A digital-twin-based intelligent city road management method

Abstract

The present application belongs to the field of three-dimensional model construction, and particularly relates to a road management method for smart city based on digital twinning. It aims to solve the problem that traditional modeling relies on two-dimensional or limited cross-section data, lacks geometric precision and is difficult to dynamically reflect road correlation, resulting in a large deviation in traffic management strategy optimization. The steps are as follows: a high-precision laser scanner is used to obtain point clouds, and rough error removal and curvature adaptive thinning preprocessing are performed; point cloud integration and registration (selecting reference features, calculating descriptor similarity matching, and accumulating cumulative error through least squares, graph optimization and Levenberg-Marquardt algorithm); constructing a digital twinning model of urban roads; and finally simulating signal timing and lane control based on the model, and outputting the optimal strategy through multi-objective optimization. The present application improves the fidelity of the model, effectively alleviates congestion and improves traffic efficiency.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional model construction, and in particular relates to a road management method for smart cities based on digital twins. Background Technology

[0002] With the continued acceleration of urbanization, urban road networks are becoming increasingly complex. The surge in motor vehicle ownership has led to frequent problems such as traffic congestion, low traffic efficiency, and delayed control and response. Urban road management is no longer solely reliant on traditional traffic monitoring and manual dispatching, but is gradually shifting towards data-driven and model-driven approaches. In this process, accurately constructing digital twin models of urban roads has become a core issue. Existing modeling methods mostly rely on two-dimensional geographic information or limited road cross-section data. This approach not only lacks geometric accuracy but also struggles to dynamically reflect the complex relationship between road structure and environmental factors, resulting in significant deviations in traffic flow simulation and management strategy optimization. Summary of the Invention

[0003] In view of the technical problems existing in the above-mentioned background art, the present invention proposes a road management method for smart cities based on digital twins.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Deploy multiple high-precision 3D laser scanners on the road to obtain laser point cloud data, and preprocess the obtained laser point cloud data to obtain preprocessed laser point cloud data.

[0006] S2. Integrate and register the preprocessed laser point cloud data obtained from multiple scanners to obtain the road laser point cloud data of the entire city.

[0007] S3. Construct a digital twin model of urban roads based on the laser point cloud data of roads throughout the city;

[0008] S4. Optimize traffic flow using a digital twin model; the traffic flow optimization is to obtain the traffic volume, speed and queue length of each road segment based on the digital twin model, dynamically simulate different traffic light timings and lane control strategies, and output the optimal control strategy.

[0009] The implementation of step S2, which involves integrating and registering the preprocessed laser point cloud data obtained from multiple scanners to obtain the road laser point cloud data for the entire city, includes:

[0010] S21. Select reference landmark features from the point cloud and extract the geometric feature descriptor for each reference feature;

[0011] S22. For any two scanner point clouds, calculate the similarity using geometric feature descriptors to match the baseline features; if the similarity value reaches a set threshold, it is determined to be a match. ,in These are geometric feature descriptors acquired by any two scanners;

[0012] S23. Based on the matched baseline feature pairs, solve the initial transformation matrix using least squares. Treat all scanner points as graph nodes and transform the matrix. A registration graph model is constructed for the edges, and cumulative errors are eliminated through graph optimization. The objective function of the registration graph model is: ,in The weight of the edge. Let E be the final transformation matrix of different point clouds in the global coordinate system. The Levenberg-Marquardt algorithm is used to minimize the objective function E, thus solving for the globally consistent transformation matrix. Output the registered global point cloud .

[0013] Preferably, the preprocessing of laser point cloud data in step S1 includes coarse point cloud removal and point cloud data thinning.

[0014] Preferably, the coarse difference cloud removal includes:

[0015] S111, On the original laser point cloud Each point in Search first The K nearest neighbors of the given point are determined by the radius constraint. Removal distance Exceed The effective neighborhood is obtained by identifying the far neighboring points;

[0016] S112, For each point and its effective neighborhood, calculate The distance to the centroid of the neighborhood is denoted as the spatial distance feature, and is calculated... The angle between the normal vector and the average normal vector of its neighborhood is denoted as the normal vector consistency feature. Calculation... The relative deviation between the reflected intensity and the average intensity of the neighborhood is denoted as the reflected intensity deviation characteristic;

[0017] S113. The spatial distance feature, normal vector consistency feature and reflection intensity deviation feature are normalized and weighted by weighted summation to obtain a weighted anomaly score. An anomaly score threshold is set. If the obtained weighted anomaly score is greater than or equal to the set threshold, it is determined to be a coarse point cloud and is removed.

[0018] Preferably, the point cloud data thinning operation includes:

[0019] S121. Point cloud after gross error removal Each point in A quadratic surface is fitted using its neighborhood, and the equation of the quadratic surface is: ,in The fitting coefficients are obtained by solving the least squares method;

[0020] S122. Solve for curvature using the eigenvalues ​​of the Hessian matrix, and calculate the curvature of the quadratic surface. Hessian matrix at the location , ;

[0021] S123, to Eigenvalue decomposition yields two eigenvalues. Then calculate the local principal curvature. Where j=1,2, ;

[0022] S124. Take the larger principal curvature of the two as... The local curvature characterization is used to calculate the adaptive thinning radius. ,in, For adaptive thinning radius, Based on the thinning radius, The curvature sensitivity coefficient, For local curvature characterization;

[0023] S125. Finally, dynamic spherical neighborhood sampling is used to achieve uniform thinning, and the point set is initialized after thinning. Empty, unsampled point set ,from Randomly select a seed point to add Calculate the adaptive thinning radius of this seed point and delete... For all points whose distance from the seed point is less than or equal to the adaptive thinning radius, repeat the selection and judgment until... Empty, ultimately This refers to the point cloud data obtained after uniform thinning.

[0024] Preferably, in step S23, the initial transformation matrix is ​​solved by least squares. The specific implementation is as follows:

[0025] S231. Set the initial rotation matrix and translation vector as the iteration starting point, and preset the maximum number of iterations and the convergence threshold to terminate the iteration.

[0026] S232, Transformation parameters for the current iteration round Iterate through all benchmark feature matching pairs and calculate the spatial error residual for each matching pair. ,in, Let mn be the spatial error residual of the mn-th feature under the current change;

[0027] S233. Calculate the weight of each matching pair according to the Huber robust loss function, and for the residual... like Less than or equal to the set threshold Then weight ,otherwise ;

[0028] S234, Rotation matrix Parameterized into a rotation vector using the Rodriguez formula. Construct a system of linear equations: ,in Rotation vector The corresponding rotation matrix is ​​obtained by solving the linear least squares problem through matrix inversion, resulting in the updated rotation vector and translation vector, and then the rotation vector is converted into a rotation matrix.

[0029] S235. Finally, calculate the difference between the current transformation parameters and the previous round: ,in Let Frobenius norm be , if The iteration terminates when the value is less than the convergence threshold or the maximum number of iterations is reached, and the transformation matrix is ​​output. .

[0030] Preferably, the implementation of traffic flow optimization using a digital twin model in step S4 includes the following steps:

[0031] S41. First, extract the current traffic flow, vehicle speed, and queue length for each road segment using a digital twin model;

[0032] S42. Construct a congestion index by weighting traffic flow, vehicle speed and queue length, and use the congestion index as the objective function;

[0033] S43. A multi-objective optimization algorithm is used to solve the Pareto optimal solution. The traffic light timing scheme and lane strategy are encoded as chromosomes. The implementation effect of the scheme is simulated through a digital twin model. The objective function value is calculated, and high-quality individuals are selected based on the congestion distance.

[0034] S44. When the iteration reaches the preset number of iterations, output the optimal control strategy.

[0035] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This invention acquires point clouds using high-precision three-dimensional laser scanning, and improves data quality through gross error removal and curvature adaptive thinning preprocessing; the multi-scan point clouds are registered using graph optimization combined with the Levenberg-Marquardt algorithm to eliminate accumulated errors, and a high-fidelity digital twin model of urban roads is constructed, overcoming the problem of insufficient geometric accuracy of traditional two-dimensional or finite cross-section data; it also utilizes a multi-objective optimization algorithm to dynamically simulate traffic light timing and lane control based on the model, outputting the optimal strategy to effectively alleviate congestion and improve traffic efficiency. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the structural process of a road management method for smart cities based on digital twins. Detailed Implementation

[0038] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0039] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0040] In practice, existing modeling methods mainly rely on two-dimensional maps, road cross-section measurements, or partial surveillance video data. These methods are insufficient in representing road geometric features, traffic facility distribution, and dynamic operating environments. For example, two-dimensional information cannot reflect the spatial undulations and three-dimensional structure of roads. Traditional cross-section measurement methods acquire data with limited granularity and lack real-time performance, resulting in inaccurate models for traffic flow simulation, road load assessment, and emergency dispatch, making it difficult to meet the needs of smart city development. To achieve high-precision reconstruction of the three-dimensional road structure and realize high-fidelity mapping between the physical road entity and the virtual model, a road management method based on digital twins is proposed. This method uses three-dimensional laser point clouds as the data foundation. Through point cloud preprocessing and global registration, a high-precision digital twin model of the road is constructed. Then, combined with real-time traffic data, dynamic simulation and optimization control are achieved. The specific implementation process is as follows: Figure 1 As shown.

[0041] First, multiple high-precision 3D laser scanners are deployed on the road to obtain laser point cloud data. Although the raw point cloud data contains geometric information about the road and its surrounding environment, due to the complexity of the environment, it often contains noise points, occlusion errors, and data redundancy. Direct use of this data would lead to distorted modeling results. Therefore, the obtained laser point cloud data is preprocessed to obtain preprocessed laser point cloud data. The preprocessing of the laser point cloud data includes removing coarse point cloud data and thinning the point cloud data.

[0042] The gross point cloud removal includes the removal of the original laser point cloud. Each point in Search first The K nearest neighbors of the given point are determined by the radius constraint. Removal distance Exceed Find the far neighbor of each point to obtain the effective neighborhood; for each point and its effective neighborhood, calculate The distance to the centroid of the neighborhood is denoted as the spatial distance feature, and is calculated... The angle between the normal vector and the average normal vector of its neighborhood is denoted as the normal vector consistency feature. Calculation... The relative deviation between the reflection intensity and the average intensity of the neighborhood is denoted as the reflection intensity deviation feature. The spatial distance feature, normal vector consistency feature, and reflection intensity deviation feature are normalized and weighted by weighted summation to obtain a weighted anomaly score. An anomaly score threshold is set; if the obtained weighted anomaly score is greater than or equal to the set threshold, it is determined to be a coarse point cloud and removed. This scheme can effectively identify and remove isolated points, external interference points, and anomalies caused by local geometric abrupt changes, significantly improving the overall quality and geometric continuity of point cloud data.

[0043] Secondly, to avoid excessively dense point clouds leading to increased computational complexity, an adaptive thinning method based on curvature is employed. The point cloud data thinning operation includes: processing the point cloud after removing gross errors... Each point in A quadratic surface is fitted using its neighborhood, and the equation of the quadratic surface is: ,in The fitting coefficients are obtained by solving the least squares method; the curvature is solved by the eigenvalues ​​of the Hessian matrix, and the quadratic surface is calculated. Hessian matrix at the location , ;right Eigenvalue decomposition yields two eigenvalues. Then calculate the local principal curvature. Where j=1,2, Take the largest principal curvature of the two as... The local curvature characterization is used to calculate the adaptive thinning radius. ,in, For adaptive thinning radius, Based on the thinning radius, The curvature sensitivity coefficient, For local curvature representation; finally, dynamic spherical neighborhood sampling is used to achieve uniform thinning, and the thinned point set is initialized. Empty, unsampled point set ,from Randomly select a seed point to add Calculate the adaptive thinning radius of this seed point and delete... For all points whose distance from the seed point is less than or equal to the adaptive thinning radius, repeat the selection and judgment until... Empty, ultimately This refers to the point cloud data obtained after uniform thinning. This method effectively controls the point cloud size while preserving geometric feature details, reducing computational complexity, and maintaining the uniformity of the point cloud globally, thus providing high-quality input data for subsequent registration and digital twin modeling.

[0044] Then, the preprocessed laser point cloud data obtained from multiple scanners are integrated and registered to obtain the road laser point cloud data of the entire city. The implementation of integrating and registering to obtain the road laser point cloud data of the entire city includes:

[0045] This step involves selecting benchmark landmark features from the point cloud and extracting geometric feature descriptors for each benchmark feature. Specifically, to establish a reliable correspondence between point cloud data acquired by different scanners, this step first requires selecting benchmark landmark features from the point cloud data. Benchmark landmark features refer to feature points in an urban road environment that possess significant geometric characteristics and maintain stable visibility under different scanning angles and positions. This embodiment uses geometric feature descriptors to encode the benchmark points. Geometric feature descriptors, by statistically analyzing the distribution of points and their neighborhoods in terms of spatial location, normal vectors, and curvature, transform local geometric structures into fixed-length feature vectors, exhibiting rotation and scale invariance.

[0046] For any two scanner point clouds, a similarity score is calculated using geometric feature descriptors to match baseline features; if the similarity value reaches a set threshold, the pair is considered a match. ,in These are geometric feature descriptors acquired by any two scanners. Specifically, for any two sets of scanner point clouds, a set of descriptors for baseline feature points is extracted. The similarity between the two feature vectors is determined by calculating the Euclidean distance or cosine similarity. When the similarity value is higher than a set threshold, the pair of feature points is considered a matching pair. To avoid false matching, this embodiment introduces bidirectional nearest neighbor verification based on the similarity calculation; that is, a valid match is only considered when the best match for point A is point B and the best match for point B is also point A. In addition, the RANSAC (Random Sample Consensus) method can be used to remove outliers from the matching set, thereby ensuring the stability and accuracy of baseline feature matching. Through this scheme, a robust initial correspondence can be established between point clouds from different scanners, ensuring good convergence and accuracy in the subsequent registration process.

[0047] Based on the matched baseline feature pairs, the initial transformation matrix is ​​solved by least squares. Treat all scanner points as graph nodes and transform the matrix. A registration graph model is constructed for the edges, and cumulative errors are eliminated through graph optimization. The objective function of the registration graph model is: ,in The weight of the edge. Let E be the final transformation matrix of different point clouds in the global coordinate system. The Levenberg-Marquardt algorithm is used to minimize the objective function E, thus solving for the globally consistent transformation matrix. Output the registered global point cloud Where N is the total number of globally consistent transformation matrices. Specifically, after obtaining matching point pairs between point clouds from different scanners, initial registration is required, followed by global optimization to eliminate accumulated errors. Specifically, the initial rigid body transformation matrix is ​​first estimated using the least squares method based on the matched reference feature pairs. This process is solved by minimizing the Euclidean distance between corresponding point pairs, ensuring that the two point clouds fit as closely as possible during initial alignment. Since urban road scenes involve data collected by multiple scanners, relying solely on pairwise registration can easily lead to accumulated errors, causing misalignment or distortion in the overall model. Therefore, in this embodiment, all scanner point clouds are treated as nodes in a graph model, and the initial transformation matrices between different point clouds are treated as edges in the graph, thus constructing a point cloud registration graph. To achieve global consistency, this graph needs to be optimized. An optimization function is set, and by minimizing this objective function, all point clouds can achieve optimal consistency in the global coordinate system. In this embodiment, the Levenberg-Marquardt algorithm is used to perform nonlinear minimization iterations on the objective function. This algorithm combines the global stability of gradient descent with the fast convergence of Gauss-Newton's method, effectively avoiding local extremum traps and ensuring convergence speed. The specific implementation involves first representing the objective function as the sum of squared residuals of matching point pairs in the global coordinate system. This means that the coordinates of each matching point pair in different point clouds are aligned using the currently estimated transformation matrix, and then the positional differences are calculated. The residuals of all matching points are vectorized to form the overall residual function. Subsequently, a first-order Taylor expansion is performed on the residual function to obtain an approximate linear relationship, which is then used to construct the incremental update equation. The LM algorithm adaptively adjusts the damping factor λ between the Gauss-Newton method and the gradient descent method, balancing convergence speed and stability during iteration: when the current step direction is reliable, λ is decreased to accelerate convergence; when the descent is not significant or divergent, λ is increased to enhance robustness. In each iteration, the Jacobian matrix is ​​calculated and combined with λ to construct the normal equation. After solving for the parameter increment, the rotation matrix and translation vector are updated, gradually optimizing the global transformation parameters. Through multiple iterations, the algorithm stops when the decrease in the objective function is less than a set threshold or when the maximum number of iterations is reached, finally outputting the converged globally consistent transformation matrix. Finally, the global transformation matrix corresponding to each point cloud is obtained, and all point cloud data are mapped to a unified coordinate system to output the registered global point cloud. The effect of this scheme is that it can not only achieve high-precision fusion of multi-source point cloud data, but also eliminate the cumulative error caused by stepwise registration, ensuring the continuity, integrity and geometric accuracy of the entire city road point cloud, and providing reliable data support for subsequent digital twin modeling.

[0048] The initial rigid body transformation matrix is ​​implemented by first setting an initial rotation matrix and translation vector as the iteration starting point, and presetting a maximum number of iterations and a convergence threshold to terminate the iteration; the transformation parameters for the current iteration are... Iterate through all benchmark feature matching pairs and calculate the spatial error residual for each matching pair. ,in, Let mn be the spatial error residual of the feature pair under the current change; calculate the weight of each matching pair according to the Huber robust loss function, and for the residual... like Less than or equal to the set threshold Then weight ,otherwise ; rotate matrix Parameterized into a rotation vector using the Rodriguez formula. Construct a system of linear equations: ,in Rotation vector The corresponding rotation matrix is ​​used to solve the linear least squares problem by matrix inversion, yielding the updated rotation and translation vectors. The rotation vectors are then converted into rotation matrices. Finally, the difference between the current transformation parameters and the previous round is calculated. ,in It is the Frobenius norm. The rotation and translation matrices from the previous round are: Let the rotation and translation matrices be the current round's rotation and translation matrices. The iteration terminates if the value is less than the convergence threshold or the maximum number of iterations is reached, and the transformation matrix is ​​output. .

[0049] To achieve a high-fidelity recreation of physical roads in virtual space, a digital twin model of the city's roads is constructed based on laser point cloud data of the entire city's roads. First, point cloud segmentation technology is used to distinguish elements such as lanes, curbs, and traffic facilities. Then, a deep learning semantic segmentation network (such as PointNet++) is used to perform semantic recognition on the point clouds, thereby obtaining semantic labels for road structures and traffic elements. During model construction, parametric modeling methods are used to transform the point clouds into computable 3D objects. For example, road edges are fitted with spline curves, lane surfaces are generated into meshes using planar fitting, and traffic facilities are modeled using geometric primitives. Finally, a complete 3D road scene is formed in the digital twin platform.

[0050] Finally, to alleviate traffic congestion and optimize road resource allocation, traffic flow simulation and control are performed within the constructed digital twin model. The digital twin model is used to optimize traffic flow; this optimization involves obtaining traffic volume, speed, and queue length for each road segment based on the digital twin model, dynamically simulating different traffic light timings and lane control strategies, and outputting the optimal control strategy.

[0051] Specifically, the process begins by extracting current traffic flow, speed, and queue length for each road segment using a digital twin model. This digital twin model works in real-time with on-site sensing devices, collecting raw data from video surveillance, geomagnetic sensors, microwave detectors, and floating car GPS. This raw data is then mapped onto road nodes and segments within the twin model. This process accurately obtains the current traffic flow, average speed, and queue length for each road segment. Traffic flow is calculated based on the number of vehicles passing through the detection section per unit time; speed is determined by tracking changes in license plates or vehicle shapes across consecutive frames; and queue length is obtained by detecting the distance vehicles accumulate before the stop line at red lights. All data is then visually overlaid onto the virtual road within the twin model, creating a real-time traffic status map.

[0052] A congestion index is constructed by weighting traffic flow, vehicle speed, and queue length, and the congestion index is used as the objective function. Specifically, traffic flow, vehicle speed, and queue length are first normalized to obtain normalized traffic flow, vehicle speed, and queue length. Then, they are combined by weighted summation, where the normalized term for vehicle speed is negative. By weighted summation of the congestion indices for all road segments, the global congestion level is obtained as the objective function for optimization.

[0053] A multi-objective optimization algorithm is employed to find the Pareto optimal solution. Traffic light timing schemes and lane strategies are encoded as chromosomes. The execution effect of the schemes is simulated using a digital twin model, and the objective function value is calculated. High-quality individuals are selected based on congestion distance. Specifically, traffic light timing schemes (such as green light duration and cycle length for each phase) and lane strategies (such as tidal lane direction adjustment and lane priority allocation) are parameterized and encoded as chromosomes. First, the traffic light timing parameters are numerically expressed. For example, an intersection is divided into multiple phases, with the green light duration for each phase denoted as g. Red and yellow light times are set as fixed or constrained conditions, and the total cycle length is calculated. Thus, the duration of each phase can be represented as a gene segment. Second, the lane strategies are parameterized. Tidal lane direction is set as a binary variable (0 represents the original direction, 1 represents the reverse direction), and lane priority allocation is represented by integer variables indicating priority or passage weight. These parameters are concatenated in a fixed order to form a vector, which is the complete chromosome. The initial population is set by a combination of random generation and historical experience to ensure solution diversity. Each candidate solution is simulated using a digital twin model to calculate its corresponding congestion index and average delay time, among other objective function values. Crowding distance is used to select high-quality individuals to maintain population diversity and avoid premature convergence. Genetic operations include selection, crossover, and mutation: the selection phase employs a Pareto-dominant ranking mechanism; the crossover phase combines parameters of different individuals to explore better solutions; and the mutation phase randomly perturbs parameters to prevent getting trapped in local optima. The effectiveness of this approach lies in its ability to efficiently explore multiple possible control strategies within a vast search space, and its feasibility is verified through twin model simulations, ensuring both globality and reliability in the optimization process.

[0054] The optimal control strategy is output when the pre-set number of iterations is reached. Specifically, the algorithm terminates when the optimization iteration reaches the set number or when the population fitness changes less than a threshold over several consecutive generations. At this point, the set of non-dominated solutions in the population is the Pareto optimal set. The strategy is verified through simulation using a digital twin model, demonstrating predictable effects and feasibility. It is then implemented at actual intersections via a road traffic control system. This approach achieves seamless integration from the digital twin space to the physical road, ensuring the optimization results are implemented in reality. The ultimate effect is to significantly alleviate road congestion, improve road resource utilization, and enhance overall traffic efficiency.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A road management method for smart cities based on digital twins, characterized in that, Includes the following steps: S1. Deploy multiple high-precision 3D laser scanners on the road to obtain laser point cloud data, and preprocess the obtained laser point cloud data to obtain preprocessed laser point cloud data. S2. Integrate and register the preprocessed laser point cloud data obtained from multiple scanners to obtain the road laser point cloud data of the entire city. S3. Construct a digital twin model of urban roads based on the laser point cloud data of roads throughout the city; S4. Optimize traffic flow using a digital twin model; the traffic flow optimization is to obtain the traffic volume, vehicle speed and queue length of each road segment based on the digital twin model, dynamically simulate different traffic light timings and lane control strategies, and output the optimal control strategy. The implementation of step S2, which involves integrating and registering the preprocessed laser point cloud data obtained from multiple scanners to obtain the road laser point cloud data for the entire city, includes: S21. Select reference landmark features from the point cloud and extract the geometric feature descriptor for each reference feature; S22. For any two scanner point clouds, calculate the similarity using geometric feature descriptors to match the baseline features; if the similarity value reaches a set threshold, it is determined to be a match. ,in These are geometric feature descriptors acquired by any two scanners; S23. Based on the matched baseline feature pairs, solve the initial transformation matrix using least squares. Treat all scanner points as graph nodes and transform the matrix. A registration graph model is constructed for the edges, and cumulative errors are eliminated through graph optimization. The objective function of the registration graph model is: ,in The weight of the edge. Let E be the final transformation matrix of different point clouds in the global coordinate system. The Levenberg-Marquardt algorithm is used to minimize the objective function E, thus solving for the globally consistent transformation matrix. Output the registered global point cloud ; In step S23, the initial transformation matrix is ​​solved by least squares. The specific implementation is as follows: S231. Set the initial rotation matrix and translation vector as the iteration starting point, and preset the maximum number of iterations and the convergence threshold to terminate the iteration. S232, Transformation parameters for the current iteration round Iterate through all benchmark feature matching pairs and calculate the spatial error residual for each matching pair. ,in, Let mn be the spatial error residual of the mn-th feature under the current change; S233. Calculate the weight of each matching pair according to the Huber robust loss function, and for the residual... like Less than or equal to the set threshold Then weight ,otherwise ; S234, Rotation matrix Parameterized into a rotation vector using the Rodriguez formula. Construct a system of linear equations: ,in Rotation vector The corresponding rotation matrix is ​​obtained by solving the linear least squares problem through matrix inversion, resulting in the updated rotation vector and translation vector, and then the rotation vector is converted into a rotation matrix. S235. Finally, calculate the difference between the current transformation parameters and the previous round: ,in For Frobenius norm, if The iteration terminates if the value is less than the convergence threshold or the maximum number of iterations is reached, and the transformation matrix is ​​output. .

2. The road management method for smart cities based on digital twins according to claim 1, characterized in that, The preprocessing of laser point cloud data in step S1 includes coarse point cloud removal and point cloud data thinning.

3. The road management method for smart cities based on digital twins according to claim 2, characterized in that, The coarse point cloud removal includes: S111, On the original laser point cloud Each point in Search first The K nearest neighbors of the given point are determined by the radius constraint. Removal distance Exceed The effective neighborhood is obtained by identifying the far neighboring points; S112, For each point and its effective neighborhood, calculate The distance to the centroid of the neighborhood is denoted as the spatial distance feature, and is calculated... The angle between the normal vector and the average normal vector of its neighborhood is denoted as the normal vector consistency feature. Calculation... The relative deviation between the reflected intensity and the average intensity of the neighborhood is denoted as the reflected intensity deviation characteristic; S113. The spatial distance feature, normal vector consistency feature and reflection intensity deviation feature are normalized and weighted by weighted summation to obtain a weighted anomaly score. An anomaly score threshold is set. If the obtained weighted anomaly score is greater than or equal to the set threshold, it is determined to be a coarse point cloud and is removed.

4. A road management method for smart cities based on digital twins according to claim 2, characterized in that, The point cloud data thinning operation includes: S121. Point cloud after gross error removal Each point in A quadratic surface is fitted using its neighborhood, and the equation of the quadratic surface is: ,in The fitting coefficients are obtained by solving the least squares method; S122. Solve for curvature using the eigenvalues ​​of the Hessian matrix, and calculate the curvature of the quadratic surface. Hessian matrix at the location , ; S123, to Eigenvalue decomposition yields two eigenvalues. Then calculate the local principal curvature. Where j=1,2, ; S124. Take the larger principal curvature of the two as... The local curvature characterization is used to calculate the adaptive thinning radius. ,in, For adaptive thinning radius, Based on the thinning radius, The curvature sensitivity coefficient, For local curvature characterization; S125. Finally, dynamic spherical neighborhood sampling is used to achieve uniform thinning, and the point set is initialized after thinning. Empty, unsampled point set ,from Randomly select a seed point to add Calculate the adaptive thinning radius of this seed point and delete... For all points whose distance from the seed point is less than or equal to the adaptive thinning radius, repeat the selection and judgment until... Empty, ultimately This refers to the point cloud data obtained after uniform thinning.

5. A road management method for smart cities based on digital twins according to claim 1, characterized in that, The implementation of traffic flow optimization using a digital twin model in step S4 includes the following steps: S41. First, extract the current traffic flow, vehicle speed, and queue length for each road segment using a digital twin model; S42. Construct a congestion index by weighting traffic flow, vehicle speed and queue length, and use the congestion index as the objective function; S43. A multi-objective optimization algorithm is used to solve the Pareto optimal solution. The traffic light timing scheme and lane strategy are encoded as chromosomes. The implementation effect of the scheme is simulated through a digital twin model. The objective function value is calculated, and high-quality individuals are selected based on the congestion distance. S44. When the iteration reaches the preset number of iterations, output the optimal control strategy.

Images