurements, are subtracted and their absolute values ​​are taken. Then, macroscopic aggregation is performed to obtain the "average deviation," which is the sum of the individual deviations of all road segments divided by the total number of road segments. At the microscopic level, the individual deviation of each road segment is checked one by one to determine if it exceeds the 0.5m threshold, used to detect local anomalies at bends, etc. To ensure the correctness of the topology, a topology integrity index is defined, where "connectivity" is "the ratio of the actual number of connected edges to the expected number of edges derived from the degree distribution." The expected number of edges is given by the degree sum theorem in graph theory, namely, "the sum of the degrees of all nodes equals twice the number of edges," which allows the degree distribution to be converted into the expected number of edges. Regarding boundary quality, "closure error" is defined, taking the separation distance between the start and end points of the coordinate sequence of each boundary polygon, and calculating this separation distance using line string distance calculations from a geometry library. In engineering implementation, the index construction adopts a parallel processing mechanism, using four working processes to calculate road segments in parallel. The overall processing time increases with the number of road segments by multiplying the logarithm N. With the increase in scale, when the number of road segments is one thousand, the overall calculation time can be controlled within one second; macro indicators are used for rapid physical examination of overall consistency, while micro indicators focus on fine anomaly location of single segments. The two complement each other, ensuring both coverage and meeting the requirements of timeliness and stability.

[0109] After constructing the indicators, the accuracy of the calculation results is verified using preset thresholds: for each road segment, the deviation is first defined in words. Its value is the output of S4. Compared with reference benchmark The absolute value of the difference between them ( (Taken from GIS benchmarks or actual measurements with an accuracy of less than 0.1 m), and then... Compared with an engineering threshold of ±0.5 m, the positive deviation... Interpreted as "overestimation" (calculation results are too broad, often due to noise-induced boundary expansion), negative bias This is interpreted as "underestimation" (the calculated result is too narrow, often due to occlusion causing edge loss); the specific judgment logic is: if If the absolute value is greater than 0.5m, then the road segment is incorrectly marked. Set to "Yes" otherwise set to "No"; at the macro level, this invention further performs statistical analysis on the overall deviation distribution and calculates the standard deviation of the deviation. Its meaning is "to give each Deviation from the overall average The difference is squared, summed, averaged by the number of road segments, and then the square root is taken to obtain the quantity. Based on this, an overall risk threshold is set, i.e., when... When the value is greater than 0.3 m, the entire network will be marked as "high risk"; to enhance the interpretability of review and traceability, the inspection process integrates anomaly classification and simultaneously records causal clues: " Cases where the result is positive and exceeds the threshold are categorized as "overestimation" and accompanied by warnings about noise, artifacts, etc. The case of "negative and exceeding the threshold" is classified as "underestimation" and the possible sources such as occlusion and missing edges are indicated. In terms of implementation path, vectorized batch processing with NumPy arrays is adopted to complete the above segment-by-segment verification, overall statistics and classification labeling in the same batch, which is significantly faster than the pure Python implementation of looping one by one. In engineering practice, the time to process N=1000 road segments can be stably controlled in the hundreds of milliseconds. It should be noted that 0.5m and 0.3m are engineering thresholds set under the ISO 19157 data quality framework in this embodiment. Therefore, the thresholds can be parameterized and adjusted according to the project level or scenario to maintain a consistent evaluation caliber while taking into account the differences in different road networks and measurement conditions.

[0110] Following the width check, the topology integrity verification phase begins. This phase verifies the logical consistency of the road network structure along the main lines of "connectivity—closure—layered verification—anomaly self-repair," preventing global misjudgments caused by width calculations deviating from the topology. First, connectivity verification is performed, constructing the road network graph G, and then utilizing NetworkX's... The number of connected components is counted to determine whether structural breaks exist; when the number of components is greater than 1, the connectivity is considered to be less than 1, and this is indicated by "...". The overall connectivity level is measured using a method based on literature topology testing (joint connectivity and closure rate), verified through simulation (local loss rate <5%) and standard specifications, with 0.95 as the engineering threshold (a disconnection rate exceeding 5% triggers an alarm when the rate is below this threshold). A "weak connectivity" test is then performed on the entire map, equivalent to ignoring edge directions to determine overall connectivity, used to quickly eliminate macroscopic structural breaks. Closure error verification is then performed. For each boundary polygon, Shapely's geometric validity check is used to determine if it is a legal closed geometry. If not closed, the Euclidean distance between the first and last coordinates of the outer ring is used as the reconstruction error and calculated using Shapely's distance calculation, with a threshold of 0.1m (exceeding the threshold is considered a significant closure deviation). The verification process adopts a layered strategy: at the macro level, weak connectivity testing is performed on the entire G; at the micro level, fine connectivity and closure checks are performed on key subnets such as intersections. When topological inconsistencies (such as self-intersections, hanging edges, or minor cracks) are found, the system triggers an automatic repair process, prioritizing geometric purification through small buffers, and, if necessary, combining Shapely's... The system is capable of generating legal geometry. In terms of algorithm implementation, the traversal and connectivity determination adopt a graph search strategy that is linearly related to the number of vertices and edges. The time complexity is proportional to the number of nodes plus the number of edges. With a scale of about 2000 edges, the verification can be completed in the order of hundreds of milliseconds. Therefore, when verifying the width of an intersection, once the closure error reaches 0.2m, the system marks the anomaly and backtracks to the data and topology reconstruction stage to ensure that the consistency of the macro structure and the accuracy of the micro local precision are controlled at the same time, thereby avoiding the amplification of isolated deviations in width calculation in subsequent processes.

[0111] During the verification process, a backtracking mechanism is triggered once a deviation exceeds a preset threshold. A backtracking condition is established when the absolute value of the width deviation of any road segment is greater than 0.5 meters, the connectivity index is lower than 0.95, or the boundary closure error is greater than 0.1 meters. After triggering, the system collects the corresponding erroneous road segment data and feeds it back to the S3 multimodal network, adaptively adjusting the attention weights in the fusion process. For example, when the deviation is mainly determined to be due to remote sensing noise, the spatial weight is reduced; specifically, the "new spatial weight" is set to " The learning rate is set to 0.1, and the error ratio is equal to the ratio of the deviation of the road segment to the threshold. Then, gradient descent is used to optimize the weights. The optimizer is Adam, the learning rate is 0.001, and the fusion vector is recalculated in small steps of up to 5 rounds. A "limited number of times" mechanism is set for backtracking to avoid oscillation and loop. A single sample can be backtracked a maximum of 3 times, and the optimization log is fully recorded (including the trajectory of weight changes in each round and the triggering reason). In practice, when the width deviation of a certain intersection is verified and the topology is inconsistent, the system will automatically backtrack and recalculate the various weights of S3, so that the weights related to space and topology are increased by 0.1, thereby improving the overall quality and ensuring efficiency. In typical scenarios, the time taken for a single backtracking is less than 2 seconds, which can support real-time road network applications. After the backtracking is completed, a traceable verification output is generated and connected to the downstream. The verification result of each road segment is encapsulated into JSON report data. The "pass" result enters S6 for iterative update, and the "fail" result triggers a full-link warning and retains the backtracking record for auditing and review.

[0112] S6: Output optimization results and iteratively update them, adjust fusion parameters based on verification feedback to form a closed-loop mechanism, and support the dynamic application of road network data. Specific implementation details are as follows:

[0113] The system first completes the integrated readiness of "input configuration for verification feedback - initialization of the optimization module": specifically, it receives the quality verification evaluation results from step S5 and loads the multi-scale consistency indicators as structured text; for example, it loads " (Width deviation) is recorded as 0.3. The "(Topological Integrity Score)" is recorded as 0.92, and in the " The feedback dataset is recorded item by item in the "(Feedback Point Set)" to ensure semantic consistency of cross-module parameter passing. Subsequently, the size of the feedback dataset is set according to the road network coverage, with approximately 1000 feedback records configured in a 50 square kilometer urban area. The input batch size is set to 30, based on empirical conclusions from memory optimization tests. When the batch size exceeds 64, the overflow rate exceeds 5%, aiming to achieve a balance between throughput and stability. The optimization module is initialized using a PyTorch 1.12 environment, configured with the AdamW optimizer, and the learning rate is set to "". The weight decay is set to "1%", a value derived from gradient descent control tests (when the learning rate is below a certain threshold). "When the convergence time is extended by more than 20%, the initial state of the fusion parameters is preloaded, such as the attention weight matrix from S3, with a dimension of 512 by 512; at the same time, a trigger threshold is set for subsequent key verification and local recalculation, the deviation threshold is 0.5m (according to navigation accuracy requirements, when the deviation exceeds 0.5m, the navigation error increases by more than 10%), and the integrity threshold is 0.95 (according to topology connectivity test, when this threshold is lower than 0.9, the connectivity loss ratio exceeds 5%); on the engineering time efficiency side, the startup time is reduced to less than 0.5 seconds through a lightweight initialization path, thereby avoiding the response lag caused by traditional static output links, laying a stable starting point for high-quality input and adjustable parameters for subsequent iterative updates of large-scale road networks such as highways;

[0114] Upon entering the optimization result generation stage, based on the feedback from S5, the adjusted road segment width is directly generated and input for the next iteration is prepared. Specifically, the quality verification results are merged line by line with the quantized width vector output by S4 (fields are "Road Segment ID", "Width Value", and "Confidence Score") to form a vector with "...". (Optimized width) and " The GeoJSON optimized output file contains attributes such as "(verification score)". The optimized width is calculated as "optimized width equals original width multiplied by (one minus deviation ratio) plus correction term", where "deviation ratio equals actual deviation divided by threshold" and "correction term is obtained by averaging the deviation information of feedback points and multiplying by a weight of 0.2". The output uses the OGR library to write to the vector layer to ensure read and write compatibility with mainstream GIS, and the coordinates are defined according to CGCS2000 projection (EPSG:4490) to meet the coordinate consistency requirements of the engineering side. In high-sensitivity scenarios such as urban curved road networks, it can automatically correct road segments with deviations greater than 0.3m and generate an optimized dataset containing about one thousand road segments in one batch processing. At the same time, it retains the merged vectorized input for subsequent iterations to achieve a closed-loop update of "verification - correction - re-verification".

[0115] After generating the results, the fusion parameters are adjusted based on the feedback returned in step S5. The fusion parameters in step S3 are dynamically optimized as follows: First, the feedback bias vector and integrity score are concatenated sequentially and input into a small multilayer perceptron (three layers, each with 256 hidden units and ReLU activation function). This network outputs the attention weight increment for this round. Then, the increment is scaled by the learning rate multiplied by 1 to the power of 10 (negative fourth power) to obtain the change in attention weights used for updating. To suppress unstable updates, gradient clipping is applied to all trainable parameters, constraining their gradient norm 2 to not exceed 1.0. The updated weights are then bounded according to business constraints: when feedback indicates high occlusion in the remote sensing space... When the remote sensing spatial weight is lowered by 0.1 within its legal range (the allowable weight range is 0.3-0.5), when the feedback indicates that the trajectory temporal change is large, the trajectory temporal weight is increased by 0.2 within its legal range (the allowable weight range is 0.4-0.6). The GIS topology weight is kept stable in the range of 0.2-0.4 (according to stability tests, a variation of less than 0.05 can accelerate convergence). The above adjustments are performed in small iterations of 1-3 rounds until the weight change between two adjacent rounds is less than 0.01, which is considered convergence. In a typical intersection road network scenario, if the feedback deviation increases significantly, the system will automatically increase the weight of the point cloud temporal channel to greater than 0.5, thereby strengthening the characterization of dynamic elements and improving the quality of the final fused vector.

[0116] After parameter adjustment, the closed-loop mechanism construction phase begins. A feedback loop of "monitoring—triggering—backtracking—verification" is used to achieve self-optimization and continuously improve the accuracy of width calculation. Specifically, output indicators are tracked in real-time at fixed intervals, and triggers determine whether to backtrack and recalculate based on threshold rules. When the width deviation of any road segment exceeds 0.5 meters, backtracking to S3 for re-fusion is immediately initiated. The entire loop uses the comprehensive quality score as the convergence criterion; that is, as long as the comprehensive quality score remains below 0.95, parameters are updated based on the latest feedback, and the next iteration begins. The comprehensive quality score is defined in two categories. The weighted sum of indicators, with "deviation indicators accounting for 60% of the weight and integrity indicators accounting for 40% of the weight," condenses the performance of different dimensions into a single comparable indicator through this weighted composite score. In terms of engineering organization, the closed loop adopts a finite state machine design, with states including "output," "feedback," "adjustment," and "verification." Threshold-driven conditions serve as triggers for state transitions, enabling the mechanism to automatically complete two rounds of iterations and adjust the weights accordingly to adapt to the new data when dynamic changes such as construction are detected. This forms a closed-loop control logic of "self-adjustment upon feedback" and maintains the system's stable operation within the target range.

[0117] Once the closed-loop system is built, it supports the dynamic application of road network data in a "plug-and-play" manner. Optimization results are output to a GIS platform (such as ArcGIS or QGIS) via a plugin, in Shapefile or GeoJSON format, and a RESTful interface is provided for external systems to access the data on demand. The application side uses WebSocket for push notifications, outputting the latest width data to the navigation application every five minutes within the city road network update cycle, with the cycle set based on the traffic flow change rate. This frequency is maintained when the change rate exceeds 0.1 meters per second during peak hours. Navigation route recalculation uses A*. An engineered variant of the algorithm incorporates the principle that "path cost equals the sum of path length and width constraint penalties" (where the penalty is calculated by multiplying the width deviation by the penalty coefficient) into the cost model to continuously correct the route. In terms of compatibility, the output is compatible with both mobile devices and cloud services (e.g., using low-latency object storage to achieve sub-second access and keeping end-to-end access latency within 200 milliseconds). By verifying lane width and road geometry in real time, it assists vehicles in robust lateral control and predictive hazard avoidance, thus achieving seamless integration with lane centering control.

[0118] In actual deployment, the process adopts a cloud-edge collaborative architecture for iterative operation. Edge nodes are responsible for on-site data collection and rapid preprocessing, while the cloud performs parameter updates and global verification. Delivered in a containerized manner, the process is executed on an hourly basis to meet the high-frequency update needs of urban road networks. In each iteration, the fused parameters start with random initial values ​​evenly distributed between 0 and 1, with a target accuracy set at 98%. Within the update cycle of a typical urban road network, the accuracy can be improved from approximately 85% to approximately 98% through 3-5 iterations, corresponding to converging the initial error of approximately 5 meters to less than 0.5 meters. To ensure availability and horizontal scalability, the deployment supports a Kubernetes-based distributed cluster (example: 4 nodes) and a highly available control plane topology. The front-end uses Nginx for load balancing to distribute traffic, perform health checks, and handle failover, thus supporting a global runtime share exceeding 99.9%. The goal is high availability; in remote or weak network scenarios, the "edge-first" mode is enabled, which moves the near-end computing forward and only sends back the necessary digest, thereby reducing the round-trip network latency to less than 100 milliseconds; the deployment logic takes into account the portability of containerization, the elasticity and high availability of Kubernetes, the traffic management capabilities of Nginx, and the engineering advantages of edge computing in reducing backhaul bandwidth and end-to-end latency, so that iterative convergence, system availability and end-to-end real-time performance are synergistic.

[0119] It has strong applicability and portability. In lightweight scenarios such as rural road networks, the number of iteration rounds can be adjusted to 2 rounds to reduce computational overhead. In terms of process connection, the fusion vector is directly fed back from S3 to S6, forming a closed loop with "feedback-correction-re-optimization" as the core and improving end-to-end efficiency. At the same time, it fully considers the handling of edge cases. When the feedback is abnormally high (i.e. the deviation is greater than one meter), it automatically switches to the backup parameter set (pre-trained weights) to perform a rollback to ensure business continuity. The timestamp, deviation value and parameter adjustment details are recorded in JSON structure for easy subsequent auditing and traceability.

[0120] The scheme in this embodiment first collects and pre-standardizes multi-source road network data, including remote sensing imagery, vehicle-mounted LiDAR point clouds, and GIS vectors, completing coordinate unification, metadata annotation, and quality gating to generate a homogenized dataset that can be directly used for fusion. Then, deep learning is used for adaptive noise filtering and occlusion detection, combined with U-Net semantic segmentation and adversarial completion to output clean data. Based on this, a multi-source fusion framework is constructed, extracting imagery, point clouds, and topology through convolutional / PointNet++ / GraphSAGE branches, and then using an attention mechanism to calculate dynamic fusion weights. GPU parallel processing yields a unified fusion vector representing the geometric and structural semantics of road segments. Subsequently, road segment width is quantified by improving Canny edge extraction and DBSCAN density clustering, and the width attribute is accurately represented by an R-tree spatial index. The process involves accurately writing the loop network diagram, completing topology embedding and connectivity checks, and then performing quality assessments using multi-scale consistency indicators. At the macro level, the average deviation across the entire network is statistically analyzed, while at the micro level, thresholds are checked segment by segment. Simultaneously, connectivity and boundary closure errors are calculated in parallel to accelerate the output of verification conclusions. If the deviation or topology indicators fail to meet the standards, a backtracking process is triggered. Finally, based on feedback, a hot-start optimization is performed in the PyTorch / AdamW environment, dynamically adjusting the fusion weights to form a closed loop of "output—feedback—adjustment—verification." The optimization results are published externally in GeoJSON / Shapefile and JSON reports, supporting cloud-edge collaboration and lightweight, rapid initialization. In the event of extreme deviations, the system automatically reverts to a backup parameter set and records audit logs, thereby achieving high-precision, low-latency, traceable width calculation and continuous updates in road networks of different scales, such as urban and rural areas.

[0121] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0123] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0124] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0125] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] In addition, the functional modules in the embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0127] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0129] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatic calculation and quality check of road width of road network data, characterized in that, Comprise: S1: Collect multi-source road network data and perform pre-standardization processing, map and metadata annotation in a unified coordinate system, convert remote sensing images, vehicle trajectory point clouds and GIS vector data into compatible formats, and generate labeled original data sets; S2: Perform noise adaptive filtering and occlusion detection, use deep learning models to identify noise interference and occlusion areas, and apply generative adversarial networks for local completion, output the purified data set; S3: Build an efficient multi-source fusion framework, use a multi-modal network with an attention mechanism to calculate dynamic fusion weights, input the purified data set into a parallel computing module, and form a fusion feature vector; S4: Calculate the road width parameter, process the fusion feature vector through edge extraction algorithm and cluster analysis, quantify the road width value, and embed the road network topology structure; S5: Quality check and evaluation, use multi-scale consistency indicators to verify the calculation results, if the deviation exceeds the preset threshold, trigger backtracking optimization fusion weight; S6: Output the optimization results and iterative update, adjust the fusion parameters according to the feedback of the check, form a closed loop mechanism, and support the dynamic application of road network data.

2. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, Collect multi-source road network data and perform pre-standardization processing, map and metadata annotation in a unified coordinate system, convert remote sensing images, vehicle trajectory point clouds and GIS vector data into compatible formats, and generate labeled original data sets, including: Unified coordinate reference for mapping different source data to the same spatial reference system and aligning scale, resolution and timestamp; Metadata annotation includes source device or service identification, collection time, spatial coverage, resolution description and data quality identification; Perform quality gating before data is written to disk, quality gating limits include: automatic detection of clarity and contrast judgment indicators on the image side, investigation of trajectory integrity and point density anomalies on the point cloud side, topology effectiveness and field consistency test on the vector side, and trigger resampling or mark as "gray input" for review when any link fails to meet the standards; Divide the responsibilities between the edge node and the cloud: the edge node undertakes on-site collection and format unification, and the cloud undertakes batch review and master data registration.

3. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, Perform noise adaptive filtering and occlusion detection, use deep learning models to identify noise interference and occlusion areas, and apply generative adversarial networks for local completion, output the purified data set, including: Establish a "problem area annotation-local reconstruction-quality feedback" purification process, output the purified data set while carrying the spatial range, source type and confidence of the problem area, and provide binary or multi-value masks in the same resolution / same coordinate as the original data; Problem area annotation is used to identify cloud, shadow, vegetation, building occlusion, GPS drift and abnormal sparse typical interference; Local reconstruction is used to complete the missing information while preserving the boundary geometry and texture / structure continuity; Quality feedback is used to write the spatial range, confidence and quality score of the reconstructed area into the structured log of the purified data set and establish a one-to-one correspondence with the corresponding source data; When the purification quality score is lower than the set threshold or the problem area ratio is over limit, trigger backtracking identification or resampling prompt.

4. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, A high-efficiency multi-source fusion framework is constructed, a multi-modal network with an attention mechanism is used to calculate dynamic fusion weights, a purified data set is input into a parallel computing module to form a fusion feature vector, including: Parallel processing paths of image branch, point cloud branch and vector topology branch are established in a mode of shunting by mode and adaptive allocation of weight, and in the fusion layer, the features of each branch are weighted, spliced and uniformly represented according to dynamic weights; The image branch is used to extract spatial edges and texture semantics; The point cloud branch is used to represent local geometry and density changes; The vector topology branch is used to encode the connectivity and hierarchical attributes of road segments and intersections; The fusion output is organized as a "fusion feature vector carrying road segment identification", which is the only input object for subsequent width calculation and is accompanied by source mapping; The fusion framework supports the collaborative scheduling of batch processing and parallel computing in resource-constrained scenarios, allowing parameterized configuration of batches, caches and parallelism according to throughput and memory constraints.

5. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, The road width parameter is calculated, the fusion feature vector is processed by edge extraction algorithm and cluster analysis, the road width value is quantified, and the road network topology structure is embedded, including: Edge extraction is performed on the fusion features to obtain road boundary point sets, which are clustered into pairs of curb clusters according to density and geometric consistency, and the spatial proximity and normal angle consistency between curb pairs are used as constraints for width calculation to obtain a width sequence along the road segment; Based on spatial indexing, the width attribute is accurately matched with the road segment and written back to the road network graph as an edge attribute, while completing structural repair of isolated nodes, hanging edges and obviously unclosed boundaries, outlier labeling and recording of abnormal fluctuating width, and local geometric fidelity strategy for curves / complex intersections, ensuring the consistency of width parameters and topology structure; The output topology-enhanced graph includes the representative width of each edge, width fluctuation statistics and a list of abnormal points.

6. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, Quality check and evaluation are performed, and multi-scale consistency indicators are used to verify the calculation results. If the deviation exceeds the preset threshold, the fusion weight is triggered for backtracking optimization, including: A multi-scale consistency index system combining macro and micro is established, supplemented by topology integrity index and boundary closure index, supporting the simultaneous identification of whole network and single segment problems; Macro consistency is used to statistically analyze the average deviation of the whole network and give the overall risk threshold; Micro consistency is used to check whether the deviation exceeds the limit for each segment and locate local anomalies in curves; Topology integrity is used to measure whether the connected components and connectivity meet the engineering constraints; Boundary closure is used to measure whether the spatial separation of the first and last points of the boundary polygon is acceptable; When any indicator exceeds the limit, the backtracking process is triggered in the order of "local first, then global, and then backtracking", and the over-limit road segment, problem type and trigger threshold are recorded in the structured feedback to support subsequent parameter optimization; The index system allows parallel calculation and maintains controllable computational complexity when the scale expands.

7. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, The optimization results are output and iteratively updated, the fusion parameters are adjusted according to the check feedback, a closed-loop mechanism is formed, and dynamic application of road network data is supported, including: The check feedback is structured data that follows a closed loop of "initialization - parameter fine-tuning - local recalculation - result review". Initialization is used to load the last round of fusion parameters and adjustable thresholds; Parameter fine-tuning is used to adjust the processing threshold of cross-modal weight between edge and cloud according to the principle of minimum necessity; Local recalculation is used to regenerate fusion features and width results within the scope of problem sections; Result review is used to determine whether the convergence threshold is reached and whether to publish; When there are still out-of-limit sections after optimization, rollback to the backup parameter set is allowed to ensure task continuity, and the rollback reason, impact range and recovery strategy are included in the audit log; The output is provided in the form of GIS-friendly vector data and in the form of structured report with version, threshold and difference explanation.

8. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, Including: Logical fragmentation of edge nodes according to geographical areas, plants or road levels, and maintenance of routing and health status; When new or removed nodes are added, seamless scaling is achieved through task redistribution and state migration, and mapping context and data consistency are maintained during migration; When a single point failure or network isolation occurs, adjacent nodes temporarily take over the affected area's collection and calculation tasks according to the preset strategy, and after recovery, perform bidirectional reconciliation and difference merging; During cross-regional coordination, attribute consistency of the same road feature in different domains is ensured through master data synchronization and version arrangement, and the impact range and effective order of cross-domain corrections are controlled and managed.

9. The method for automatic calculation and quality check of road width of road network data according to claim 1, characterized in that, Including: Record the time, subject and associated data summary of key events such as width calculation, quality evaluation, parameter adjustment, link switching and rollback in edge and cloud respectively, and collect them in a unified log model; To meet the needs of higher level compliance or multi-subject coordination, the log model can be connected to a tamper-proof storage with a trusted time source to solidify key change records and support cross-domain verification; When a dispute or security incident occurs, quickly reconstruct the event sequence, replay version evolution and locate the responsibility according to the log chain.

Images