Curve road section identification and lane line extraction method and system based on millimeter wave radar

Abstract

The invention relates to the technical field of intelligent traffic, in particular to a curve road section identification and lane line extraction method and system based on millimeter wave radar. According to the invention, the point cloud data of a road traffic target is subjected to preprocessing and dynamic noise compensation through distance-Doppler combined filtering, and lane features are enhanced through multi-frame point cloud fusion; through track point cloud cluster extraction of a moving vehicle target, a lane center point is accurately positioned in combination with an improved transverse variation kernel density estimation algorithm. Variational mode decomposition is introduced to solve the problem of multi-lane aliasing, and segmented cubic polynomial fitting is adopted to achieve conversion from discrete points to a continuous lane line model. According to the method, the problem of detection failure caused by lane line abrasion and shielding in a complex road scene is solved, the curve area identification error is reduced, high precision is kept in a straight line scene and a curve scene, reliable technical support can be provided, the robustness and accuracy of all-weather road perception are improved, and the method has important application value and practical significance.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and system for identifying curved road sections and extracting lane lines based on millimeter-wave radar. Background Technology

[0002] With the advancement of smart city construction, the accuracy requirements for lane detection and curve identification in intelligent transportation systems are increasing. Existing technologies mainly rely on visual sensors and lidar, but these have significant limitations under complex weather conditions. While millimeter-wave radar offers all-weather advantages, it faces challenges such as sparse point clouds and dynamic noise interference. Traditional methods, such as the DBSCAN clustering algorithm, produce trajectory drift errors (average error exceeding 0.3 meters) when curve curvature changes abruptly. Least-squares fitting schemes do not model the dynamic changes in lane width, leading to high boundary missegmentation rates. Furthermore, fixed-parameter filtering struggles to adapt to dynamic scenes, and single-frame point cloud processing cannot overcome occlusion and wear issues.

[0003] Millimeter-wave radar autonomous control systems face three major technical bottlenecks in complex urban scenarios: point cloud blurring caused by moving targets significantly reduces the signal-to-noise ratio; fixed-parameter filtering struggles to adapt to dynamic road environments; and spectral aliasing occurs under multi-target interference. Traditional clustering algorithms cannot address the nonlinear decay of point cloud density with distance, resulting in unstable feature extraction for long-range targets, and clustering radius parameters struggle to balance near-field accuracy and far-field sensitivity. Furthermore, single-control architectures exhibit significant response delays, and computational resource contention during multi-target tracking leads to timing asynchrony. These technical deficiencies severely restrict the perception performance and control reliability of millimeter-wave radar in complex urban scenarios such as dense traffic flow and intersections.

[0004] Existing research shows that lane line extraction methods based on DBSCAN clustering suffer from trajectory drift when there are abrupt changes in curve curvature; least squares fitting schemes fail to model lane characteristics, leading to incorrect boundary segmentation; and traditional polynomial fitting is prone to overfitting and exhibits poor curve continuity. These shortcomings limit the practical application effectiveness of millimeter-wave radar in complex road scenarios.

[0005] Therefore, a method and system for identifying curved road sections and extracting lane lines based on millimeter-wave radar is needed. Summary of the Invention

[0006] To address the problems in existing technologies, this invention provides a method and system for identifying curved road sections and extracting lane lines based on millimeter-wave radar. The specific technical solution is as follows: A method for identifying curved road sections and extracting lane lines based on millimeter-wave radar includes the following steps: Step S1: Collect point cloud data of road traffic targets using millimeter-wave radar equipment. The point cloud data of road traffic targets includes the target ID number, radial distance, radial velocity, and angle of the traffic targets relative to the millimeter-wave radar equipment. Step S2: The point cloud data of road traffic targets is preprocessed based on the range-Doppler joint filtering method to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment; Step S3: Construct a sliding window mechanism, divide the sliding window at fixed intervals along the lane rectangular coordinate system parallel to the vehicle driving direction, and divide the grid in the lane rectangular coordinate system perpendicular to the vehicle driving direction within each window to form a two-dimensional processing unit; Step S4: Within each two-dimensional processing unit, the spatial distribution of the point cloud data of the moving vehicle target is statistically modeled. The point cloud data distribution within each two-dimensional processing unit is statistically analyzed to generate a lane feature map. The lane feature map contains several trajectory point cloud clusters of the moving vehicle target. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle target relative to the millimeter-wave radar device is considered. Adjust the neighborhood radius to perform clustering and extract candidate point clusters for lane centers; Step S5: Extract the lane center point set from the candidate lane center point cluster using an improved lateral variational kernel density estimation algorithm; Step S6, use variational mode decomposition to decompose the lane center point set into... K A subset of lane center points; Step S7, based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model for each lane center point subset, thereby realizing the modeling of curves and straight road segments.

[0007] Preferably, the preprocessing of the point cloud data of road traffic targets based on the range-Doppler joint filtering method in step S2 to obtain the moving vehicle point cloud data in the lane rectangular coordinate system specifically includes the following steps: In step S21, the millimeter-wave radar device transmits M linear frequency modulated pulses to the road traffic target within one coherent processing interval. The number of sampling points for each linear frequency modulated pulse is N. The echo signals received by the millimeter-wave radar device constitute an N×M complex matrix. x(n,m), in n Index value representing the distance dimension ，n= 0,1,... ,N-1 ; m The index value represents the dimension of the linear frequency modulated pulse. m= 0,1,... ,M-1 ; Step S22: Perform a Fast Fourier Transform on the echo signal received by the millimeter-wave radar equipment to convert the signal from the time domain to the range domain. The calculation formula is: ; ; In the formula, r This represents the range cell index in the range domain, indicating the radial distance of a road traffic target relative to a millimeter-wave radar device. Among them, distance resolution , c At the speed of light, B For millimeter-wave radar equipment; Indicates the first r The signal of the m-th linear frequency modulated pulse in a distance unit; Step S23: Perform a Fast Fourier Transform on the signal sequence of each range unit in the range domain between different linear frequency modulated pulses to obtain the range-Doppler matrix. The calculation formula is as follows: ; in, Index representing a Doppler cell; For the first r The distance unit is the The distance-Doppler matrix of a Doppler unit; Step S24 involves calculating the phase difference between the range-Doppler matrices of the same range cell and the same Doppler cell in two adjacent frames and comparing it with a set phase threshold to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets. Specifically: No. k Frame and the k The distance-Doppler matrices of the same range cell and the same Doppler cell in the +1 frame are respectively and The time interval between two adjacent frames is Take the complex value at the same position in the distance-Doppler matrix of two adjacent frames, and calculate the phase difference of the complex signal. The calculation formula is: ; in, Indicates taking the conjugate. This indicates a phase take operation. The unit is radians, and the range is ; For each The unit performs phase difference calculations and makes the following judgments to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets; ; in, The set phase threshold; Step S25: Process the statically separated range-Doppler matrix to extract effective point cloud information of the moving target. Specifically, apply the CFAR detection algorithm to the amplitude of the statically separated range-Doppler matrix to filter out effective detection points with energy higher than the background noise, and obtain the coordinates of the effective detection points in the statically separated range-Doppler matrix. ; Step S26, extracting valid detection points The power spectrum is obtained by beamforming the corresponding echo signal in the spatial dimension. Estimate the angle between the moving target and the illumination direction of the millimeter-wave radar equipment. ， Power spectrum The calculation formula is as follows: ; In the formula, The first part represents the millimeter-wave radar equipment. l The root antenna is in the corresponding Complex signals on the unit, for , This refers to the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment. L The total number of antenna channels for a millimeter-wave radar device; measured by power spectrum. The angle relationship corresponding to the maximum peak value: The angle between the moving target and the illumination direction of the millimeter-wave radar equipment is obtained. The estimated value ; Step S27, assuming the location of the millimeter-wave radar equipment is the origin of the coordinate system, within the plane containing the road surface, x The coordinate axes are parallel to the vehicle's direction of travel. y If the coordinate axes are perpendicular to the vehicle's direction of travel, then the spherical coordinates of the moving target are... With respect to the radar rectangular coordinate system The coordinate transformation relationship is as follows: ; in, This indicates the position of the moving target in the radar Cartesian coordinate system, parallel to the vehicle's direction of travel. This indicates the position of a moving target in the radar Cartesian coordinate system, perpendicular to the vehicle's direction of travel. Step S28, establish a lane rectangular coordinate system, then in the lane rectangular coordinate system With radar coordinate system The coordinate transformation relationship is as follows: ; Finally, each moving target is represented as a quadruple containing position and velocity information. ; Step S29: Perform dynamic noise compensation and motion distortion correction on each moving target to obtain point cloud data of the moving vehicle target in the lane rectangular coordinate system.

[0008] Preferably, the phase threshold set in step S24 The determination method is as follows: Establish phase difference radial velocity relative to the traffic target by the millimeter-wave radar device The specific relationships are as follows: ; in, Indicates the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment; Set the radial velocity of the traffic target relative to the millimeter-wave radar equipment speed threshold This is used to distinguish between stationary and moving targets; The set phase threshold The calculation method is as follows: .

[0009] Preferably, step S29, which involves performing dynamic noise compensation and motion distortion correction on each moving target to obtain point cloud data of the moving vehicle target in the lane rectangular coordinate system, specifically includes the following steps: Step S291, construct the noise power function, as follows: ; in, It refers to the radial velocity of a traffic target relative to the millimeter-wave radar device during measurement. v and radial distance R The noise variance value determined jointly; Step S292, construct a Butterworth filter, wherein the dynamic cutoff frequency of the Butterworth filter is... Specifically as follows: ; in, This refers to the system sampling frequency; ; Step S293: Combine the main energy distribution frequency band of each moving target signal with the dynamic cutoff frequency of the Butterworth filter. Comparison: If the energy of the moving target signal is mainly concentrated at the dynamic cutoff frequency of the Butterworth filter The following frequency bands are considered to represent real moving vehicle targets and are retained. If the energy of the moving target signal is significantly distributed at the dynamic cutoff frequency of the Butterworth filter If any of the above are found, the moving target is determined to be residual noise or a false target and is therefore eliminated. Step S294: Following step S293, valid moving vehicle targets are filtered out, and each moving vehicle target is ultimately represented as a quadruple containing position and velocity information. , representing the point cloud data of moving vehicle targets in the lane rectangular coordinate system.

[0010] Preferably, step S4, which uses an adaptive Euclidean clustering algorithm to process trajectory point cloud clusters within a two-dimensional processing unit and adjusts the neighborhood radius based on the relative distance to the moving vehicle target for clustering, specifically includes the following steps: Step S41: Calculate the value of each data point within the two-dimensional processing unit using the Gaussian kernel function. Local density and relative distance The calculation formula is: ; In the formula, Represents the i-th data point The local density estimate; Represents the i-th data point The minimum distance to a higher density point; where Let i and j represent the i-th and j-th data points within the two-dimensional processing unit, respectively. This represents the Euclidean distance between the i-th data point and the j-th data point; h The bandwidth parameter represents the Gaussian kernel function; Step S42: Traverse all data points within the two-dimensional processing unit, and calculate the radial distance of each moving vehicle target relative to the millimeter-wave radar device. The neighborhood radius, i.e. the initial cluster radius, is dynamically adjusted, and clustering is performed using the initial cluster radius to obtain several clusters; Step S43: Perform a cluster merging operation based on density similarity on several clusters to obtain the merged clusters; Step S44, for each cluster obtained by merging The points in the middle, according to the timestamp t Sort the points in ascending order, then sort them according to their direction of travel in the lane Cartesian coordinate system, forming a point sequence. The representative points of each cluster are calculated using a weighted average method based on local density. coordinates That is, the candidate point for the center of the lane, the representative point. The formula for calculating the coordinates is: Calculate representative points x Axis coordinates : ; Calculate representative points y Axis coordinates : ; in, express The number of points in For point The local density estimate.

[0011] Preferably, step S5, which uses an improved lateral variational kernel density estimation algorithm to extract the lane center point set from the candidate lane center point cluster, specifically includes the following steps: Step S51, for each data point in the candidate cluster of lane centers x The probability density estimation of the axis position distribution is calculated using the following formula: ; In the formula, Indicates that the data points are in x The probability density estimate at point ; n This indicates the total number of candidate center points for the lane; Represents the Gaussian kernel function; Indicates the first i Lane center candidate point x Axis coordinates; Indicates an indicator function, when the condition is met. The value is 1 if the condition is met, and 0 otherwise. Represents all data points within a two-dimensional processing unit. x Average of axis coordinates; Indicates lane width; h The bandwidth parameter represents the Gaussian kernel function; Step S52: Calculate the probability density distribution for each data point in the candidate lane center point cluster, where the lane center point corresponds to the probability density distribution. The local maxima are identified, and the local density peaks are used as the final lane center points, thus obtaining the lane center point set. The detection conditions for the local density peaks are as follows: ; In the formula, for The first derivative, for The second derivative, The parameter used to test the significance of density; Step S53: Obtain the x-axis position of the local density peak point detected in each lane center candidate point cluster, satisfying the detection conditions for local density peak points. The x-axis position is selected from the original candidate lane center points and the x-axis position is compared with the x-axis position. Points whose distance to the center point is less than a set x-axis distance threshold are determined as the final lane center points, thus obtaining the lane center point set. .

[0012] Preferably, the bandwidth parameter of the Gaussian kernel function h Based on the total number of target points of moving vehicles in step S2 Adaptive adjustment, the specific correspondence is: .

[0013] Preferably, in step S6, variational mode decomposition is used to decompose the lane center point set into... K Specifically, the lane center point subset is obtained by using variational mode decomposition to extract the mixed lane center point set containing information from all lanes. Separate by lane to obtain K Set of center points of each lane Specifically: Step S61, based on the lane center point set ,along y Statistically analyze the distribution of all points along the axis to generate a probability density function. The calculation formula is: ; in, n For the set of lane center points The total number of midpoints; The center point of the i-th lane y Axis coordinates; Step S62, based on probability density function The probability density function is obtained by using a variational mode decomposition model. Decomposed into K An independent mode The decomposition is achieved by solving the following constraints to optimize the objective function: ; ; in, The term represents minimizing the gradient of the modal change; The term represents the calculation of modal components. of Norm; This represents the first regularization parameter; Number of lanes K Determined based on the following energy gradient ratio criterion: ; In the formula, , representing belonging to the first k The energy contained in the point set distribution of each lane; This represents all currently identified [items / entities]. K The total energy of all modes; It is the gradient vector Norm, representing the first norm k Density distribution of candidate lanes The overall fluctuation range; Step S63, setting K =1, start and execute the solution of the energy gradient ratio criterion, and obtain the energy gradient ratio. When two consecutive times Change When the iteration ends, the increment stops. K At this time K This is the optimal estimated number of lanes; Step S64, for each modal component ,turn up K The local maximum values ​​correspond to y The axis position is obtained. K Peak position of each y-axis ; Step S65, for the set of lane center points any point in Calculate its y-axis coordinate to the th k Peak position of each y-axis lateral European distance ; Step S66, based on the lateral Euclidean distance Points where the center points of the lanes are concentrated Assign to the nearest location that is less than the assignment distance threshold. The lane, thus obtainingK A subset of lane center points separated by lane. .

[0014] Preferably, in step S7, based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model using a subset of lane center points. The specific steps for modeling curves and straight road segments are as follows: Step S71, let the first... k The origin of the local coordinate system for each lane sub-section is... The local coordinate expression of the lane centerline within the lane is: ; In the formula, Let be the vector of polynomial coefficients to be determined; Step S72, perform weighted least squares method to optimize polynomial coefficients, the first... k The objective function for each lane sub-interval is: ; In the formula, This represents the total number of lane center points falling within the k-th lane sub-section; Let be the coordinates of the center point of the i-th lane corresponding to this lane; The weighting coefficient, which is positively correlated with point cloud density, is calculated using the following formula: ; For the smoothing regularization term, where Represents the curvature term, which mathematically characterizes the degree of curvature of a curve; This represents the second regularization parameter; Step S73: For each lane, execute the cubic polynomial parameterized model of the lane centerline established in steps S71 and S72 respectively. Use the cubic polynomial fitting method on the lane interval to reconstruct the discrete lane center points into a continuous and smooth lane line topology model, thereby realizing the modeling of curves and straight road segments.

[0015] A system for identifying curved road sections and extracting lane lines based on millimeter-wave radar, employing the method described above, includes: The data acquisition module is used to acquire point cloud data of road traffic targets through millimeter-wave radar equipment. The point cloud data of road traffic targets includes the target ID number, radial distance, radial velocity and angle of the traffic target relative to the millimeter-wave radar equipment. The data processing module is used to preprocess the point cloud data of road traffic targets based on the range-Doppler joint filtering method, so as to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment; The mesh generation module is used to construct a sliding window mechanism, which divides the lane into rectangular coordinates parallel to the vehicle's direction of travel at fixed intervals, and within each window, it divides the lane into rectangular coordinates perpendicular to the vehicle's direction of travel to form a two-dimensional processing unit. The clustering module is used to statistically model the spatial distribution of point cloud data of moving vehicle targets within each two-dimensional processing unit, statistically analyze the point cloud data distribution within each two-dimensional processing unit, and generate a lane feature map. The lane feature map contains several trajectory point cloud clusters of moving vehicle targets. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle targets relative to the millimeter-wave radar equipment is considered. Adjust the neighborhood radius to perform clustering and extract candidate point clusters for lane centers; The center point extraction module is used to extract the set of lane center points from the cluster of lane center candidate points using an improved lateral variational kernel density estimation algorithm. The lane decomposition module is used to decompose the lane center point set into lanes using variational mode decomposition. K A subset of lane center points; Lane line fitting module, used for based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model for each lane center point subset, thereby realizing the modeling of curves and straight road segments.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention preprocesses and dynamically compensates for point cloud data of road traffic targets using distance-Doppler joint filtering, and enhances lane features through multi-frame point cloud fusion. It accurately locates lane center points by extracting point cloud clusters from the trajectory of moving vehicle targets and combining this with an improved lateral variational kernel density estimation algorithm. A three-stage processing framework is constructed, comprising target point cloud density distribution analysis, improved kernel density estimation, and piecewise cubic polynomial modeling. Variational mode decomposition is introduced to solve the multi-lane aliasing problem, and piecewise cubic polynomial fitting is used to convert discrete points into continuous lane line models. This invention solves the detection failure problem caused by lane line wear and occlusion in complex road scenarios, reduces identification errors in curved areas, maintains high accuracy in both straight and curved scenarios, provides reliable technical support, and improves the robustness and accuracy of all-weather road perception, possessing significant application value and practical significance. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a flowchart illustrating the preprocessing of point cloud data of road traffic targets.

[0020] Figure 3 This is a flowchart of variational kernel density estimation and variational mode decomposition.

[0021] Figure 4 This is a schematic diagram of the process for extracting lane line information and fitting lane models using millimeter-wave radar point clouds, as provided by the present invention.

[0022] Figure 5 This is a schematic diagram of lane center extraction using cubic polynomial fitting in a straight road scenario, based on the present invention.

[0023] Figure 6 This is a schematic diagram of the lane information extraction results of the present invention on straight roads and curves.

[0024] Figure 7 This is a schematic diagram showing the comparison results of lane-level error indices in straight and curved road scenarios according to the present invention.

[0025] Figure 8 This is a system schematic diagram of the present invention. Detailed Implementation

[0026] 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, not all, of the embodiments of the present invention. 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.

[0027] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] Example 1: like Figure 1 As shown, this embodiment provides a method for identifying curved road sections and extracting lane lines based on millimeter-wave radar. Addressing the technical bottlenecks of sparse millimeter-wave radar point cloud data, significant noise interference, and difficulties in extracting curved geometric features in complex road environments, it reconstructs the lane topology by constructing a multi-level data processing and adaptive modeling framework. Specifically, it includes the following steps: Step S1 involves acquiring point cloud data of road traffic targets using millimeter-wave radar equipment. This point cloud data includes the target ID number, radial distance, radial velocity, and angle of the traffic target relative to the millimeter-wave radar equipment. The number of traffic targets can be determined based on the target ID number. The millimeter-wave radar equipment uses a Texas Instruments (TI) IWR1843 high-frequency millimeter-wave radar sensor as the core data acquisition unit, and acquires point cloud data through a millimeter-wave radar acquisition module.

[0031] Step S2: The point cloud data of road traffic targets is preprocessed based on the range-Doppler joint filtering method to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment.

[0032] Because the relative motion of vehicles causes spectral broadening and dynamic noise, preprocessing of point cloud data of road traffic targets is necessary. The flowchart is as follows: Figure 2 As shown, the specific steps include: In step S21, the millimeter-wave radar device transmits M linear frequency modulated pulses to the road traffic target within one coherent processing interval. The number of sampling points for each linear frequency modulated pulse is N. The echo signals received by the millimeter-wave radar device constitute an N×M complex matrix. x(n,m), in n Index value representing the distance dimension ，n= 0,1,... ,N-1 ; m The index value represents the dimension of the linear frequency modulated pulse. m= 0,1,... ,M-1 .

[0033] Step S22: Perform a Fast Fourier Transform on the echo signal received by the millimeter-wave radar equipment to convert the signal from the time domain to the range domain. The calculation formula is as follows: (1) (2) In the formula, r This represents the range cell index in the range domain, indicating the radial distance of a road traffic target relative to a millimeter-wave radar device. Among them, distance resolution , c At the speed of light, B For millimeter-wave radar equipment; Indicates the first r The signal of the m-th linear frequency modulated pulse in a distance unit.

[0034] Step S23: Perform a Fast Fourier Transform on the signal sequence of each range unit in the range domain between different linear frequency modulated pulses to obtain the range-Doppler (RD) matrix. The calculation formula is as follows: (3) in, Index representing a Doppler cell; For the first r The distance unit is the The distance-Doppler matrix of each Doppler unit.

[0035] Step S24 involves calculating the phase difference between the range-Doppler matrices of the same range cell and the same Doppler cell in two adjacent frames and comparing it with a set phase threshold to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets. Specifically: No. k Frame and the k The distance-Doppler matrices of the same range cell and the same Doppler cell in the +1 frame are respectively and The time interval between two adjacent frames is Take the complex value at the same position in the distance-Doppler matrix of two adjacent frames, and calculate the phase difference of the complex signal. The calculation formula is: (4) in, Indicates taking the conjugate. This indicates a phase take operation. The unit is radians, and the range is .

[0036] For each The unit performs phase difference calculations and makes the following judgments to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets; (5) in, The set phase threshold.

[0037] Set phase threshold The determination method is as follows: Based on radar principles, establish phase difference radial velocity relative to the traffic target by the millimeter-wave radar device The specific relationships are as follows: (6) in, This indicates the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment.

[0038] Set the radial velocity of the traffic target relative to the millimeter-wave radar equipment speed threshold =5m / s is used to distinguish between stationary and moving targets; The set phase threshold The calculation method is as follows: ;(7)

[0039] Step S25: Process the statically separated range-Doppler matrix to extract effective point cloud information of the moving target. Specifically, apply the CFAR detection algorithm to the amplitude of the statically separated range-Doppler matrix to filter out effective detection points with energy higher than the background noise, and obtain the coordinates of the effective detection points in the statically separated range-Doppler matrix. .

[0040] Step S26, extracting valid detection points The power spectrum is obtained by beamforming the corresponding echo signal in the spatial dimension. Estimate the angle between the moving target and the illumination direction of the millimeter-wave radar equipment. ， Power spectrum The calculation formula is as follows: (8) In the formula, The first part represents the millimeter-wave radar equipment. l The root antenna is in the corresponding Complex signals on the unit, for , This refers to the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment. L This represents the total number of antenna channels for the millimeter-wave radar equipment, set to 12; based on the power spectrum... The angle relationship corresponding to the maximum peak value: The angle between the moving target and the illumination direction of the millimeter-wave radar equipment is obtained. The estimated value .

[0041] Step S27, it is hereby stated that this millimeter-wave radar equipment is installed upright, and the radar installation height is [missing information]. H Adjust the installation height to ensure coverage of the entire area of ​​interest. When installing the radar, tilt it downwards at a certain angle. α This refers to the radar tilt angle, taken as -3°. Assuming the millimeter-wave radar equipment is located at the origin of the coordinate system, within the plane of the road surface... x The coordinate axes are parallel to the vehicle's direction of travel. y If the coordinate axes are perpendicular to the vehicle's direction of travel, then the spherical coordinates of the moving target are... With respect to the radar rectangular coordinate system The coordinate transformation relationship is as follows: (9) in, This indicates the position of the moving target in the radar Cartesian coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving target in the radar Cartesian coordinate system, perpendicular to the vehicle's direction of travel.

[0042] Step S28, establish a lane rectangular coordinate system, then in the lane rectangular coordinate system With radar coordinate system The coordinate transformation relationship is as follows: (10) Finally, each moving target is represented as a quadruple containing position and velocity information. ; Step S29 involves performing dynamic noise compensation and motion distortion correction on each moving target to obtain point cloud data of the moving vehicle targets in the lane rectangular coordinate system. This specifically includes the following steps: Step S291: Construct a noise power function, where the noise power is proportional to the 1.5th power of the velocity and proportional to the distance. This model quantifies the noise attenuation characteristics as the target moves and the distance increases. Specifically: (11) in, It refers to the radial velocity of a traffic target relative to the millimeter-wave radar device during measurement. v and radial distance R The noise variance value determined jointly; Step S292, construct a Butterworth filter, wherein the dynamic cutoff frequency of the Butterworth filter is... Specifically as follows: (12) in, This refers to the system sampling frequency; The radial velocity of a traffic target relative to the millimeter-wave radar device during measurement. v and radial distance R The noise standard deviation is determined jointly.

[0043] Step S293: Combine the main energy distribution frequency band of each moving target signal with the dynamic cutoff frequency of the Butterworth filter. Comparison: If the energy of the moving target signal is mainly concentrated at the dynamic cutoff frequency of the Butterworth filter The following frequency bands are considered to represent real moving vehicle targets and are retained. If the energy of the moving target signal is significantly distributed at the dynamic cutoff frequency of the Butterworth filter For example, if high-frequency components are generated due to noise or interference, the moving target is determined to be residual noise or a false target and is therefore eliminated. Step S294: Following step S293, valid moving vehicle targets are filtered out, and each moving vehicle target is ultimately represented as a quadruple containing position and velocity information. , representing the point cloud data of moving vehicle targets in the lane rectangular coordinate system.

[0044] By filtering out valid moving vehicle targets using the aforementioned frequency bands, each moving vehicle target can ultimately be represented as a quadruple containing position and velocity information. This represents a point cloud of moving vehicles with accurate positioning and reliable speed in the lane rectangular coordinate system.

[0045] To address signal distortion caused by motion, a power noise function with velocity and distance as bivariates is established. And through the parameters of the adaptive Butterworth filter The cutoff frequency is dynamically adjusted to suppress the spectral broadening caused by Doppler shift. Based on the power noise model, an adaptive Butterworth filter is used, whose cutoff frequency is adjusted according to the calculated noise power. The signal-to-noise ratio of the filtered signal is improved, thereby effectively suppressing signal distortion.

[0046] Step S3: Traditional fixed-window data processing methods are ill-suited to the drastic fluctuations in point cloud density and distribution characteristics caused by changes in vehicle speed and road curvature radius. Multi-frame fusion is employed to accumulate and enhance the pre-processed moving vehicle target point cloud. A sliding window mechanism is constructed to intelligently accumulate point cloud data from multiple consecutive frames. Sliding windows are divided at fixed intervals of 5m along the lane Cartesian coordinate system, parallel to the vehicle's direction of travel. Within each window, a grid is created perpendicular to the vehicle's direction of travel in the lane Cartesian coordinate system, with a resolution of 0.3m × 0.3m, forming a two-dimensional processing unit.

[0047] Step S4: Within each two-dimensional processing unit, the spatial distribution of the point cloud data of the moving vehicle target is statistically modeled (approximately Gaussian distribution). The point cloud data distribution within each two-dimensional processing unit is statistically analyzed to generate a high-density, enhanced lane feature map. The lane feature map contains several trajectory point cloud clusters of moving vehicle targets. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle target relative to the millimeter-wave radar device is considered. Clustering is performed by adjusting the neighborhood radius to accurately extract a set of discrete candidate lane center points. Specifically, this includes the following steps: Step S41: Calculate the value of each data point within the two-dimensional processing unit using the Gaussian kernel function. Local density and relative distance The calculation formula is: (13) In the formula, Represents the i-th data point The local density estimate; Represents the i-th data point The minimum distance to a higher density point; where Let i and j represent the i-th and j-th data points within the two-dimensional processing unit, respectively. This represents the Euclidean distance between the i-th data point and the j-th data point; h The bandwidth parameter represents the Gaussian kernel function; Step S42: Traverse all data points within the two-dimensional processing unit, and calculate the radial distance of each moving vehicle target relative to the millimeter-wave radar device. The neighborhood radius, i.e. the initial cluster radius, is dynamically adjusted, and clustering is performed using the initial cluster radius to obtain several clusters.

[0048] Specifically, when the moving vehicle target is At that time, the neighborhood radius is set to 0.3. When the target is At that time, the neighborhood radius is set to 0.6. ; where the maximum distance threshold Preliminary clustering was performed, focusing on clusters in close proximity (…). Using small cluster radii to achieve fine segmentation and prevent target adhesion; at long distances ( Use a large cluster radius to achieve inclusive aggregation and avoid the target being fragmented due to sparse point cloud.

[0049] Step S43: Perform a cluster merging operation based on density similarity on several clusters to obtain merged clusters. Specifically, after completing the initial clustering, many clusters are obtained, and for local density differences... Clusters with a density similarity that are strictly less than 20% of the larger of the two local densities and whose relative distance is strictly less than 60% of the dynamic threshold are merged based on density similarity. This avoids clustering of targets at close range and segmentation of targets at long range. If any of these conditions are not met, merging is not allowed. The specific conditions for determining cluster merging are as follows: (14) (15) In the formula, d Indicates the effective radius of the cluster; This represents the absolute difference in local density between two clusters to be merged, i.e., the local density difference. This represents the core density of the first and second clusters to be merged. , The dynamic distance thresholds for the two clusters are used, and cluster feature extraction is performed using the pcsegdist function in MATLAB to calculate the average distance of each cluster. That is, the average relative distance of all points within the cluster; based on the average distance The value is compared with the set threshold of 60; if If the value is less than 60, the function will return a dynamic distance threshold of 0.1 for that cluster. ;like If the value is greater than or equal to 60, the function will return a dynamic distance threshold of 0.05 for that cluster. .

[0050] Step S44, for each cluster obtained by merging The points in the middle, according to the timestamp t Sort the points in ascending order, then sort them according to their direction of travel in the lane Cartesian coordinate system, forming a point sequence. The representative points of each cluster are calculated using a weighted average method based on local density. coordinates That is, the candidate point for the center of the lane, the representative point. The formula for calculating the coordinates is: Calculate representative points x Axis coordinates : (16) Calculate representative points y Axis coordinates : (17) in, express The number of points in threshold ; For point The local density estimate is obtained from the Gaussian kernel density estimate.

[0051] The calculations above are respectively x coordinates and y Combining the coordinates, we obtain the two-dimensional coordinates of the representative point of this cluster (i.e., the candidate point for the lane center): (18) For clusters The representative point is the candidate point for the center of the lane.

[0052] It is worth emphasizing that the sliding window segmentation method used in the above processing performs well in straight areas with small changes in road curvature; however, when vehicles enter curves with significant curvature, the fixed window segmentation struggles to accurately capture the rapidly changing geometric features of lane lines. Therefore, a curvature-adaptive window segmentation strategy is further introduced. By calculating the road's radius of curvature, the size of the longitudinally segmented window is dynamically adjusted, significantly improving the recognition accuracy of curved road sections. This is because at curves, lane line directions change rapidly, and smaller processing units can more sensitively capture and follow this curve shape, thus ensuring that the generated lane centerline candidate points accurately match the actual road alignment. Specifically: For a discrete, ordered sequence of lane center candidate points Indicates the candidate points for the lane center (already arranged along the road direction, i.e.) x (Coordinates are strictly monotonically increasing, sorted) for any point radius of curvature at The calculation formula is: (19) In the formula, Point radius of curvature at point, subscript i This indicates the index of the point in the sequence; Point The first derivative at the trajectory is dimensionless. Point Second derivative at the trajectory, unit: .

[0053] Take the median of the radii of curvature of all points in the sequence as the radius of curvature. : (20) Based on the calculated representative radius of curvature It performs adaptive decision-making, dynamically adjusting the vertical division size of the sliding window, specifically: Based on the "Highway Route Design Specifications" and a large amount of measured data, a radius of curvature threshold is set. m, this threshold is an engineering experience value that distinguishes between gentle curves and sharp curves, and can effectively divide different road sections and adaptive decision-making logic.

[0054] when When determining whether the current road is a straight or gently curved section, the road direction changes gradually and the lane lines deviate slowly laterally. The decision is to use a 5m longitudinal division window size for division.

[0055] when When the current section is determined to be a sharp bend, the road direction changes drastically and the lane lines deviate significantly laterally within a short distance. Switch to a 3m longitudinal division window size for division.

[0056] This step further improves the reliability of the candidate lane center point set, providing a high-quality data foundation for the subsequent construction of a continuous and accurate lane line model. Ultimately, this step outputs a series of filtered candidate point clusters that represent the lane center location, providing crucial data features for the entire lane line extraction process.

[0057] Step S5 involves extracting the lane center point set from the candidate lane center point cluster using an improved lateral variational kernel density estimation algorithm. The core task of this step is to accurately locate the center position of each lane. The key technology lies in effectively extracting the lane center point distribution that corresponds to the actual lane from the candidate lane center points, while overcoming interference caused by multi-lane overlap and environmental noise. Figure 3 As shown, the specific steps include: Step S51: Perform lateral variational kernel density estimation on lane center candidate points for each data point in the lane center candidate point cluster. x The probability density estimation of the axis position distribution is calculated using the following formula: ;(twenty one) In the formula, Indicates that the data points are in x The probability density estimate at point ; n This indicates the total number of candidate center points for the lane; Represents the Gaussian kernel function. ; Indicates the first i Lane center candidate point x Axis coordinates; Indicates an indicator function, when the condition is met. The value is 1 if the condition is met, and 0 otherwise. Represents all data points within a two-dimensional processing unit. x Average of axis coordinates; Indicates lane width; h The bandwidth parameter of the Gaussian kernel function is set to 0.5m.

[0058] Bandwidth parameter of Gaussian kernel function h Based on the total number of target points of moving vehicles in step S2 Adaptive adjustment ensures the stability of lane center point extraction. Specifically, it extracts all target points of the current frame from the preprocessed point cloud stream and counts the total number of target points in the entire perception scene. Since the scene range is relatively fixed, there is no need for area normalization; it can be used directly. As a measure of density, bandwidth parameter h The baseline value is 0.5m, which will be adjusted based on the total number of target points in the scene. The range of quantities is used to perform parameter optimization and adaptive adjustment, and the specific correspondence is as follows: ;(twenty two).

[0059] Based on extensive measured data and engineering experience from numerous road scenarios, the target point cloud quantity is divided into five intervals with clear physical meaning: extremely sparse (1.25), sparse (1.0), normal (0.5), dense (0.35), and extremely dense (0.15). Each interval corresponds to a preset bandwidth adjustment coefficient. For example, in scenarios with very few target points, such as at night or in severe weather, a larger bandwidth coefficient will be automatically selected. h To achieve a strong smoothing effect, suppress noise, and ensure the stability of the estimation; however, in congested road sections with a large number of target points, a smaller bandwidth coefficient will be selected. h This improves the resolution of kernel density estimation and allows for finer separation of features in adjacent lanes.

[0060] To ensure smooth parameter adjustments and avoid frequent parameter jumps due to minor inter-frame fluctuations in point cloud density, density changes in adjacent frames are monitored, and a hysteresis interval of ±5 points is set to prevent frequent switching caused by slight density fluctuations. Parameter switching is only triggered when the monitored data from multiple consecutive frames stably exceeds the current interval boundary and crosses this tolerance band, effectively filtering out transient interference and ensuring the continuity and smoothness of the system output. Therefore, bandwidth parameters are constructed based on the target point cloud density environment. h The optimization mechanism dynamically adjusts algorithm parameters through strategies to ensure the stability of lane center point extraction.

[0061] Step S52: Calculate the probability density distribution for each data point in the candidate lane center point cluster, where the lane center point corresponds to the probability density distribution. The local maxima are identified, and the local density peaks are used as the final lane center points, thus obtaining the lane center point set. The detection conditions for the local density peaks are as follows: For lane center candidate point The necessary conditions for the existence of a peak are that the first derivative is zero and the second derivative is less than 0. Using the central difference method, the calculation formula is as follows: ;(twenty three) ;(twenty four) In the formula, Take 0.1m; satisfy: (25) In the formula, for The first derivative, for The second derivative, =0.2 is the parameter for testing the significance of density.

[0062] Step S53: Obtain the x-axis position of the local density peak point detected in each lane center candidate point cluster, satisfying the detection conditions for local density peak points. The x-axis position is selected from the original candidate lane center points and the x-axis position is compared with the x-axis position. Points whose distance to the center point is less than a set x-axis distance threshold are determined as the final lane center points, thus obtaining the lane center point set. The x-axis distance threshold is set to 0.1m, and the final formula for calculating the lane center point is: (26) After filtering, the final output is the set of lane center points. for: ;(27)

[0063] This yields a set of points representing the center positions of all lanes, which in x The points are roughly evenly distributed in direction. On multi-lane roads, this set of points is a mixture of all lane center points, without distinguishing the specific lane to which each point belongs. To solve this multi-lane aliasing problem, variational mode decomposition (VMD) is introduced, aiming to transform the mixed lane center point set containing information from all lanes. Separate by lane to obtain K A subset of lane center points .

[0064] Unlike standard methods, this invention innovatively introduces a lane physical width constraint mechanism. Specifically, when calculating the contribution of each candidate point to the overall density distribution, the algorithm uses an indicator function to determine the average value of that point relative to all points within its corresponding two-dimensional processing unit. x Axis coordinates Is the lateral offset less than the standard lane width? Half of the initial probability density distribution. This mechanism can effectively suppress interference from abnormal trajectory points from adjacent lanes or far from the lane center, thereby improving the estimation of the initial probability density distribution. This significantly improved the purity of lane center features. It also resulted in the acquisition of a preliminary lateral density distribution. Subsequently, the algorithm identifies local density peaks representing the center of the lane using strict mathematical conditions. A peak must simultaneously satisfy the following conditions: its first derivative is zero (located at a peak), its second derivative is negative, and its density value is not less than a certain proportion of the global maximum density, determined by a significance threshold. =0.2 control. For each detected peak location, the algorithm finds its nearest neighbor from the original candidate points, and the matching distance threshold is 0.1 meters as the final output lane center point, thereby generating a mixed set of lane center points containing all lane centers but not distinguishing which lane they belong to. .

[0065] Step S6, use variational mode decomposition to decompose the lane center point set into... K Each lane center point subset is specifically derived by using variational mode decomposition to obtain a hybrid lane center point set containing information from all lanes. Separate by lane to obtain K Set of center points of each lane Reference process Figure 3 Specifically: Step S61, based on the lane center point set ,along y Statistically analyze the distribution of all points along the axis to generate a probability density function. The calculation formula is: (28) in, n For the set of lane center points The total number of midpoints; The center point of the i-th lane y Axis coordinates.

[0066] Step S62, based on probability density function The probability density function is obtained by using a variational mode decomposition model. Decomposed into K An independent mode The decomposition is achieved by solving the following constraints to optimize the objective function: (29) (30) in, The term represents minimizing the gradient of the modal change; The term represents the calculation of modal components. of Norm; This represents the first regularization parameter; adjustable regularization parameters are used. Weight them, This helps to separate independent lanes in space.

[0067] This indicates a constraint condition, namely that the sum of all modal components must be exactly equal to the original signal. To ensure no information is lost.

[0068] The constrained problem is solved iteratively in the frequency domain using the augmented Lagrange method combined with the alternating direction multiplier method, and can ultimately be decomposed into... K individual modalities .

[0069] Number of lanes K The energy gradient ratio is adaptively determined based on the following criterion: (31) In the formula, , representing belonging to the first k The energy contained in the point set distribution of each lane; This represents all currently identified [items / entities]. K The sum of the energies of each mode, as... K As it increases, this value will approach the total energy of the original signal. ; It is the gradient vector Norm, representing the first norm k Density distribution of candidate lanes The overall fluctuation range; Step S63, setting K =1, start and execute the solution of the energy gradient ratio criterion, and obtain the energy gradient ratio. When two consecutive times Change When the iteration ends, the increment stops. K At this time K This is the optimal estimated number of lanes; Step S64, for each modal component ,turn up K The local maximum values ​​correspond to y The axis position is obtained. K Peak position of each y-axis ; (32) Step S65, for the set of lane center points any point in Calculate its y-axis coordinate to the th k Peak position of each y-axis lateral European distance ; (33) Step S66, based on the lateral Euclidean distance Points where the center points of the lanes are concentrated Assign to the nearest location that is less than the assignment distance threshold. The lane, thus obtaining K A subset of lane center points separated by lane. This provides good input data for subsequent lane fitting modeling.

[0070] To address the challenge of point set aliasing in the lateral space in multi-lane scenarios, this invention further introduces variational mode decomposition (VMD) technology to achieve accurate lane-level separation. The algorithm first separates the points along the longitudinal direction of the road (…). y (Axial direction) for point set Kernel density estimation is performed on the distribution to obtain a signal describing the overall distribution of the multi-lane network. Subsequently, by solving a constrained variational optimization problem, the... Adaptively decomposed into K eigenmode functions The constraint of this optimization problem is that the sum of all modes equals the original signal, thus ensuring no information loss during mode separation. The regularization parameter is... Set it to 0.8 to balance the concentration of modes with the integrity of separation.

[0071] Number of lanes K The adaptive determination is another innovative aspect of this step. The algorithm employs an iterative method based on the energy gradient ratio criterion: from... K Starting with 1, the number of assumed lanes is gradually increased, and the current value is calculated after each decomposition. K The energy gradient ratio γ ( K When γ( ) in two consecutive iterations K When the change is less than the set threshold (0.05), the benefit of adding a new lane is considered no longer significant, and the iteration stops. K This represents the optimal estimated number of lanes. This method effectively avoids the problems of over-decomposition or under-decomposition caused by manually pre-setting the number of lanes. After decomposition, each independent mode... This corresponds to the longitudinal distribution of a lane. The algorithm extracts the peak positions of each mode. This serves as the longitudinal reference for the center of the lane. Finally, based on the "nearest neighbor" principle, the mixed point set is... Each point in the list is assigned to the nearest point that is less than the assignment threshold. The peak value of the lane, thus ultimately outputting K A topologically reliable subset of lane center points separated by lane. This provides high-quality input for subsequent continuous lane line modeling.

[0072] Step S7, based on K A piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center point subsets into a continuous and smooth lane line topology model, enabling the modeling of curves and straight road segments. Specifically, the piecewise cubic polynomial fitting algorithm involves: based on the lane center point subsets separated by lane... Division K Establish a cubic polynomial parameterized model of the lane centerline for each lane sub-section. Specifically, this includes the following steps: Step S71, let the first... k The origin of the local coordinate system for each lane sub-section is... The local coordinate expression of the lane centerline within the lane is: (34) In the formula, Let be the vector of polynomial coefficients to be determined; Step S72: Perform weighted least squares method to optimize polynomial coefficients, optimizing the fitting bias caused by uneven point cloud density. k The objective function for each lane sub-interval is: (35) In the formula, This represents the total number of lane center points falling within the k-th lane sub-section; Let be the coordinates of the center point of the i-th lane corresponding to this lane; The weighting coefficient, which is positively correlated with point cloud density, is calculated using the following formula: ; For the smoothing regularization term, where Represents the curvature term, which mathematically characterizes the degree of curvature of a curve; This represents the second regularization parameter, set to 0.75; Step S73: For each lane, execute the cubic polynomial parameterized model of the lane centerline established in steps S71 and S72 respectively. Use the cubic polynomial fitting method on the lane interval to reconstruct the discrete lane center points into a continuous and smooth lane line topology model, thereby realizing the modeling of curves and straight road segments.

[0073] The flowchart for extracting lane line information and fitting a lane model, and the simulation diagram for lane center extraction are shown below. Figure 4 and Figure 5 As shown. The subset of discrete lane center points output from step S6, which has been precisely separated by lane. This process reconstructs a smooth, continuous, and parameterized lane topology model. This step involves independently and in parallel executing subsequent modeling processes on the point set of each lane, ultimately outputting... KThe model uses parallel lane lines. The core challenge lies in the fact that while these point subsets are precise, they are inherently discrete and non-uniform. A mathematical model that balances fitting accuracy, geometric smoothness, and physical plausibility is needed to reconstruct the continuous road topology hidden behind these points. This step employs a piecewise cubic polynomial weighted least squares fitting algorithm. This algorithm is a complete optimization framework incorporating weighted fitting and smoothness constraints. For the k-th lane interval, a local coordinate system is established within it, using the coordinates of the interval's starting point... The origin is used. Within this lane section, the geometry of the lane centerline is defined by... A cubic polynomial is used for parameterization. Further, weighted least squares is applied to optimize the coefficients, mitigating fitting bias caused by uneven point cloud density. Through mathematical theory and calculation, a minimization problem is solved to obtain the optimal polynomial coefficients for each lane interval. A cubic polynomial fitting technique is then applied to each lane interval to fit discrete lane center points into continuous, smooth lane centerlines, enabling lane line modeling for both curves and straight sections.

[0074] Lane identification and extraction results for straight and curved road sections are as follows: Figure 6 As shown, the comparison results of lane-level error indices in straight and curved road scenarios are as follows: Figure 7 As shown, the optimization process that constitutes the entire lane line extraction and curved road segment identification is the core adaptive mechanism that enables the method to maintain high accuracy and robustness in real-world complex environments.

[0075] Empirical analyses were conducted on datasets for both straight and curved road scenarios. The lane identification and extraction results for straight and curved road sections are as follows: Figure 6 As shown. Figure 6 The left image shows the detection results for straight road sections, successfully identifying 5 lane boundary lines and deriving 4 lane centerlines; Figure 6 The right image shows the analysis results for a curved road section, extracting four lane markings and reconstructing three lane centerlines. In the image, red markers correspond to the lane center coordinates predicted by the algorithm, blue fitted curves represent the lane geometry reconstructed by the mathematical model, black baselines represent manually annotated actual lane information, and black dashed lines represent manually annotated lane centerlines. Figure 7 The comparison of lane-level error indices in straight and curved road scenarios shows that the present invention exhibits significant advantages in different road forms. As shown in the figure, the mean absolute error (MAE) for straight and curved road segments is 0.12m, and the root mean square error (RMSE) is 0.19m, which is superior to traditional methods. This indicates that the millimeter-wave radar-based method for identifying curved road segments and extracting lane lines demonstrates good applicability in road scenarios. The lane lines estimated by this method closely match the actual lane lines, verifying the reliability of the method.

[0076] Finally, based on the aforementioned multi-layer processing and optimization strategies, the proposed method achieves accurate and stable construction of the topology of curved road segments and lane lines by performing dynamic noise filtering, multi-frame fusion enhancement, high-precision extraction of lane centers, and piecewise polynomial modeling on millimeter-wave radar point clouds. Even in complex driving environments with sparse point clouds, abrupt changes in curve curvature, or vehicle occlusion, the method can ensure the robustness and accuracy of lane line extraction through adaptive parameter adjustment, providing reliable lane-level environmental perception capabilities and reliable technical support.

[0077] Example 2: like Figure 8 As shown, based on the same inventive concept as Embodiment 1, this embodiment provides a millimeter-wave radar-based system for identifying curved road sections and extracting lane lines. The method described includes: The data acquisition module is used to acquire point cloud data of road traffic targets through millimeter-wave radar equipment. The point cloud data of road traffic targets includes the target ID number, radial distance, radial velocity and angle of the traffic target relative to the millimeter-wave radar equipment. The data processing module is used to preprocess the point cloud data of road traffic targets based on the range-Doppler joint filtering method, so as to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment; The mesh generation module is used to construct a sliding window mechanism, which divides the lane into rectangular coordinates parallel to the vehicle's direction of travel at fixed intervals, and within each window, it divides the lane into rectangular coordinates perpendicular to the vehicle's direction of travel to form a two-dimensional processing unit. The clustering module is used to statistically model the spatial distribution of point cloud data of moving vehicle targets within each two-dimensional processing unit, statistically analyze the point cloud data distribution within each two-dimensional processing unit, and generate a lane feature map. The lane feature map contains several trajectory point cloud clusters of moving vehicle targets. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle targets relative to the millimeter-wave radar equipment is considered. Adjust the neighborhood radius to perform clustering and extract candidate point clusters for lane centers; The center point extraction module is used to extract the set of lane center points from the cluster of lane center candidate points using an improved lateral variational kernel density estimation algorithm. The lane decomposition module is used to decompose the lane center point set into lanes using variational mode decomposition. K A subset of lane center points; Lane line fitting module, used for based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model for each lane center point subset, thereby realizing the modeling of curves and straight road segments.

[0078] Those skilled in the art will recognize that the modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0079] In the embodiments provided by this invention, it should be understood that the division of modules is only a logical functional division. In actual implementation, there may be other division methods, such as multiple modules can be combined into one module, one module can be split into multiple modules, or some features can be ignored.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for identifying curved road sections and extracting lane lines based on millimeter-wave radar, characterized in that, Includes the following steps: Step S1: Collect point cloud data of road traffic targets using millimeter-wave radar equipment. The point cloud data of road traffic targets includes the target ID number, radial distance, radial velocity, and angle of the traffic targets relative to the millimeter-wave radar equipment. Step S2: The point cloud data of road traffic targets is preprocessed based on the range-Doppler joint filtering method to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment; Step S3: Construct a sliding window mechanism, divide the sliding window at fixed intervals along the lane rectangular coordinate system parallel to the vehicle driving direction, and divide the grid in the lane rectangular coordinate system perpendicular to the vehicle driving direction within each window to form a two-dimensional processing unit; Step S4: Within each two-dimensional processing unit, the spatial distribution of the point cloud data of the moving vehicle target is statistically modeled. The point cloud data distribution within each two-dimensional processing unit is statistically analyzed to generate a lane feature map. The lane feature map contains several trajectory point cloud clusters of moving vehicle targets. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle target relative to the millimeter-wave radar device is considered. Adjust the neighborhood radius to perform clustering and extract candidate point clusters for lane centers; Step S5: Extract the lane center point set from the candidate lane center point cluster using an improved lateral variational kernel density estimation algorithm; Step S6, use variational mode decomposition to decompose the lane center point set into... K A subset of lane center points; Step S7, based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model for each lane center point subset, thereby realizing the modeling of curves and straight road segments.

2. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 1, characterized in that, The step S2, which preprocesses the point cloud data of road traffic targets using a range-Doppler joint filtering method to obtain moving vehicle point cloud data in the lane rectangular coordinate system, specifically includes the following steps: Step S21: The millimeter-wave radar device transmits M linear frequency modulated pulses to the road traffic target within one coherent processing interval. The number of sampling points for each linear frequency modulated pulse is N. The echo signals received by the millimeter-wave radar device constitute an N×M complex matrix. x(n,m), in n Index value representing the distance dimension ，n= 0,1,... ,N-1 ; m The index value represents the dimension of the linear frequency modulated pulse. m= 0,1,... ,M-1 ; Step S22: Perform a Fast Fourier Transform on the echo signal received by the millimeter-wave radar equipment to convert the signal from the time domain to the range domain. The calculation formula is: ； ； In the formula, r This represents the range cell index in the range domain, indicating the radial distance of a road traffic target relative to a millimeter-wave radar device. Among them, distance resolution , c At the speed of light, B For millimeter-wave radar equipment; Indicates the first r The signal of the m-th linear frequency modulated pulse in a distance unit; Step S23: Perform a Fast Fourier Transform on the signal sequence of each range unit in the range domain between different linear frequency modulated pulses to obtain the range-Doppler matrix. The calculation formula is as follows: ； in, Index representing a Doppler cell; For the first r The distance unit is the The distance-Doppler matrix of a Doppler unit; Step S24 involves calculating the phase difference between the range-Doppler matrices of the same range cell and the same Doppler cell in two adjacent frames and comparing it with a set phase threshold to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets. Specifically: No. k Frame and the k The distance-Doppler matrices of the same range cell and the same Doppler cell in the +1 frame are respectively and The time interval between two adjacent frames is Take the complex value at the same position in the distance-Doppler matrix of two adjacent frames, and calculate the phase difference of the complex signal. The calculation formula is: ； in, Indicates taking the conjugate. This indicates a phase take operation. The unit is radians, and the range is... ; For each The unit performs phase difference calculations and makes the following judgments to effectively filter and separate moving targets and static environmental clutter in the point cloud data of road traffic targets; ； in, The set phase threshold; Step S25: Process the statically separated range-Doppler matrix to extract effective point cloud information of the moving target. Specifically, apply the CFAR detection algorithm to the amplitude of the statically separated range-Doppler matrix to filter out effective detection points with energy higher than the background noise, and obtain the coordinates of the effective detection points in the statically separated range-Doppler matrix. ; Step S26, extracting valid detection points The power spectrum is obtained by beamforming the corresponding echo signal in the spatial dimension. Estimate the angle between the moving target and the illumination direction of the millimeter-wave radar equipment. ， Power spectrum The calculation formula is as follows: ； In the formula, The first part represents the millimeter-wave radar equipment. l The root antenna is in the corresponding Complex signals on the unit, for , This refers to the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment. L The total number of antenna channels for a millimeter-wave radar device; measured by power spectrum. The angle relationship corresponding to the maximum peak value: The angle between the moving target and the illumination direction of the millimeter-wave radar equipment is obtained. The estimated value ; Step S27, assuming the location of the millimeter-wave radar equipment is the origin of the coordinate system, within the plane containing the road surface, x The coordinate axes are parallel to the vehicle's direction of travel. y If the coordinate axes are perpendicular to the vehicle's direction of travel, then the spherical coordinates of the moving target are... With respect to the radar rectangular coordinate system The coordinate transformation relationship is as follows: ； in, This indicates the position of the moving target in the radar Cartesian coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving target in the radar Cartesian coordinate system, perpendicular to the vehicle's direction of travel. Step S28, establish a lane rectangular coordinate system, then in the lane rectangular coordinate system With radar coordinate system The coordinate transformation relationship is as follows: ； Finally, each moving target is represented as a quadruple containing position and velocity information. ; Step S29: Perform dynamic noise compensation and motion distortion correction on each moving target to obtain point cloud data of the moving vehicle target in the lane rectangular coordinate system.

3. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 2, characterized in that, The phase threshold set in step S24 The determination method is as follows: Establish phase difference radial velocity relative to the traffic target by the millimeter-wave radar device The specific relationships are as follows: ； in, Indicates the wavelength of the electromagnetic waves emitted by millimeter-wave radar equipment; Set the radial velocity of the traffic target relative to the millimeter-wave radar equipment speed threshold This is used to distinguish between stationary and moving targets; The set phase threshold The calculation method is as follows: 。 4. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 2, characterized in that, Step S29, which involves performing dynamic noise compensation and motion distortion correction on each moving target to obtain point cloud data of the moving vehicle target in the lane rectangular coordinate system, specifically includes the following steps: Step S291, construct the noise power function, as follows: ； in, It refers to the radial velocity of a traffic target relative to the millimeter-wave radar device during measurement. v and radial distance R The noise variance value determined jointly; Step S292, construct a Butterworth filter, wherein the dynamic cutoff frequency of the Butterworth filter is... Specifically as follows: ； in, This refers to the system sampling frequency; ; Step S293: Combine the main energy distribution frequency band of each moving target signal with the dynamic cutoff frequency of the Butterworth filter. Comparison: If the energy of the moving target signal is mainly concentrated at the dynamic cutoff frequency of the Butterworth filter The following frequency bands are considered to represent real moving vehicle targets and are retained. If the energy of the moving target signal is significantly distributed at the dynamic cutoff frequency of the Butterworth filter If any of the above are found, the moving target is determined to be residual noise or a false target and is therefore eliminated. Step S294: Following step S293, valid moving vehicle targets are filtered out, and each moving vehicle target is ultimately represented as a quadruple containing position and velocity information. , representing the point cloud data of moving vehicle targets in the lane rectangular coordinate system.

5. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 1, characterized in that, Step S4 employs an adaptive Euclidean clustering algorithm to process trajectory point cloud clusters within a two-dimensional processing unit, and adjusts the neighborhood radius based on the relative distance to the moving vehicle target for clustering, extracting candidate lane center point clusters. This specifically includes the following steps: Step S41: Calculate the value of each data point within the two-dimensional processing unit using the Gaussian kernel function. Local density and relative distance The calculation formula is: ； In the formula, Represents the i-th data point The local density estimate; Represents the i-th data point The minimum distance to a higher density point; where Let i and j represent the i-th and j-th data points within the two-dimensional processing unit, respectively. This represents the Euclidean distance between the i-th data point and the j-th data point; h The bandwidth parameter represents the Gaussian kernel function; Step S42: Traverse all data points within the two-dimensional processing unit, and calculate the radial distance of each moving vehicle target relative to the millimeter-wave radar device. The neighborhood radius, i.e. the initial cluster radius, is dynamically adjusted, and clustering is performed using the initial cluster radius to obtain several clusters; Step S43: Perform a cluster merging operation based on density similarity on several clusters to obtain the merged clusters; Step S44, for each cluster obtained by merging The points in the middle, according to the timestamp t Sort the points in ascending order, then sort them according to their direction of travel in the lane Cartesian coordinate system, forming a point sequence. The representative points of each cluster are calculated using a weighted average method based on local density. coordinates That is, the candidate point for the center of the lane, the representative point. The formula for calculating the coordinates is: Calculate representative points x Axis coordinates : ； Calculate representative points y Axis coordinates : ； in, express The number of points in For point The local density estimate.

6. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 5, characterized in that, Step S5, which uses an improved lateral variational kernel density estimation algorithm to extract the lane center point set from the candidate lane center point cluster, specifically includes the following steps: Step S51, for each data point in the candidate cluster of lane centers x The probability density estimation of the axis position distribution is calculated using the following formula: ； In the formula, Indicates that the data points are in x The probability density estimate at point ; n This indicates the total number of candidate center points for the lane; Represents the Gaussian kernel function; Indicates the first i Lane center candidate point x Axis coordinates; Indicates an indicator function, when the condition is met. The value is 1 if the condition is met, and 0 otherwise. Represents all data points within a two-dimensional processing unit. x Average of axis coordinates; Indicates lane width; h The bandwidth parameter represents the Gaussian kernel function; Step S52: Calculate the probability density distribution for each data point in the candidate lane center point cluster, where the lane center point corresponds to the probability density distribution. The local maxima are identified, and the local density peaks are used as the final lane center points, thus obtaining the lane center point set. The detection conditions for the local density peaks are as follows: ； In the formula, for The first derivative, for The second derivative, The parameter used to test the significance of density; Step S53: Obtain the x-axis position of the local density peak point detected in each lane center candidate point cluster, satisfying the detection conditions for local density peak points. The x-axis position is selected from the original candidate lane center points and the x-axis position is compared with the x-axis position. Points whose distance to the center point is less than a set x-axis distance threshold are determined as the final lane center points, thus obtaining the lane center point set. .

7. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 6, characterized in that, Bandwidth parameter of Gaussian kernel function h Based on the total number of target points of moving vehicles in step S2 Adaptive adjustment, the specific correspondence is: 。 8. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 6, characterized in that, In step S6, variational mode decomposition is used to decompose the lane center point set into... K Specifically, the lane center point subset is obtained by using variational mode decomposition to extract the mixed lane center point set containing information from all lanes. Separate by lane to obtain K set of center points of each lane Specifically: Step S61, based on the lane center point set ,along y Statistically analyze the distribution of all points along the axis to generate a probability density function. The calculation formula is: ； in, n For the set of lane center points The total number of midpoints; The center point of the i-th lane y Axis coordinates; Step S62, based on probability density function The probability density function is obtained by using a variational mode decomposition model. Decomposed into K An independent mode The decomposition is achieved by solving the following constraints to optimize the objective function: ； ； in, The term represents minimizing the gradient of the modal change; The term represents the calculation of modal components. of Norm; This represents the first regularization parameter; Number of lanes K Determined based on the following energy gradient ratio criterion: ； In the formula, , representing belonging to the first k The energy contained in the point set distribution of each lane; This represents all currently identified... K The total energy of all modes; It is the gradient vector Norm, representing the first norm k Density distribution of candidate lanes The overall fluctuation range; Step S63, setting K =1, start and execute the solution of the energy gradient ratio criterion, and obtain the energy gradient ratio. When two consecutive times Change When the iteration ends, the increment stops. K At this time K This is the optimal estimated number of lanes; Step S64, for each modal component ,turn up K The local maximum values ​​corresponding to y The axis position is obtained. K Peak position of each y-axis ; Step S65, for the set of lane center points any point in Calculate its y-axis coordinate to the th k Peak position of each y-axis lateral European distance ; Step S66, based on the lateral Euclidean distance Points where the center points of the lanes are concentrated Assign to the nearest location that is less than the assignment distance threshold. The lane, thus obtaining K A subset of lane center points separated by lane. .

9. The method for identifying curved road sections and extracting lane lines based on millimeter-wave radar according to claim 1, characterized in that, In step S7, based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model using a subset of lane center points. The specific steps for modeling curves and straight road segments are as follows: Step S71, let the first... k The origin of the local coordinate system for each lane sub-section is... The local coordinate expression of the lane centerline within the lane is: ； In the formula, Let be the vector of polynomial coefficients to be determined; Step S72, perform weighted least squares method to optimize polynomial coefficients, the first... k The objective function for each lane sub-interval is: ； In the formula, This represents the total number of lane center points falling within the k-th lane sub-section; Let be the coordinates of the center point of the i-th lane corresponding to this lane; The weighting coefficient, which is positively correlated with point cloud density, is calculated using the following formula: ; For the smoothing regularization term, where Represents the curvature term, which mathematically characterizes the degree of curvature of a curve; This represents the second regularization parameter; Step S73: For each lane, execute the cubic polynomial parameterized model of the lane centerline established in steps S71 and S72 respectively. Use the cubic polynomial fitting method on the lane interval to reconstruct the discrete lane center points into a continuous and smooth lane line topology model, thereby realizing the modeling of curves and straight road segments.

10. A system for identifying curved road sections and extracting lane lines based on millimeter-wave radar, characterized in that, The method applied to any one of claims 1 to 9 includes: The data acquisition module is used to acquire point cloud data of road traffic targets through millimeter-wave radar equipment. The point cloud data of road traffic targets includes the target ID number, radial distance, radial velocity and angle of the traffic target relative to the millimeter-wave radar equipment. The data processing module is used to preprocess the point cloud data of road traffic targets based on the range-Doppler joint filtering method, to obtain the point cloud data of moving vehicle targets in the lane rectangular coordinate system, where each moving vehicle target is represented as a quadruple containing position and velocity information. ,in, This represents the position of the moving vehicle target in the lane's rectangular coordinate system, parallel to the vehicle's direction of travel. This indicates the position of the moving vehicle target in the lane's rectangular coordinate system, perpendicular to the vehicle's direction of travel. This represents the radial velocity of the moving vehicle target relative to the millimeter-wave radar equipment. This indicates the radial distance between a moving vehicle target and the millimeter-wave radar equipment; The mesh generation module is used to construct a sliding window mechanism, which divides the lane into rectangular coordinates parallel to the vehicle's direction of travel at fixed intervals, and within each window, it divides the lane into rectangular coordinates perpendicular to the vehicle's direction of travel to form a two-dimensional processing unit. The clustering module is used to statistically model the spatial distribution of point cloud data of moving vehicle targets within each two-dimensional processing unit, statistically analyze the point cloud data distribution within each two-dimensional processing unit, and generate a lane feature map. The lane feature map contains several trajectory point cloud clusters of moving vehicle targets. An adaptive Euclidean clustering algorithm is used to process the trajectory point cloud clusters within the two-dimensional processing unit, and the radial distance of the moving vehicle targets relative to the millimeter-wave radar equipment is considered. Adjust the neighborhood radius to perform clustering and extract candidate point clusters for lane centers; The center point extraction module is used to extract the set of lane center points from the cluster of lane center candidate points using an improved lateral variational kernel density estimation algorithm. The lane decomposition module is used to decompose the lane center point set into lanes using variational mode decomposition. K A subset of lane center points; Lane line fitting module, used for based on K The piecewise cubic polynomial fitting algorithm is used to reconstruct the discrete lane center points into a continuous and smooth lane line topology model for each lane center point subset, thereby realizing the modeling of curves and straight road segments.

Images