A road boundary line reconstruction method and system fusing trajectory multi-feature regression

Abstract

The application provides a road boundary reconstruction method and system based on fusion trajectory multi-feature regression, and relates to the technical field of automatic driving environment perception. The method comprises the following steps: obtaining vehicle trajectory flow clusters in a target road section; extracting a trajectory multi-feature vector comprising a vehicle type weighted lateral distribution variance, an average vehicle speed, a road curvature and a trajectory point density; determining a free traffic, a constrained traffic or an abnormally sparse traffic state based on the vector; constructing a corresponding trajectory regression boundary, a substitute boundary or a geometric prior boundary candidate according to the state; and finally, selecting, fusing and rolling back the candidate boundary based on the traffic state, and outputting a continuous road boundary line. The application realizes the coordinated improvement of the accuracy, robustness and safety of road boundary reconstruction in complex scenes such as visual failure by deeply mining trajectory features and intelligently judging their reliability.

Description

Technical Field

[0001] This invention relates to the field of environmental perception technology for autonomous driving, and in particular to a method and system for reconstructing road boundary lines by fusing trajectory multi-feature regression. Background Technology

[0002] In engineering practice in the field of autonomous driving environmental perception, the acquisition of road boundaries mainly relies on the following technical approaches: First, image semantic segmentation based on vision, inferring road structure by recognizing road surface lane line texture; second, lane line extraction methods based on features such as lidar echo intensity; and third, centerline estimation based on vehicle trajectory clustering or map priors, and generating road boundaries through methods such as "fixed lane width expansion".

[0003] However, existing technologies have significant limitations when dealing with complex real-world scenarios. First, vision-based methods struggle to function in environments where vision is impaired, such as heavy rain, snow cover, or strong glare at night. Furthermore, snow and water can interfere with the echo characteristics of lidar, leading to instability or failure of intensity-based detection methods. Second, in road scenarios with worn, missing, or unmarked lane lines, all algorithms relying on explicit lane line features face fundamental challenges. Finally, while some trajectory-based methods do not rely on vision, they typically extrapolate from a fixed lane width assumption when generating boundaries. This fails to fully consider the complex influence of various dynamic factors such as vehicle type, speed, and road curvature on the trajectory distribution in actual traffic flow. Consequently, the generated road boundaries often fail to accurately reflect the real, dynamic traversable area, making it difficult to meet the stringent requirements of autonomous driving for accuracy and safety redundancy in perception results. Summary of the Invention

[0004] To address the technical problem mentioned above, where traditional methods cannot reconstruct road boundaries stably and accurately in scenarios such as visual failure or missing lane lines, this invention introduces three collaboratively improved technical features: trajectory multi-feature extraction based on vehicle type classification, road traffic status determination based on consistency scoring, and boundary candidate construction and fusion based on state selection. This provides a road boundary reconstruction method and system that integrates trajectory multi-feature regression, aiming to overcome the problem of insufficient robustness of road boundary perception in complex environments.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for reconstructing road boundary lines by fusing trajectory multi-feature regression, comprising:

[0006] Obtain vehicle trajectory flow clusters within the target road segment;

[0007] Based on the trajectory flow cluster, multiple feature vectors of the trajectory are extracted. The multiple feature vectors of the trajectory include at least the vehicle type-weighted lateral distribution variance of the trajectory, average vehicle speed, road curvature, and trajectory point density.

[0008] Based on the trajectory multi-feature vector, the road traffic status is determined, including free traffic status, constrained traffic status, and abnormal bias or trajectory sparse status.

[0009] Based on the road traffic status, corresponding road boundary candidates are constructed. In the free traffic state, trajectory regression boundary candidates based on nonlinear regression are constructed. In the constrained traffic state, alternative boundary candidates based on the lower bound model of the traffic space are constructed. In the abnormal bias or trajectory sparse state, trajectory-driven modeling is disabled and prior boundary candidates based on weak geometric features or historical stable results are constructed.

[0010] Based on the road traffic status, the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates are selected, merged, and backtracked, and a continuous road boundary line is output.

[0011] Furthermore, vehicle trajectory flow clusters within the target road segment are obtained, specifically including:

[0012] Simultaneously acquire lidar point cloud data, inertial measurement unit data, wheel speed and odometer data, and the pose of the positioning and mapping system;

[0013] A locally consistent coordinate system is constructed based on the fusion of multiple sensor sources, and the lidar point cloud is transformed into the locally consistent coordinate system.

[0014] The transformed point cloud is then subjected to outlier filtering and ground point removal to obtain a non-ground point cloud.

[0015] Vehicle target detection and cross-frame tracking are performed based on the non-ground point cloud to obtain time-series vehicle trajectories;

[0016] Within a sliding time window, the time-series vehicle trajectories of the vehicle itself and surrounding vehicles are accumulated to form the trajectory stream cluster.

[0017] Furthermore, based on the trajectory flow cluster, multiple feature vectors of the trajectory are extracted, specifically including:

[0018] The trajectory flow cluster is divided into multiple road segment units along the road travel direction;

[0019] Vehicle model classification is performed based on the 3D bounding box dimensions of the vehicle target;

[0020] Within each road segment unit, the lateral distribution variance of the trajectory weighted by the vehicle type is calculated based on the vehicle type classification results;

[0021] Within each road segment unit, calculate the average vehicle speed, road curvature, and trajectory point density;

[0022] Construct the trajectory multi-feature vector, which includes the vehicle type weighted trajectory lateral distribution variance, the average vehicle speed, the road curvature, and the trajectory point density.

[0023] Furthermore, based on the aforementioned trajectory multi-feature vector, the road traffic status is determined, specifically including:

[0024] Based on the multiple feature vectors of the trajectory, a trajectory reachability consistency scoring function is constructed. The scoring function is coupled with lateral dispersion, trajectory point density, curve cutting risk factors, heavy vehicle impact and overall trajectory bias index.

[0025] Based on the value range of the scoring function, the road traffic status is determined to be one of the following: free traffic status, constrained traffic status, abnormal bias, or sparse trajectory status.

[0026] Furthermore, based on the road traffic status, corresponding road boundary candidates are constructed, specifically including:

[0027] When the road traffic status is free traffic, the trajectory multi-feature vector is input into a pre-trained nonlinear regression model to predict the effective traffic half-width of the road, and trajectory regression boundary candidates are generated based on the prediction results.

[0028] When the road traffic state is a constrained traffic state, a lower bound half-width of the traffic space is constructed based on the trajectory multi-feature vector, and alternative boundary candidates are generated based on the lower bound half-width of the traffic space.

[0029] When the road traffic status is abnormally biased or the trajectory is sparse, weak geometric features are extracted from the lidar point cloud or historical stable boundaries are called to generate prior boundary candidates.

[0030] Furthermore, based on the road traffic status, the candidate trajectory regression boundary, the candidate alternative boundary, and the candidate prior boundary are selected, fused, and backed down, and a continuous road boundary line is output, specifically including:

[0031] Based on the road traffic status, the dominant boundary is uniquely determined from the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates.

[0032] When the dominant boundary is either the trajectory regression boundary candidate or the alternative boundary candidate, the prior boundary candidate is used to perform stability correction or continuity constraint correction on the dominant boundary based on the trajectory-driven confidence weight.

[0033] When the dominant boundary is the prior boundary candidate, the trajectory regression boundary candidate and the alternative boundary candidate are explicitly disabled;

[0034] The corrected dominant boundaries of each road segment unit are longitudinally spliced ​​and smoothed along the road travel direction to output a continuous road boundary line.

[0035] A second aspect of the present invention provides a road boundary line reconstruction system that integrates trajectory multi-feature regression, comprising:

[0036] The trajectory flow cluster acquisition module is used to acquire vehicle trajectory flow clusters within the target road segment;

[0037] The multi-feature vector extraction module is used to extract multiple feature vectors of the trajectory based on the trajectory flow cluster. The multiple feature vectors of the trajectory include at least the vehicle type-weighted lateral distribution variance of the trajectory, the average vehicle speed, the road curvature, and the trajectory point density.

[0038] The traffic status determination module is used to determine the road traffic status based on the trajectory multi-feature vector. The road traffic status includes free traffic status, constrained traffic status, and abnormal bias or trajectory sparse status.

[0039] The boundary candidate construction module is used to construct corresponding road boundary candidates according to the road traffic status. Specifically, in the free traffic state, trajectory regression boundary candidates based on nonlinear regression are constructed; in the constrained traffic state, alternative boundary candidates based on the lower bound model of the traffic space are constructed; and in the abnormal bias or trajectory sparse state, trajectory-driven modeling is disabled and prior boundary candidates based on weak geometric features or historical stable results are constructed.

[0040] The boundary fusion output module is used to select, fuse, and backtrack the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates based on the road traffic status, and output continuous road boundary lines.

[0041] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in the first aspect of the present invention.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in the first aspect of the present invention.

[0043] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in the first aspect of the present invention.

[0044] Compared with existing technologies, the road boundary line reconstruction method and system provided by this invention, which integrates trajectory multi-feature regression, has the following beneficial effects:

[0045] (1) By employing vehicle type-weighted lateral distribution variance of the trajectory and constructing multiple feature vectors by combining average vehicle speed, road curvature, and trajectory point density, the feature representation accuracy and robustness of the road passability range are significantly improved. Classifying vehicles based on their three-dimensional dimensions and applying differentiated weighting to lateral statistics ensures that the features accurately reflect the lane-occupancy impact of different vehicle types in mixed traffic flows. Simultaneously, vehicle speed and curvature are integrated to quantify and compensate for the curve-cutting effect, avoiding misjudging vehicle dynamics as road physical contraction. Trajectory point density provides a direct measure of data reliability. The synergy of these features makes the information extracted from the trajectory more comprehensive and more consistent with physical reality, providing a reliable data foundation for subsequent decision-making.

[0046] (2) By introducing a traffic status determination mechanism based on multiple feature vectors and constructing differentiated boundary candidate generation strategies according to different states, the transformation of road boundary reconstruction from a fixed algorithm to situation-adaptive approaches has been realized, enhancing the system's intelligence level and fault tolerance in complex scenarios. The designed consistency scoring function integrates multi-dimensional features into a comprehensive index to evaluate the reliability and representativeness of trajectory statistics, and classifies traffic status into free, constrained, or abnormal states accordingly. For each state, the system uses different boundary generation logics in parallel: the free state uses a regression model to pursue accurate estimation, the constrained state uses a conservative lower bound model of traffic space to ensure a safety baseline, and the abnormal state relies entirely on geometric priors. This paradigm of evaluation before execution enables the system to proactively identify and properly handle various challenging conditions such as snow accumulation, construction, and sparse traffic flow.

[0047] (3) By establishing a unique selection rule for the dominant boundary based on the passage state and setting insurmountable geometric hard constraints under constrained conditions, the clarity of the final boundary output logic, the certainty of the decision, and the rigidity of the safety level are ensured. The system strictly selects the dominant output from multiple candidate boundaries based on the determined state, rather than performing a simple weighted fusion. Under constrained conditions, the alternative boundary generated by the lower bound model of the passage space is given the highest priority, and no final boundary result may intrude into the minimum safe area defined by the model, thus establishing absolute safety at the geometric level. Other boundary candidates can only be used for smoothing or stability fine-tuning without violating the dominant logic and hard constraints. This mechanism effectively eliminates the ambiguity or dangerous output that may be caused by the fuzzy fusion of multi-source information. Attached Figure Description

[0048] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0049] Figure 1 This is a flowchart of a road boundary line reconstruction method based on multi-feature regression of fused trajectory provided in Embodiment 1 of the present invention;

[0050] Figure 2 This is a comparative diagram of the failure of traditional methods and trajectory flow distribution under visual failure conditions provided in Embodiment 1 of the present invention. Figure 2 In the diagram, A represents the detection failure of the traditional method under visual impairment. Figure 2 B in the diagram represents a schematic representation of the boundary line of a passable road that can be inferred from the trajectory flow.

[0051] Figure 3 A schematic diagram of the shearing effect trajectory provided in Embodiment 1 of the present invention (the trajectory narrows towards the inner path as a whole).

[0052] Figure 4 This is a comparison diagram of the continuity of the road segment unit boundary before and after smoothing, provided in Embodiment 1 of the present invention. Figure 4 In the diagram, A represents the road segment boundary line segment before smoothing. Figure 4 B in the diagram is a smoothed representation of the road segment boundary line.

[0053] Figure 5 This is an architecture diagram of the road boundary line reconstruction system based on multi-feature regression of fused trajectory provided in Embodiment 2 of the present invention. Detailed Implementation

[0054] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0055] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0056] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0057] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0058] Terminology Explanation:

[0059] 1. Visual failure scenario: The camera cannot stably output lane lines / boundaries (heavy rain, snow cover, strong glare, low light at night, etc.).

[0060] 2. Implicit Lane Keeping: This refers to the behavior of human drivers who, when unable to clearly see lane lines, subconsciously keep their vehicle in the "virtual center" of the road based on the road edge, the trajectory of the vehicle in front, and driving experience.

[0061] 3. Lateral Distribution Variance: This refers to the degree of lateral dispersion of the trajectory points of multiple vehicles relative to the fitted centerline on a specific road section. This index is positively correlated with the physical width of the lane.

[0062] 4. Oversized Transport Vehicles: The minimum dimensions of vehicles are based on the Chinese standard GA802—2019 "Road Traffic Management Motor Vehicle Types," typical announced vehicle types (such as large trucks, large buses, semi-trailers, and container trucks), and are combined with the US FHWA's minimum size definition for Single-Unit Trucks (SUTs) and the EU Directive 96 / 53 / EC's regulations on the external dimensions of heavy vehicles. A vehicle is designated as an "oversized transport vehicle" if it meets any of the following conditions: total length L ≥ 6.0m; total width W ≥ 2.3m; or total height H ≥ 3.0m.

[0063] 5. Weak Geometric Features: These refer to geometric feature points or sets in LiDAR point clouds that have insignificant reflection intensity, unclear texture or semantic features, but exhibit continuous or locally consistent height differences, slope variations, or structural abrupt changes in spatial geometry. These weak geometric features typically correspond to road edges, shoulder steps, snow accumulation edges, construction access boundaries, or other road geometric structures with insignificant reflection. The extraction of these weak geometric features can be based on height gradients in the point cloud, local elevation abrupt changes, normal variation rates, surface continuity analysis, or combinations thereof. This invention does not limit the specific feature calculation operators, threshold forms, or implementation procedures.

[0064] 6. Effective Width Regression: Effective width refers to the actual lateral width of the road that vehicles can safely pass under the current environmental and traffic conditions (it can be smaller than the physical width, such as a snow trough). It involves calculating the most probable physical width of the current road segment by inputting statistical features of vehicle trajectories and using a machine learning model.

[0065] 7. Confidence: A quantification of the credibility of the "trajectory regression boundary", which varies with trajectory density and is used for fusion decision-making.

[0066] Example 1

[0067] like Figure 1 This embodiment provides a method for reconstructing road boundary lines by fusing trajectory multi-feature regression, including:

[0068] S1. Obtain the vehicle trajectory flow cluster within the target road segment;

[0069] S2. Based on the trajectory flow cluster, extract multiple feature vectors of the trajectory. The multiple feature vectors of the trajectory include at least the vehicle type-weighted lateral distribution variance of the trajectory, average vehicle speed, road curvature, and trajectory point density.

[0070] S3. Based on the trajectory multi-feature vector, determine the road traffic status, which includes free traffic status, constrained traffic status, and abnormal bias or trajectory sparse status.

[0071] S4. Based on the road traffic status, construct corresponding road boundary candidates, wherein, in the free traffic state, construct trajectory regression boundary candidates based on nonlinear regression, in the constrained traffic state, construct alternative boundary candidates based on the lower bound model of the traffic space, and in the abnormal bias or trajectory sparse state, disable trajectory-driven modeling and construct prior boundary candidates based on weak geometric features or historical stable results.

[0072] S5. Based on the road traffic status, select, merge, and backtrack the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates, and output a continuous road boundary line.

[0073] S1. Obtain the vehicle trajectory flow cluster within the target road segment, specifically including:

[0074] S1.1 synchronously acquires lidar point cloud data, inertial measurement unit data, wheel speed odometer data, and the pose of the positioning and mapping system.

[0075] At any moment Simultaneously collect the following multi-source sensor data:

[0076] S1.1.1 LiDAR Point Cloud Data

[0077] Collection time The lidar point cloud data is represented as:

[0078]

[0079] in, No. The three-dimensional spatial coordinates of a laser point in the radar coordinate system This represents the number of point cloud points at that moment.

[0080] S1.1.2 Inertial Measurement Unit (IMU) Data

[0081] IMU at all times The output angular velocity and linear acceleration observations are expressed as follows:

[0082]

[0083] The IMU data is used to describe the attitude change trend and motion continuity of the vehicle over a short time scale.

[0084] S1.1.3 Wheel Odometry Data

[0085] Wheel speed odometer at time Output the linear velocity and heading changes of the vehicle in the vehicle coordinate system:

[0086]

[0087] in, The forward speed of the vehicle. The heading angle increment between adjacent time points is used to constrain the vehicle's translational scale and heading drift. The wheel speed odometer data is used to provide velocity scale constraints for vehicle motion, suppressing drift caused by pure inertial integration.

[0088] S1.1.4 Simultaneous Localization and Mapping (SLAM) System Pose Output

[0089] SLAM system at time Output low-frequency but globally consistent pose observations:

[0090]

[0091] in, Represents the rotation matrix. This represents the translation vector. The SLAM pose is used to provide a low-frequency but globally consistent pose reference.

[0092] S1.2 Construct a locally consistent coordinate system based on multi-source sensor fusion, and transform the lidar point cloud to the locally consistent coordinate system.

[0093] S1.2.1 Vehicle State Prediction and Multi-Source Observation Correction. To provide a unified representation of the above multi-source motion information, the vehicle state at time... The state vector is:

[0094]

[0095] in, For the vehicle in a locally consistent coordinate system The position vector below; the locally consistent coordinate system The reference coordinate system is updated as the vehicle moves, and its origin and attitude are determined at the current moment by the multi-source fusion positioning results. The vehicle speed is a forward velocity scalar, the initial scale of which is provided by the wheel speed odometer; The vehicle attitude parameters (such as Euler angles or equivalent attitude representations) are obtained by integrating the IMU angular velocity.

[0096] Between adjacent time points, the vehicle state is updated using the angular velocity and acceleration data from the IMU for short-term prediction.

[0097]

[0098] in, Indicates the predicted state. The function represents the state propagation process based on the vehicle kinematics model, used to generate prior estimates of the vehicle state.

[0099] Obtaining the predicted state Subsequently, wheel speed odometer and SLAM pose were introduced as observations to correct the state.

[0100] Within the state estimation framework, it is necessary to determine "if the vehicle state is..." based on the predicted state. "The theoretically observable result of the wheel speed odometer", theoretical speed observation:

[0101]

[0102] Modeling relationships:

[0103]

[0104] Therefore, its observation model is expressed as:

[0105]

[0106] in, This indicates the wheel speed odometer observations; Indicates the state in the prediction Below, the velocity in the vehicle state vector with posture The theoretical observation values ​​obtained by calculation; To address wheel speed odometer observation noise, the actual wheel speed odometer readings are compared with theoretical readings based on predicted conditions. This corrects for velocity scale and heading errors in the predicted conditions, thereby suppressing drift caused by long-term IMU integration.

[0107] Simultaneously, the vehicle pose output by the SLAM system is introduced as another type of observation, theoretical pose observation:

[0108]

[0109] Modeling relationships:

[0110]

[0111] Therefore, its observation model is expressed as:

[0112]

[0113] in, Indicates the SLAM system in a locally consistent coordinate system The vehicle pose observations are output below; Indicates the state in the prediction Below, the position in the vehicle state vector with posture The theoretical pose observation values ​​obtained by calculation; This refers to the noise in SLAM pose observations. The SLAM observations are used to perform low-frequency but globally consistent corrections to vehicle position and attitude to suppress accumulated errors from inertia and wheel speed odometers.

[0114] The predicted state obtained based on the motion model is then converted using the Extended Kalman Filter (EKF) or equivalent state estimation methods. The vehicle state is fused and updated with wheel speed and odometer observations and SLAM pose observations to obtain a posterior estimate. .

[0115] The fusion update process includes: in the prediction state Based on this, the theoretical observation values ​​that each sensor should obtain at the current moment are calculated according to the predicted state, and compared with the actual observation results provided by the wheel speed odometer and SLAM system. According to the magnitude of the prediction error and the credibility of different observation sources, the position, velocity and attitude parameters in the predicted state are corrected, so as to obtain vehicle state estimation results with higher consistency with multi-source observation.

[0116] S1.2.2 Extract the unified vehicle pose. From the updated state vector Extracting the vehicle in a locally consistent coordinate system The following pose, where:

[0117] Vehicle position vector is

[0118]

[0119] The vehicle attitude rotation matrix is ​​obtained from the updated attitude parameters. It is calculated through the mapping relationship from attitude to rotation matrix, and is denoted as:

[0120]

[0121] Thus, the fused vehicle pose estimate is obtained:

[0122]

[0123] Among them, vehicle position Take the radar coordinate system relative to the locally consistent coordinate system The position; This is the fused vehicle attitude rotation matrix. For the vehicle in a locally consistent coordinate system The position vector below.

[0124] The This serves as the basis for subsequently constructing a locally consistent coordinate system and completing the pose representation of point clouds across time.

[0125] The vehicle pose obtained by fusing at the current moment Defined locally consistent coordinate system The coordinate system has the vehicle's current position as its origin, and the coordinate axes are aligned with the vehicle's heading.

[0126] Point cloud acquired by lidar in the radar coordinate system:

[0127]

[0128] Mapping from pose transformation to a locally consistent coordinate system :

[0129]

[0130] This method projects point cloud data from different time points and sources onto the same local reference frame, ensuring that a spatially consistent road structure representation can still be obtained even under visual failure conditions.

[0131] S1.3 Perform outlier filtering and ground point removal on the transformed point cloud to obtain a non-ground point cloud; perform vehicle target detection and cross-frame tracking based on the non-ground point cloud to obtain the time-series vehicle trajectory.

[0132] S1.3.1 Point Cloud Unification and Denoising. To address outlier interference caused by environmental factors such as rain, snow, splashing, and water mist, a local consistency coordinate system is used. Point cloud data Perform Statistical Outlier Removal (SOR) to remove points whose distance from the mean of their neighbors is significantly different, resulting in a cleaned point cloud:

[0133]

[0134] in, This represents the point cloud data points after denoising. This step improves the stability of the point cloud spatial structure, providing reliable input for subsequent weak geometric edge extraction. The weak geometric edges are as described in the terminology explanation section.

[0135] S1.3.2 Ground point removal. This is done after obtaining the purified point cloud. Next, ground point removal and road plane separation are performed. RANSAC plane fitting is then used for ground model estimation.

[0136]

[0137] Will satisfy Points that are identified as ground points are discarded to obtain a non-ground point cloud. .in, For planar parameters, This is the ground distance threshold.

[0138] S1.3.3 Vehicle Detection and Tracking. For non-terrestrial point clouds... Vehicle target extraction can be performed using PointPillars for learned 3D target detection: a point cloud 3D detection network outputs a set of 3D bounding boxes. .

[0139] For cluster point set detection boxes Calculate its 3D bounding box dimensions:

[0140]

[0141] in, These represent the target's length, width, and height, respectively.

[0142] To generate temporally consistent targets suitable for trajectory flow construction, cross-frame association and tracking of the targets are further performed. Kalman filtering based on a uniform motion model can be used for state prediction.

[0143]

[0144] in, For the first The target status of a vehicle (position, speed, etc.). Here is the state transition matrix. This is process noise.

[0145] The Hungarian algorithm is then used to match the predicted target with the current target based on the cost matrix (e.g., center distance / IOU) to obtain the vehicle target trajectory. A continuous observation sequence.

[0146] S1.3.4 Preliminary Vehicle Classification. Based on this, the target vehicles can be preliminarily grouped according to preset size thresholds: when the following conditions are met... or or When the target is in use, mark it as a large transport vehicle; otherwise, mark it as a small / medium-sized vehicle. This is the threshold parameter.

[0147] S1.4 Within the sliding time window, accumulate the temporal vehicle trajectories of the vehicle itself and surrounding vehicles to form the trajectory stream cluster. Based on the obtained vehicle target temporal trajectory... The maintenance length is A sliding time window (e.g., 10 minutes) is used to uniformly project the trajectory data of the vehicle and surrounding vehicles acquired through vehicle-to-everything (V2X) communication or onboard perception systems onto a locally consistent coordinate system within this time window. In the process, trajectory flow clusters are formed:

[0148]

[0149] in, Indicates time This is a collection of all vehicle trajectory points acquired internally. By accumulating multi-time-lapse and multi-vehicle trajectories, it enhances the indirect perception of road structure under conditions where lane lines are missing or invisible. Trajectory stream cluster output. It will be used as the input to S2.

[0150] S2. Based on the trajectory flow cluster, extract multiple feature vectors of the trajectory, specifically including:

[0151] S2.1 Divide the trajectory flow cluster into multiple road segment units along the road travel direction. (e.g., 5m) is divided into units :

[0152]

[0153] In each road segment unit Within this space, the set of trajectory points that fall within this spatial range converges to form a corresponding subset of trajectories. .

[0154] S2.2 Vehicles are classified based on the 3D bounding box size of the vehicle target; within each road segment unit, the vehicle type weighted trajectory lateral distribution variance is calculated based on the vehicle type classification results.

[0155] S2.2.1 Calculate basic statistics. Calculate road segment units. All trajectory points within To the center line lateral distance Mean lateral distance and the unweighted lateral distribution variance of the trajectory .

[0156] Specifically, for road segment units Inner trajectory points The lateral distance from the centerline is defined as:

[0157]

[0158] in, For road segment units Inner A trajectory point; This is the shortest distance operator from a point to a curve; This represents the lateral distance of the trajectory point relative to the centerline (used in this invention to characterize the lateral lane occupation of vehicles and the passable width of the road). Further, it can be... The unit normal direction at that point is defined as And by marking the left and right sides Extended to a tagged lateral offset to distinguish between left-side and right-side lane encroachment; the tagged lateral offset is still denoted as... Its absolute value is consistent with the aforementioned shortest distance.

[0159] The mean lateral distance is defined as:

[0160]

[0161] in, For road segment units Number of internal trajectory points; This represents the mean horizontal distance.

[0162] The variance of the lateral distribution of the trajectory is defined as:

[0163]

[0164] in, For road segment units The variance of the lateral distribution of the inner trajectory is used to measure the degree of lateral dispersion of the trajectory relative to the center line.

[0165] S2.2.2 Vehicle Classification Statistics. Based on the vehicle classification results in S1.3.4, the proportion of oversized transport vehicles is statistically analyzed. and the proportion of small / medium-sized vehicles .

[0166] Specifically, based on the target detection results of LiDAR point cloud, the 3D bounding box size features of the vehicle point cloud are extracted, including length, width, and height information. Vehicles are then classified into two vehicle types according to a preset threshold rule: small and medium-sized vehicles, and heavy-duty transport vehicles. The proportion of each vehicle type in the trajectory stream is statistically analyzed.

[0167]

[0168]

[0169] in, This indicates the number of vehicles used for transporting oversized items. This indicates the number of small and medium-sized vehicles. This represents the total number of vehicles participating in the trajectory statistics. The vehicle type ratio reflects the impact of different types of vehicles on lateral lane-occupancy behavior on the road.

[0170] S2.2.3 Calculate the weighted variance. Arrange the trajectory points according to vehicle type. Grouping vehicles into {small / medium-sized vehicles, oversized transport vehicles} and calculating the vehicle type-weighted lateral distribution variance:

[0171]

[0172] in:

[0173] Represents road segment unit It belongs to the vehicle category The number of trajectory points;

[0174] Indicates vehicle type The corresponding lateral weights are used to reflect the differences in lateral lane occupation and driving stability among different vehicle models;

[0175] Indicates vehicle type The The trajectory points to the center line The horizontal distance;

[0176] Indicates vehicle type The mean lateral distance.

[0177] The weighted lateral distribution variance of vehicle type is the core input for subsequent feature vector construction and traffic status determination.

[0178] The vehicle type weight can be based on the proportion of heavy-duty transport vehicles. Configure adaptive settings:

[0179] ,

[0180] in, This is the weighting adjustment coefficient. The percentage of oversized transport vehicles is used to reflect the impact of heavy vehicles on lateral lane occupation in mixed traffic flow.

[0181] Vehicle-weighted lateral distribution variance To enhance the ability to depict the effective traffic boundaries of roads under different vehicle combination scenarios.

[0182] S2.3 Within each road segment unit, calculate the average vehicle speed, road curvature, and trajectory point density.

[0183] Specifically, the average speed of vehicles on this road section:

[0184]

[0185] in, For the first The instantaneous speed of the vehicle corresponding to each trajectory point; For road segment units Average speed of vehicles inside the vehicle.

[0186] The density of trajectory points is defined as:

[0187]

[0188] in, The length of the road segment unit. The preset reference width is used to normalize the number of trajectory points to a density index per unit area. Its value is independent of the actual physical width of the road and can be set as a constant. The reference width is only used for density normalization, and its specific value can be determined according to the local lane width. It does not affect the subsequent traffic status determination and boundary generation logic. It is used to measure the sufficiency of trajectory samples within the road segment unit and serves as input for subsequent gating and fusion.

[0189] Calculate curvature (from the local centerline) The curvature estimation of the fitted curve on the unit is obtained, which is used to characterize the degree of road curvature and for subsequent curve deviation compensation and condition determination.

[0190] S2.4 Construct the trajectory multi-feature vector, which includes the vehicle type weighted trajectory lateral distribution variance, the average vehicle speed, the road curvature, and the trajectory point density.

[0191] Specifically, the trajectory multi-feature vector construction:

[0192]

[0193] in, The trajectory features are multi-feature vectors, which serve as inputs for subsequent S3 traffic status determination and S4 width estimation / boundary generation.

[0194] S3. Based on the aforementioned trajectory multi-feature vector, determine the road traffic status, specifically including:

[0195] S3.1 Based on the multiple feature vectors of the trajectory, a trajectory reachability consistency scoring function is constructed. The scoring function is coupled with lateral dispersion, trajectory point density, curve cutting risk factors, heavy vehicle impact, and overall trajectory bias index.

[0196] Specifically, in road segment units Within this framework, based on the statistics obtained in S2.3, a trajectory reachability consistency scoring function is constructed. :

[0197]

[0198] in, For road segment units The trajectory can achieve a consistency score; Weighted lateral distribution variance for vehicle models; It is a lateral discrete normalization scale parameter used to normalize the lateral distribution variance of different road segments to a uniform order of magnitude. Its value can be determined based on the statistical mean of historical free passage samples or empirical calibration results. For the density of trajectory points; For the curvature of the road segment; The average vehicle speed per road segment; The proportion of vehicles used for transporting oversized items; This is the adjustment coefficient; This is the overall trajectory bias indicator.

[0199] To identify overall trajectory deviation caused by driving in the wrong lane, detouring around obstacles, or temporarily occupying lanes, an overall trajectory deviation index is defined based on the lateral distance statistics in S2.3.2:

[0200]

[0201] in, This is an indicator of the overall trajectory bias. The mean of the horizontal distance; The variance of the lateral distribution of the trajectory; To prevent extremely small positive numbers with a denominator of zero. When When the value is large, it indicates that the trajectory has shifted significantly to one side of the road. The trajectory statistics may reflect temporary detours or detours, and should not be used directly for road boundary inversion.

[0202] S3.2 Based on the value range of the scoring function, the road traffic status is determined to be one of the following: free traffic status, constrained traffic status, abnormal bias, or sparse trajectory status.

[0203] Specifically, based on the range of values ​​for the comprehensive scoring function, the road traffic status of a road segment unit is determined into one of the following three categories:

[0204]

[0205] in:

[0206] This indicates a free-traffic state, suggesting that vehicle trajectories are highly representative and can be used for road half-width estimation based on statistical regression.

[0207] This indicates a constrained traffic state, suggesting that trajectory statistics have systematically contracted or become restricted. In such cases, an alternative lower bound model for the traffic space should be used to construct candidate boundaries.

[0208] This indicates an abnormal bias or sparse trajectory state, meaning that the trajectory set is not representative of the road structure. Trajectory-driven modeling should be disabled and the model should be rolled back to the prior boundary candidate.

[0209] For trajectory reachability consistency scoring function The state boundary parameter is used to divide the continuous scoring space into three types of traffic state intervals: free passage, constrained passage, and abnormal bias.

[0210] The boundary parameter is not a hard threshold set for a single statistic, but rather a state space partitioning for the comprehensive score after multi-feature coupling. Its specific value can be determined based on the statistical characteristics of historical data or system calibration results. In one implementation, These can be taken as quantile statistics in the historical free-travel samples and the abnormal samples, respectively.

[0211] S3.3 Calculate the trajectory-driven confidence weight. To achieve continuous fusion of different boundary candidate results in subsequent steps and to gating the influence of the trajectory-driven results, a scoring function is used. Constructing trajectory-driven confidence-gated weights:

[0212]

[0213] in, The gate weights represent the confidence level of the trajectory-driven results; This is the gate slope coefficient, used to adjust... Sensitivity to changes in the scoring function; The trajectory reachability consistency score defined for S3.1; This refers to the state boundary parameters in S3.3. When Significantly greater than hour, Taking a smaller value indicates lower trajectory reliability, and the influence of trajectory-driven boundaries in subsequent fusion should be significantly reduced; when No more than hour, A larger value indicates higher trajectory reliability, which can be assigned higher weight in subsequent boundary construction and fusion.

[0214] S4. Based on the road traffic status, construct corresponding road boundary candidates, specifically including:

[0215] S4.1 When the road traffic state is free traffic state, the trajectory multi-feature vector is input into a pre-trained nonlinear regression model to predict the effective traffic half-width of the road, and trajectory regression boundary candidates are generated based on the prediction results.

[0216] Specifically, the generation of road boundary candidates in free-traffic mode ( )

[0217] As a road boundary modeling path applicable under the free-traffic assumption, this invention defines the free-traffic state. The road boundary candidate construction method is as follows: It is assumed that vehicle trajectories are sufficiently distributed in the lateral direction and that there is no significant overall bias, allowing for direct road traffic half-width regression estimation based on multi-vehicle statistical features. The trajectory multi-feature vectors constructed in S2.4 are then used... As input to the regression model, a pre-trained nonlinear regression function is used. Output the estimated half-width of the road (half of the effective traffic width):

[0218]

[0219] in, This represents the effective half-width of the road, obtained based on the statistical characteristics of free-traffic trajectories.

[0220] Based on the aforementioned traffic half-width, with the road segment unit centerline Based on this, symmetrical expansion is performed along the normal direction to generate candidate road boundaries under free-traffic conditions:

[0221]

[0222] in, This indicates that the road boundary is generated by expanding outwards to the left and right sides along the normal direction of the centerline.

[0223] S4.2 When the road traffic state is a constrained traffic state, construct the lower bound half-width of the traffic space based on the trajectory multi-feature vector, and generate alternative boundary candidates based on the lower bound half-width of the traffic space.

[0224] Specifically, candidate generation of alternative road boundaries under constrained traffic conditions ( ).

[0225] As a road boundary modeling path applicable to the assumption of constrained trajectory statistics, this invention defines constrained traffic states. The alternative road boundary candidate construction method is as follows: This invention believes that the statistical distribution of vehicle trajectories in the lateral direction no longer satisfies the modeling assumption that "it can directly reflect the road boundary".

[0226] In this state, the vehicle trajectory may systematically shrink or deviate due to the following reasons: ① Incomplete constraints of the vehicle under curve driving conditions; ② Conservative driving behavior caused by the increased proportion of heavy transport vehicles; ③ Compression of driving lanes caused by environmental constraints (snow compaction, construction lanes).

[0227] Therefore, this invention no longer adopts the road traffic half-width estimation model based on lateral distribution regression, but instead constructs an alternative lower bound model for traffic space in scenarios where trajectory statistics fail, in order to avoid systematic underestimation of the actual road traffic range.

[0228] In S3, when the road segment unit The trajectory that satisfies the consistency scoring function satisfy:

[0229]

[0230] This indicates that the statistical distribution of vehicle trajectories in the lateral direction has systematically contracted or shifted, but is not yet completely ineffective. Continuing to use a road width estimation model based on trajectory lateral distribution regression carries the risk of systematically underestimating the actual passable road area. Under this constrained traffic state, this invention no longer attempts to directly regress the actual road width, but instead constructs a lower bound model for the traffic space to characterize the minimum safe lateral space required for vehicle passage under the current trajectory consistency conditions. Based on the trajectory statistical characteristics obtained in S2.3, road segment units are defined. The lower half-width of the passage space inside is:

[0231] in, This represents the lower half-width of the road passage space under constrained traffic conditions. The square root of the weighted lateral distribution variance for vehicle type is used to characterize the basic lateral lane occupancy of the vehicle in the current road segment. For the curvature of the road segment unit, The average vehicle speed is used to jointly characterize the cornering trend under curve conditions; The percentage of oversized transport vehicles is used to reflect the amplified effect of heavy vehicles' demand for lateral safety space under restricted conditions. The variance of the unweighted lateral distribution of the trajectory; This represents the trajectory point density, used to measure sample sufficiency. This is the adjustment coefficient; This is a numerically stable term used to prevent the variance of the lateral distribution from degenerating to zero in extreme cases. To prevent extremely small positive numbers with a denominator of zero.

[0232] It should be noted that: the aforementioned It is not used to accurately estimate the actual road width, but rather to provide a lower bound for the passage space that is no less than the actual physical lane occupancy requirement of vehicles in scenarios where trajectory statistics fail, thereby avoiding boundary misjudgment caused by trajectory compression under constrained conditions.

[0233] Based on the lower bound of the passage space With the center line Based on this, generate alternative passage boundary candidates along its normal direction:

[0234]

[0235] in, This represents the alternative road boundary result under constrained traffic conditions.

[0236] The alternative boundary result is more conservative than the free state regression result, which can effectively avoid the problem of road boundaries being incorrectly shrunk in scenarios such as snow troughs and construction passages.

[0237] S4.3 When the road traffic status is abnormally biased or the trajectory is sparse, weak geometric features are extracted from the lidar point cloud or historical stable boundaries are called to generate prior boundary candidates.

[0238] Specifically, prior boundary candidate generation under anomalous bias or sparse trajectory states ( ).

[0239] As a road boundary modeling path applicable to assumptions of trajectory anomalies or sparse samples, this invention defines anomaly states. The prior boundary candidate construction method is as follows: This invention believes that the current trajectory set does not have the statistical representativeness of road boundaries, and continuing to use any trajectory-based width estimation or boundary regression model may introduce systematic errors.

[0240] Therefore, under this condition, the present invention explicitly prohibits the use of trajectory-driven road boundary modeling results, including but not limited to: ① regression width based on lateral distribution statistics; ② modified or weighted models based on trajectory features.

[0241] In this state, the geometric edge detection results or historical stable road boundaries are directly used as candidates for prior boundaries. The aforementioned It can be derived from weak geometric edge extraction results, historical frame boundaries, or map priors.

[0242] S4.4 Assign boundary confidence weights.

[0243] Specifically, to support consistent decision-making and continuous processing of multiple road boundary candidates in subsequent S5, this invention is based on the road traffic status output in S3. and trajectory reachability consistency scoring function Confidence weights are assigned to different boundary candidates. Specifically, road segment units are defined. The credibility weights for various boundary candidates are as follows:

[0244]

[0245] in, Indicates the current road segment unit The corresponding trajectory-driven boundary confidence weights; These represent free passage, restricted passage, and abnormal offset or sparse trajectory states, respectively. The trajectory-driven confidence gating weights defined in S3.4 have values ​​that change continuously with the scoring function.

[0246] In free passage status Under these conditions, trajectory statistics are highly representative, and trajectory-driven boundary candidates can be fully trusted; in constrained travel states... Under certain conditions, trajectory statistics are subject to uncertainty and are only allowed to participate in subsequent boundary stability corrections under the lower bound constraint of the travel space; in cases of abnormal bias or sparse trajectory states... In this case, trajectory-driven boundaries are explicitly disabled, and their credibility weight is set to zero. Through the credibility assignment mechanism bound to the traffic status and consistency scoring function, the multiple boundary candidates constructed in S4 have clear priorities and constraints in the subsequent fusion and backoff processes, thereby ensuring that the final road boundary output is consistent with the current road traffic status.

[0247] It should be noted that the various road boundary results described in S4 are all candidate boundaries, and their construction methods exist in parallel at the method level. They are not exclusively prohibited due to the current traffic status of the road segment unit. Whether each candidate boundary is used as the final output is only uniformly selected and rolled back in S5 based on the road traffic status and consistency constraints.

[0248] S5. Based on the road traffic status, select, merge, and backtrack the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates, and output continuous road boundary lines, specifically including:

[0249] In S4, for road segment units Several road boundary candidate results have been defined and constructed, including: road boundary candidates under the free passage assumption. Alternative road boundary candidates under the constrained traffic assumption And prior boundary candidates under the assumptions of anomalous bias or trajectory sparsity. .

[0250] It should be noted that the above boundary results exist in parallel at the methodological level, constituting the candidate boundary set for road segment units:

[0251]

[0252] Whether it is used as the final road boundary output is not directly determined in S4, but rather by the road traffic status obtained in step 3 in this step. And unified decision-making based on consistency constraint rules.

[0253] S5.1 Based on the road traffic status, uniquely determine the dominant boundary among the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates.

[0254] Based on the road traffic status obtained from the segment unit determination in S3, the dominant boundary consistent with this status is uniquely determined from the boundary candidate set, and its mapping relationship is defined as follows:

[0255]

[0256] in, Indicates a free passage status. Indicates a restricted passage status. This indicates an abnormal bias or sparse trajectory state. This indicates the dominant boundary that is uniquely selected within the current road segment unit and used to generate the final road boundary. It is important to emphasize that the above mapping relationship is not a re-determination of the state, but a direct application of the state results already determined in S3, used to clarify the boundary types allowed as the basis for the final output under different traffic conditions.

[0257] S5.2 When the dominant boundary is either the trajectory regression boundary candidate or the alternative boundary candidate, the stability correction or continuity constraint correction of the dominant boundary is performed using the prior boundary candidate based on the trajectory-driven confidence weight; when the dominant boundary is the prior boundary candidate, the trajectory regression boundary candidate and the alternative boundary candidate are explicitly disabled.

[0258] Specifically, at the dominant boundary Given that the dominant boundary is uniquely determined, the remaining boundary candidates will no longer participate in peer-to-peer fusion as parallel output candidates, but will only play a supporting role in the following situations: ① when the dominant boundary is At that time, prior boundary Used only for stability reference to prevent abrupt changes in trajectory regression results due to local noise or short-term anomalies; ② When the dominant boundary is At that time, prior boundary Used only for continuity reference, to suppress discontinuous changes in alternative boundaries between adjacent road segments; ③ When the dominant boundary is, all trajectory-driven boundary results (including and All of these are explicitly disabled and cannot participate in the generation of the final road boundary.

[0259] By defining the roles mentioned above, it is ensured that only one boundary result is used as the final output basis within any road segment unit, thereby avoiding the uncertainty caused by the parallel fusion of multiple source boundaries.

[0260] The stability correction under free passage conditions is as follows:

[0261] when At that time, the dominant boundary is In this state, to avoid boundary jitter caused by local sample fluctuations or noise interference in trajectory statistics, a stability consistency detection mechanism is introduced. The boundary difference measurement function is defined as follows:

[0262]

[0263] in, This measure of the geometric difference between two boundaries can be implemented using either maximum normal deviation or mean square deviation. When:

[0264]

[0265] At that time, it is assumed that the trajectory regression boundary is geometrically consistent with the prior boundary, and is directly taken as follows:

[0266]

[0267] When the above conditions are not met, trajectory-driven confidence gating weights may be introduced without changing the dominant boundary type. Stability correction for the regression boundary:

[0268]

[0269] in, This represents a stability fusion operator that prioritizes the regression boundary and references the prior boundary.

[0270] The constraint consistency output under constrained passage conditions is as follows:

[0271] when At that time, the dominant boundary is .because The lower bound model of the passage space defined in S4.2.1 geometrically represents the minimum safe passage range of a vehicle under the current constraints. Therefore, this step requires that the final road boundary result must strictly satisfy the inviolable constraint of this lower bound. To this end, under constrained passage conditions, this invention describes the generation process of the final road boundary as a constrained optimization problem under prior boundary constraints, with the alternative boundary as the hard constraint lower bound. Its consistency constraint conditions are defined as follows:

[0272]

[0273] in, The centerline of the road segment unit obtained by fitting in S2.3; This indicates that the final road boundary is located along the centerline. Lateral offset in the normal direction; Indicates the replacement of road boundaries The lateral offset corresponding to the same position.

[0274] The above inequality constraint means that at any position along the centerline, the offset of the final road boundary in the normal direction must not be less than the offset of the alternative boundary determined by the lower bound model of the traffic space, thereby ensuring that the final road boundary does not intrude into the minimum traffic space.

[0275] In one implementation, to suppress potential local discontinuities in the alternative boundary between adjacent road segment units, while satisfying the aforementioned hard constraints, a priori boundary is introduced. As a continuity reference, the alternative boundary is smoothly corrected by a constraint consistency operator:

[0276]

[0277] in, This represents a constraint correction operator whose inputs are the alternative boundary and the prior boundary. Under the hard constraint that the lower bound of the passage space is not exceeded, the boundary result is adjusted and the corrected boundary is output. The allowable geometric deviation tolerance is used to limit the smooth adjustment range of the final boundary relative to the alternative boundary and the prior boundary.

[0278] It should be emphasized that, under constrained traffic conditions, the hard constraints take precedence over any boundary correction strategy based on trajectory statistics or regression models; even in the presence of highly credible trajectory-driven boundary candidates, the final road boundary must not violate the aforementioned lower bound constraint of the traffic space.

[0279] S5.3 performs longitudinal splicing and smoothing of the corrected dominant boundaries of each road segment unit along the road travel direction, outputting a continuous road boundary line.

[0280] Specifically, after obtaining the final road boundaries of each road segment unit, the boundary results of adjacent road segment units are longitudinally stitched together along the road travel direction, and a continuity smoothing process is applied to obtain the continuous boundary output of the entire road segment:

[0281]

[0282] in, This represents a smoothing operator that performs longitudinal stitching and continuum processing on discrete road segment boundary results. A cubic method can be used. Spline smoothing or local least squares smoothing can be used to achieve this. This represents the number of road segment units.

[0283] In one specific embodiment, the test vehicle is equipped with an onboard LiDAR, inertial measurement unit, wheel speed odometer, and onboard positioning and mapping system, simultaneously acquiring LiDAR point cloud data and vehicle motion state information. The system constructs a locally consistent coordinate system based on the multi-source sensor fusion positioning results and uniformly maps point cloud data acquired at different times to this coordinate system. To address outlier interference caused by environmental factors such as rain, snow, water mist, and splashing, statistical outlier filtering is performed on the point cloud data under the locally consistent coordinate system, and ground points are removed through ground plane fitting to obtain stable non-ground point cloud data. This processing ensures the consistency and reliability of the point cloud's spatial representation under visual failure conditions, providing a basic input for subsequent vehicle detection and road structure perception.

[0284] Subsequently, using non-terrestrial point clouds as input, the system performs 3D detection of vehicles around the road and outputs the 3D bounding boxes of the vehicles along with their length, width, and height. To obtain temporally consistent vehicle motion information, the system performs cross-frame association and tracking of detected vehicle targets, continuously outputting the driving trajectories of each vehicle. The system maintains a sliding time window, accumulating the driving trajectories of the current vehicle and surrounding vehicles within this window to form a trajectory stream cluster. For example... Figure 2 As shown in Figure A, when snow cover or severely worn lane lines prevent the camera from stably outputting road boundaries, traditional vision-based methods fail to detect them, while... Figure 2 As shown in B, the vehicle's trajectory still forms a stable distribution within the passable area of ​​the road, thus providing implicit constraint information for the road structure.

[0285] Based on this, the system divides the trajectory flow cluster into multiple road segment units along the road travel direction, and fits the local road centerline to the trajectory points within each road segment unit. Vehicles are classified by type based on their 3D bounding box dimensions, the proportion of oversized transport vehicles is statistically analyzed, and statistical features such as the vehicle-weighted lateral distribution variance, average vehicle speed, road curvature, and trajectory point density are calculated within each road segment unit to form the trajectory multi-feature vector for that road segment unit. In curved road sections, vehicles often exhibit curve-cutting behavior, causing the trajectory to converge inwards in the lateral direction, such as... Figure 3 As shown. To avoid misjudging this phenomenon as a reduction in the actual road width, this embodiment introduces a joint curvature and speed correction mechanism in the width estimation process to compensate and correct the regression results, thereby reducing the systematic bias caused by the shearing effect.

[0286] Based on multiple feature vectors of the trajectory, the system determines the road traffic status of each road segment unit and constructs corresponding road boundary candidates according to the traffic status. Under free traffic conditions, a nonlinear regression model is used to predict the effective traffic half-width of the road, and the fitted centerline is used as a reference to expand outwards to both sides to generate trajectory regression boundary candidates. Under constrained traffic conditions, a lower bound model of the traffic space is constructed to generate conservative alternative road boundary candidates. Under trajectory anomalies or sparse sample conditions, trajectory-driven modeling is disabled, and the system regresses to the prior boundary candidates.

[0287] Simultaneously, the system extracts weak geometric feature point sets such as road edges, snow trough edges, and shoulder steps from the point cloud data to form geometric boundary candidates. For snow-covered scenarios, these weak geometric features typically manifest as continuous snow troughs or snow pile edges, whose reflection intensity is not significant, but still exhibit stable geometric abrupt changes in spatial height distribution. Based on the trajectory point density and the continuous length and unit length point density of the weak geometric features, the system assigns corresponding confidence levels to different boundary candidates, and selects, merges, or backs down trajectory regression boundaries, alternative boundaries, and geometric prior boundaries according to the consistency rules of road traffic conditions.

[0288] Finally, the road boundary results output by each road segment unit are longitudinally stitched along the road travel direction, and a continuous road boundary line is output through smoothing. For example... Figure 4 As shown in A, before smoothing, there may be slight broken lines or discontinuities between the boundary segments of each road segment; after smoothing, as shown in Figure A, the smoothing effect is achieved. Figure 4 As shown in B, the transition between adjacent road segments is more natural, significantly improving overall continuity. These continuous road boundary lines can be directly used for path planning, positioning constraints, or dynamic updates of high-precision maps for autonomous vehicles.

[0289] Example 2

[0290] like Figure 5 This embodiment provides a road boundary line reconstruction system that integrates trajectory multi-feature regression, including:

[0291] The trajectory flow cluster acquisition module is used to acquire vehicle trajectory flow clusters within the target road segment;

[0292] The multi-feature vector extraction module is used to extract multiple feature vectors of the trajectory based on the trajectory flow cluster. The multiple feature vectors of the trajectory include at least the vehicle type-weighted lateral distribution variance of the trajectory, the average vehicle speed, the road curvature, and the trajectory point density.

[0293] The traffic status determination module is used to determine the road traffic status based on the trajectory multi-feature vector. The road traffic status includes free traffic status, constrained traffic status, and abnormal bias or trajectory sparse status.

[0294] The boundary candidate construction module is used to construct corresponding road boundary candidates according to the road traffic status. Specifically, in the free traffic state, trajectory regression boundary candidates based on nonlinear regression are constructed; in the constrained traffic state, alternative boundary candidates based on the lower bound model of the traffic space are constructed; and in the abnormal bias or trajectory sparse state, trajectory-driven modeling is disabled and prior boundary candidates based on weak geometric features or historical stable results are constructed.

[0295] The boundary fusion output module is used to select, fuse, and backtrack the trajectory regression boundary candidates, the alternative boundary candidates, and the prior boundary candidates based on the road traffic status, and output continuous road boundary lines.

[0296] Example 3

[0297] Embodiment 3 of the present invention provides an electronic device.

[0298] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in Embodiment 1 of the present invention.

[0299] The detailed steps are the same as those of the road boundary line reconstruction method based on multi-feature regression of fused trajectory provided in Example 1, and will not be repeated here.

[0300] Example 4

[0301] Embodiment 4 of the present invention provides a computer-readable storage medium.

[0302] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in Embodiment 1 of the present invention.

[0303] The detailed steps are the same as those of the road boundary line reconstruction method based on multi-feature regression of fused trajectory provided in Example 1, and will not be repeated here.

[0304] Example 5

[0305] Embodiment 5 of the present invention provides a computer program product.

[0306] A computer program product includes software code, wherein the program in the software code performs the steps of the road boundary line reconstruction method of fusion trajectory multi-feature regression as described in Embodiment 1 of the present invention.

[0307] The detailed steps are the same as those of the road boundary line reconstruction method based on multi-feature regression of fused trajectory provided in Example 1, and will not be repeated here.

[0308] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0309] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0310] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0311] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0312] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A road boundary line reconstruction method based on fusion trajectory multi-feature regression, characterized in that, The method comprises the following steps: acquiring a vehicle trajectory stream cluster in a target road section; extracting a trajectory multi-feature vector based on the trajectory stream cluster, specifically including: dividing the trajectory stream cluster into multiple road section units along the driving direction of the road; classifying the vehicle type based on the three-dimensional bounding box size of the vehicle target; in each road section unit, calculating the vehicle type weighted trajectory lateral distribution variance based on the vehicle type classification result; in each road section unit, calculating the average vehicle speed, road curvature and trajectory point density; constructing the trajectory multi-feature vector, which includes the vehicle type weighted trajectory lateral distribution variance, average vehicle speed, road curvature and trajectory point density; According to the vehicle type classification result, the proportion of large transport vehicles is counted and the proportion of small / medium-sized vehicles ; Grouping trajectory points by vehicle class {small / medium vehicles, large transport vehicles} and computing the vehicle class weighted trajectory lateral distribution variance: wherein: representing a road segment unit belonging to a vehicle type category number of trajectory points; representing a vehicle model class corresponding lateral weight; representing a vehicle class the first trajectory point to the centerline lateral distance; representing the vehicle model class the lateral distance mean; a trajectory lateral distribution variance weighted for the vehicle model; The transverse weight is determined according to the proportion of large-size transport vehicles Adaptive setting is performed; based on the trajectory multi-feature vector, determining the road traffic state, specifically including: based on the trajectory multi-feature vector, constructing a trajectory reach consistency scoring function, which couples the lateral dispersion, trajectory point density, bend cutting risk factor, large vehicle influence and trajectory overall bias index; based on the value range of the scoring function, determining the road traffic state as one of free traffic state, constrained traffic state or abnormal bias or trajectory sparse state; According to the road traffic state, a corresponding road boundary candidate is constructed, wherein in the free traffic state, a trajectory regression boundary candidate based on nonlinear regression is constructed, in the constrained traffic state, an alternative boundary candidate based on a traffic space lower bound model is constructed, in the abnormal bias or trajectory sparse state, trajectory-driven modeling is disabled and a prior boundary candidate based on weak geometric features or historical stable results is constructed; for a road segment unit , three kinds of road boundary candidate results have been defined and constructed respectively, and the above boundary candidate results exist in parallel at the method level, constituting a boundary candidate set of a road segment unit; based on the road traffic state, selecting, fusing and rolling back the trajectory regression boundary candidate, the alternative boundary candidate and the prior boundary candidate, and outputting a continuous road boundary line, specifically including: based on the road traffic state, uniquely determining the dominant boundary among the trajectory regression boundary candidate, the alternative boundary candidate and the prior boundary candidate; when the dominant boundary is the trajectory regression boundary candidate or the alternative boundary candidate, based on the trajectory driven confidence weight, using the prior boundary candidate to modify the stability or continuity of the dominant boundary; when the dominant boundary is the prior boundary candidate, explicitly disabling the trajectory regression boundary candidate and the alternative boundary candidate.

2. The method of claim 1, wherein, Acquiring a vehicle trajectory stream cluster in a target road section, specifically including: synchronously collecting laser radar point cloud, inertial measurement unit data, wheel speed odometer data and positioning mapping system pose; based on multi-source sensor fusion, constructing a local consistency coordinate system, and transforming the laser radar point cloud to the local consistency coordinate system; performing outlier filtering and ground point rejection on the transformed point cloud to obtain non-ground point cloud; based on the non-ground point cloud, performing vehicle target detection and cross-frame tracking to obtain time-series vehicle trajectories; in a sliding time window, accumulating the time-series vehicle trajectories of the host vehicle and surrounding vehicles to form the trajectory stream cluster.

3. The method of claim 1, wherein, According to the road traffic state, constructing the corresponding road boundary candidate, specifically including: when the road traffic state is free traffic state, inputting the trajectory multi-feature vector into a pre-trained nonlinear regression model to predict the road effective traffic half-width, and generating a trajectory regression boundary candidate based on the prediction result; when the road traffic state is constrained traffic state, constructing a traffic space lower bound half-width based on the trajectory multi-feature vector, and generating an alternative boundary candidate based on the traffic space lower bound half-width; When the road traffic state is an abnormal bias or trajectory sparsity state, weak geometric features are extracted from the laser radar point cloud or historical stable boundaries are called to generate prior boundary candidates.

4. The method of claim 1, wherein, Based on the road traffic state, the trajectory regression boundary candidate, the alternative boundary candidate and the prior boundary candidate are selected, fused and rolled back, and a continuous road boundary line is output, further comprising: The dominant boundary of each road segment unit after correction is longitudinally spliced and smoothed along the road driving direction, and a continuous road boundary line is output. 5.A road boundary line reconstruction system fusing trajectory multi-feature regression, characterized by, Comprise: A trajectory flow cluster acquisition module is configured to acquire vehicle trajectory flow clusters in a target road segment. A multi-feature vector extraction module is configured to extract trajectory multi-feature vectors based on the trajectory flow clusters, specifically comprising: dividing the trajectory flow clusters into multiple road segment units along the road driving direction. Classify vehicle types based on the three-dimensional bounding box size of the vehicle target; In each road segment unit, calculate the vehicle type weighted trajectory lateral distribution variance based on the vehicle type classification result; In each road segment unit, calculate the average vehicle speed, road curvature and trajectory point density; Construct the trajectory multi-feature vector, which includes the vehicle type weighted trajectory lateral distribution variance, average vehicle speed, road curvature and trajectory point density; According to the vehicle type classification result, the proportion of large transport vehicles is counted and the proportion of small / medium-sized vehicles ; Grouping trajectory points by vehicle class {small / medium vehicles, large transport vehicles} and computing the vehicle class weighted trajectory lateral distribution variance: Wherein: representing road segment units belonging to a vehicle type category number of trajectory points; representing a vehicle model class corresponding lateral weight; representing a vehicle class the first trajectory point to the centerline lateral distance; representing the vehicle model class the lateral distance mean; a trajectory lateral distribution variance weighted for the vehicle model; The transverse weight is determined according to the proportion of large-size transport vehicles Adaptive setting is performed; A traffic state determination module is configured to determine the road traffic state based on the trajectory multi-feature vector, specifically comprising: constructing a trajectory reach consistency scoring function based on the trajectory multi-feature vector, which couples the lateral dispersion, trajectory point density, curve cutting risk factor, large vehicle influence and trajectory overall bias indicator; Based on the value range of the scoring function, the road traffic state is determined as one of free traffic state, constrained traffic state or abnormal bias or trajectory sparsity state. a boundary candidate constructing module, configured to construct a corresponding road boundary candidate according to the road traffic state, wherein a trajectory regression boundary candidate based on nonlinear regression is constructed in a free traffic state, a substitute boundary candidate based on a traffic space lower bound model is constructed in a constrained traffic state, trajectory-driven modeling is disabled and a prior boundary candidate based on weak geometric features or historical stable results is constructed in an abnormal bias or trajectory sparse state; for a road segment unit , three road boundary candidate results have been defined and constructed respectively, and the above boundary candidate results exist in parallel at the method level, constituting a boundary candidate set of a road segment unit; A boundary fusion output module is configured to select, fuse and roll back the trajectory regression boundary candidate, the alternative boundary candidate and the prior boundary candidate based on the road traffic state, and output a continuous road boundary line, specifically comprising: Based on the road traffic state, uniquely determine the dominant boundary among the trajectory regression boundary candidate, the alternative boundary candidate and the prior boundary candidate; When the dominant boundary is the trajectory regression boundary candidate or the alternative boundary candidate, use the prior boundary candidate to modify the stability or continuity of the dominant boundary based on the trajectory driven confidence weight; When the dominant boundary is the prior boundary candidate, explicitly disable the trajectory regression boundary candidate and the alternative boundary candidate.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the program to implement the steps of the road boundary line reconstruction method of the fusion trajectory multi-feature regression as claimed in any one of claims 1 to 4.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the steps of the road boundary line reconstruction method of the fusion trajectory multi-feature regression as claimed in any one of claims 1 to 4.

8. A computer program product comprising software code, characterized in that, The program in the software code performs the steps of the road boundary line reconstruction method of the fusion trajectory multi-feature regression as claimed in any one of claims 1 to 4.

Images