A method for detecting three-dimensional parking sight distance under laser radar data

Abstract

The application discloses a kind of detection methods of three-dimensional parking sight distance under laser radar data.Classical sight distance test lacks automated detection method, and sight distance value is more from planning design value.The method is based on the laser radar point cloud data collected in actual road scene, and the city road scene point cloud is automatically classified and attributed to attribute label using PointNet++.Secondly, the target object point cloud is framed using the ellipsoid view model of innovative extension transverse range, and the efficiency of the overall model algorithm is improved using the eigenvalue calculation and comparison algorithm for the purpose of sparse point cloud.Subsequently, the speed and road surface condition change are included, and the parking sight distance is sequentially tested according to the view point change whether it meets the specification requirements, and the three-dimensional parking sight distance test value and the point cloud of occluded line of sight are output, to realize the three-dimensional parking sight distance detection of the whole road section.The application effectively solves the problem that the current situation of parking sight distance in actual road scene is unclear, and realizes the quantitative and fine test of the current situation of city road sight distance.

Description

Technical Field

[0001] This invention relates to a method for detecting three-dimensional parking line-of-sight based on lidar data, belonging to the field of transportation engineering. Background Technology

[0002] Stopping sight distance is the shortest driving distance required to ensure a driver can make correct perception and decision-making within a short time after encountering an obstacle ahead. In road design, stopping sight distance is often checked in two dimensions based on horizontal and vertical alignment. However, with the continuous growth of roadside vegetation into the inner part of the road and the increase of traffic signs and other facilities, the three-dimensional characteristics of the road as a spatial entity are becoming increasingly important. In real road scenarios, changes in driving speed and road surface conditions (ice, snow, etc.) can lead to insufficient effective sight distance. Therefore, conducting three-dimensional stopping sight distance checks and safety assessments on common urban roads has significant theoretical and applied value for road safety driving early warning and intelligent maintenance.

[0003] Point cloud data acquired using LiDAR (Light Detection and Ranging) vehicles and terrestrial 3D laser scanners provides centimeter-level density and accuracy for reconstructing road scenes, which can be used to verify actual road sight distances. Early deep learning networks required transforming disordered point clouds into regular representations through 3D voxel or 2D raster operations before feeding them into the deep learning network. PointNet enabled end-to-end point cloud data classification and segmentation. However, it only considered global features of the point cloud while ignoring local features. The subsequently proposed PointNet++ network, which combines local and global features, effectively improved its classification performance and was well validated in road information classification and extraction. However, deep learning networks are rarely used for obstacle identification in 3D sight distance research. Furthermore, in establishing 3D sight distance verification models based on point cloud scenes, a reasonable field of view needs to be selected to comprehensively capture information about the lane ahead and roadside features (trees, signs, and buildings, etc.). One approach uses a cone shape, with the human eye as the vertex, to construct a circular field of view, but this may overlook occlusion caused by inward growth of roadside vegetation. To improve the efficiency of line-of-sight measurement and recognition, uniform thinning to reduce point cloud density can easily lead to the loss of important feature points. This invention utilizes an internal shape descriptor algorithm to rapidly compress data and retain key feature points within the scene by analyzing neighborhood relationships between point clouds, the point cloud covariance matrix, and the relationships between eigenvalues ​​of the matrix. JTG B05—2015, the "Specification for Safety Evaluation of Highway Projects," proposes evaluating stopping line-of-sight under real-world driving conditions. Therefore, this invention proposes an automatic stopping line-of-sight detection method in a lidar point cloud environment.

[0004] This invention provides a detailed description of a method for detecting parking line-of-sight distance in a lidar point cloud environment. PointNet++ is used to automatically classify and assign attribute labels to road scene point clouds, solving the problem of unknown obstacle categories due to the lack of direct ground feature attribute information in the point cloud. An ellipsoidal field-of-view model is created to automatically select the field of view based on road parameters, and feature points of major ground features are filtered through point cloud feature values ​​to improve data processing efficiency. Considering changes in speed and road conditions, parking line-of-sight distance is checked sequentially according to the viewpoint trajectory, outputting three-dimensional line-of-sight distance detection results to determine whether the current parking line-of-sight distance meets the regulatory requirements, providing quantitative detection results for road safety evaluation. Summary of the Invention

[0005] The technical solution adopted in this invention is to propose a method for detecting three-dimensional parking line-of-sight based on lidar data, comprising the following steps:

[0006] Step 1: Use PointNet++ to automatically classify the point cloud of the road scene. The PointNet++ network structure mainly includes a multi-level feature learning module and a classification and segmentation module. The multi-level feature learning module consists of multiple Set Abstraction modules. The point cloud is input in the form of N×(d+C) (where N is the number of input points, d is the coordinate dimension of the points, and C is the feature dimension). The sampling layer of each module mainly constructs a centroid set through an iterative farthest point sampling algorithm. The combination layer classifies the point cloud into multiple local point sets based on the centroid set. A spherical sampling method is then used with the centroid point as the center to output the local point cloud. PointNet layers are then added to learn the global features of the point cloud in each local region. Finally, the point sets are mapped back to the initial structure, and ground feature category labels are output. This result is used for subsequent direct determination and visualization of obstacle categories.

[0007] Step 2: Construct an ellipsoidal field of view model extending along the horizontal axis of the real-world road to automatically simulate the driver's viewpoint range in the point cloud scene.

[0008] An extended horizontal axis field of view is constructed using the mathematical and geometric model of an ellipsoid. A novel view frustum is then constructed by connecting the point cloud within the field of view to the viewpoint. This aims to improve the original view frustum field of view model, ensuring that the selected field of view encompasses both the area in front of the road and the roadside environment. The specific method for creating the point cloud model of the target object that the corresponding observation point should observe is as follows:

[0009] Referring to the viewpoint positions specified in the "Highway Route Design Specifications," the center point of the ellipsoidal viewpoint model is fixed along the lane centerline. The coordinates (x, y) of any point contained within the model are... o ,y o ,z o The mathematical model formula for selecting the field of view of an ellipsoid that satisfies the equation is as follows:

[0010]

[0011] In the formula, a and b are the equatorial radii of the ellipsoid along the x and y axes, respectively, in meters; c is the polar radius of the ellipsoid along the z axis, in meters. The resulting viewpoint point cloud dataset consists of ground point clouds and non-ground point clouds.

[0012] Step 3: In order to improve the calculation efficiency and not lose the main feature points, the point cloud in the field of view is thinned, and only the important feature points are retained.

[0013] The specific process is as follows: First, for each point m in the point cloud... t Establish a local coordinate system; use a KD tree to determine the coordinates of each point m. t Centered on a point m, search for all neighboring points m with a neighborhood radius r. u Next, considering the characteristics of urban roads, the weight ω is calculated using the following formula. tu :

[0014]

[0015] The weight result ω is obtained tu Then calculate m t covariance matrix cov(m) t ):

[0016] Secondly, for each point m t The covariance matrix is ​​further calculated to obtain its corresponding eigenvalues. The eigenvalues ​​are then arranged in descending order, and the first three eigenvalues ​​are selected. pass Quantitative relationships determine feature points; all points that satisfy the formula form a cluster of key feature points within the field of view.

[0017] Based on multiple experimental tests, the optimal thresholds ε1 = 0.8 and ε2 = 0.4 were selected (usually not exceeding 1).

[0018] Step 4: Along the driving trajectory, change the relevant variables of parking sight distance, and calculate the three-dimensional parking sight distance test value of the road segment corresponding to each viewpoint in turn, and determine whether the actual parking sight distance value meets the corresponding specification requirements.

[0019] Refer to the parking sight distance S in the "Specification" v Calculation formula:

[0020]

[0021] In the formula: V is the operating speed, km / h; T is the driver's reaction time, s; K is the braking coefficient (generally taken as 1.3); f is the friction coefficient between the road surface and the tires (abbreviated as friction coefficient). This varies depending on road conditions, and four road conditions are selected: dry, wet, muddy, and icy; i is the longitudinal slope of the road section (%); S c For safe distance, m; Sv The standard value for parking sight distance is in meters (m).

[0022] Along the driving trajectory, the three-dimensional parking sight distance verification value for each viewpoint corresponding to the road segment is calculated sequentially:

[0023] (1) Input the key feature point cloud and viewpoint coordinates within the field of view; (2) Connect the viewpoint to each feature point in the field of view and calculate the Euclidean distance; (3) Sequentially determine whether there is an environmental point cloud falling on the line connecting the viewpoint and the feature point: a) If yes, the line of sight is obstructed, output the occluded point cloud with category label, and record the minimum line of sight distance as the parking sight distance test value; b) If no, the line of sight is unobstructed, and record the maximum line of sight distance as the parking sight distance test value; (4) Loop to the next viewpoint and repeat the above steps until the entire road section test is completed. Change the driving speed and friction coefficient (dry, wet, muddy and icy) in the parking sight distance, calculate the corresponding parking sight distance value, and compare the test value under the point cloud scene to determine whether the parking meets the specification requirements.

[0024] The beneficial effects of this invention are:

[0025] This invention proposes a 3D parking line-of-sight detection method based on lidar data. Utilizing the automatic classification results of urban road scenes, an ellipsoidal field-of-view model is constructed to account for occlusion caused by inward growth of roadside vegetation. The feature point extraction method within the field of view significantly improves computational efficiency. In a point cloud environment, automatic detection of 3D parking line-of-sight and automatic perception of line-of-sight obstacles are achieved across the entire road segment. This effectively solves the problems of unclear current road line-of-sight status and ambiguous risk points, accurately depicting the current 3D line-of-sight status and risks of existing roads. The overall flowchart of this invention is shown below. Figure 1 . Attached Figure Description

[0026] Figure 1 Overall flowchart of the invention Detailed Implementation

[0027] This paper proposes a 3D parking sight distance detection method based on LiDAR data. First, PointNet++ is used to automatically classify the road scene point cloud. Then, an ellipsoidal field-of-view model extending along the horizontal axis of the real road is constructed to automatically simulate the driver's viewpoint range within the point cloud scene. Next, to improve calculation efficiency without losing key feature points, the point cloud within the field of view is thinned, retaining only important feature points. Finally, along the driving trajectory, parking sight distance-related variables are changed, and the 3D parking sight distance verification value for each viewpoint corresponding to the road segment is calculated sequentially to determine whether the actual parking sight distance value meets the corresponding specification requirements.

[0028] (1) Use PointNet++ to automatically classify point clouds of road scenes.

[0029] The PointNet++ network structure mainly includes a multi-level feature learning module and a classification and segmentation module. The multi-level feature learning module consists of multiple Set Abstraction modules. The point cloud is input in the form of N×(d+C) (where N is the number of input points, d is the coordinate dimension of the points, and C is the feature dimension). The sampling layer of each module mainly constructs a centroid set through an iterative farthest point sampling algorithm. The combination layer classifies the point cloud into multiple local point sets based on the centroid set. A spherical sampling method is then used with the centroid point as the center to output the local point cloud. PointNet layers are then added to learn the global features of the point cloud in each local region. Finally, the point sets are mapped back to the initial structure, and ground feature category labels are output. This result is used for subsequent direct determination and visualization of obstacle categories.

[0030] The training and test samples for the network were randomly assigned in a 7:3 ratio. During network training, the hardware configuration was a single GPU, an NVIDIA GeForce RTX 2060 SUPER, and a Windows 10 64-bit operating system. The initial learning rate was set to 0.001, the batch size was 16, and training was performed for 64 iterations.

[0031] (2) Construct an ellipsoidal field of view model extending horizontally under the real road scene, and automatically simulate the driver's field of view range in the point cloud scene.

[0032] Referring to the viewpoint positions specified in the "Highway Route Design Specifications," the center point of the ellipsoidal viewpoint model is fixed along the lane centerline. The coordinates (x, y) of any point contained within the model are... o ,y o ,z o The mathematical model formula for selecting the field of view of an ellipsoid that satisfies the equation is as follows:

[0033] In the formula, a and b are the equatorial radii of the ellipsoid along the x and y axes, respectively, in meters (m); c is the polar radius of the ellipsoid along the z-axis, in meters (m). For urban two-way / two-lane roads, a recommended value range is 2m to 4m, b is 3m to 5m, and c is 4m to 7m, which can effectively simulate the driver's forward field of vision. For multi-lane urban roads, the equatorial radius parameter of the model is appropriately increased based on the vehicle's lane position and lane width, with a recommended value range of 4m to 6m for a, 5m to 8m for b, and c remaining unchanged. The resulting field-of-view point cloud dataset consists of ground point clouds and non-ground point clouds. This invention designs recommended parameter ranges for an elliptical field-of-view model for urban roads.

[0034] (3) In order to improve the calculation efficiency and not lose the main feature points, the point cloud in the field of view is thinned out, and only the important feature points are retained.

[0035] Specific process: First, for each point m in the point cloud... t Establish a local coordinate system; use a KD tree to determine the coordinates of each point m. t Centered on a point m, search for all neighboring points m with a neighborhood radius r. u Next, considering the characteristics of urban roads, the weight ω is calculated using the following formula. tu :

[0036] The weight result ω is obtained tu Then calculate m t covariance matrix cov(m) t For each point, the covariance matrix is ​​further calculated to obtain its corresponding eigenvalues. The eigenvalues ​​are then arranged in descending order, and the first three eigenvalues ​​are selected. pass The quantitative relationships are used to determine feature points. Based on multiple experimental tests, the optimal thresholds ε1 = 0.8 and ε2 = 0.4 are selected (usually not exceeding 1). All points satisfying the formula form a cloud of key ground feature points within the field of view. This invention designs a field-of-view point cloud thinning algorithm and parameters suitable for ground 3D laser scanning and mobile scanning vehicle point cloud data.

[0037] (4) Along the driving trajectory, change the relevant variables of parking sight distance, and calculate the three-dimensional parking sight distance test value of each viewpoint corresponding to the road segment in turn, and determine whether the actual parking sight distance value meets the corresponding specification requirements.

[0038] Refer to the parking sight distance S in the "Specification" v Calculation formula:

[0039] In the formula: V is the operating speed, km / h; T is the driver's reaction time, s; K is the braking coefficient (generally taken as 1.3); f is the friction coefficient between the road surface and the tires (abbreviated as friction coefficient). This varies depending on road conditions, and four road conditions are selected: dry, wet, muddy, and icy; i is the longitudinal slope of the road section (%); S c For safe distance, m; S v The standard value for parking sight distance is in meters (m). The sight distance interval can be defined as S. v Alternatively, the sight distance interval can be calculated based on the vehicle speed. During the sight distance inspection, the vehicle speed and friction coefficient in the parking sight distance are changed, and the changes in the parking sight distance value under different parameter conditions are measured to determine whether it meets the specification requirements.

[0040] The method for calculating the stopping sight distance verification value for each viewpoint along the driving trajectory is as follows:

[0041] (1) Input the key feature point cloud and viewpoint coordinates within the field of view; (2) Connect the viewpoint to each feature point in the field of view and calculate the Euclidean distance; (3) Determine in turn whether there is an environmental point cloud falling on the line connecting the viewpoint and the feature point: a) If yes, the line of sight is blocked, output the occluded point cloud with category label, and record the minimum line of sight distance as the parking sight distance test value; b) If no, the line of sight is unobstructed, and record the maximum sight distance value as the parking sight distance test value; (4) Loop to the next viewpoint and repeat the above steps in turn until the test of the entire road section is completed.

[0042] When the driving speed varies from 30km / h to 80km / h, the corresponding standard stopping sight distance values ​​and the stopping sight distance values ​​in the road point cloud scene are calculated respectively. It is then determined whether the actual stopping sight distance values ​​at different driving speeds meet the standards. For road sections that do not meet the standards, specific values ​​and obstacle locations and categories are provided for maintenance management and safety analysis decision-making. Based on the standards, when the road surface conditions are dry, wet, muddy, and icy, the corresponding friction coefficients are calculated, and the corresponding standard stopping sight distance values ​​and the stopping sight distance values ​​in the road point cloud scene are calculated respectively. It is then determined whether the actual stopping sight distance values ​​at different driving speeds meet the standards. For road sections that do not meet the standards, specific values ​​and obstacle locations and categories are provided for maintenance management and safety analysis decision-making.

Claims

1. A method for detecting three-dimensional parking line-of-sight based on lidar data, characterized in that, Includes the following steps: Step 1: Automatically classify the point cloud of the road scene using the PointNet++ network; The PointNet++ network structure includes a multi-level feature learning module and a classification and segmentation module; Multi-level feature learning consists of multiple Set Abstraction modules. The point cloud is input in the form of N×(d+C), where N is the number of input points, d is the coordinate dimension of the points, and C is the feature dimension. Each module's sampling layer primarily constructs a centroid set through an iterative farthest point sampling algorithm. The combination layer classifies the point cloud into multiple local point sets based on the centroid set. After using a spherical sampling method with the centroid point as the center, the local point cloud is output. A PointNet layer is added to learn the global features of the point cloud in each local region. Finally, the point sets are mapped to the initial structure, and the ground feature category labels are output. This result is used for the direct determination and visualization of subsequent obstacle categories. Step 2: Construct an ellipsoidal field of view model extending along the horizontal axis of the real-world road, and automatically simulate the driver's field of view range in the point cloud scene; The center point of the fixed ellipsoidal view model is defined along the lane centerline; the coordinates (x, y) of any point contained within the model are defined. o ,y o ,z o The mathematical model formula for selecting the field of view of an ellipsoid that satisfies the equation is as follows: In the formula, a and b are the equatorial radii of the ellipsoid along the x and y axes, respectively, in meters; c is the polar radius of the ellipsoid along the z axis, in meters; the obtained field-of-view point cloud dataset consists of ground point clouds and non-ground point clouds. Step 3: Thin out the point cloud within the field of view, retaining only the important feature points; The specific process is as follows: First, for each point m in the point cloud... t Establish a local coordinate system; use a KD tree to determine the coordinates of each point m. t Centered on a point m, search for all neighboring points m with a neighborhood radius r. u Next, considering the characteristics of urban roads, the weight ω is calculated using the following formula. tu : The weight result ω is obtained tu Then calculate m t covariance matrix cov(m) t ): Secondly, for each point m t The covariance matrix is ​​further calculated to obtain its corresponding eigenvalues. The eigenvalues ​​are then arranged in descending order, and the first three eigenvalues ​​are selected. pass Quantitative relationships determine feature points; all points that satisfy the formula form a cluster of key ground feature points within the field of view. Step 4: Along the driving trajectory, change the relevant variables of parking sight distance, and calculate the three-dimensional parking sight distance test value of the road segment corresponding to each viewpoint in turn, and determine whether the actual parking sight distance value meets the corresponding specification requirements. Parking sight distance S v Calculation formula: Where: V is the operating speed, km / h; T is the driver's reaction time, s; K is the braking coefficient; f is the friction resistance coefficient between the road surface and the tire; it varies with different road conditions, and four road conditions are selected: dry, wet, muddy, and icy. i represents the longitudinal slope of the road segment, in percentage; S c For safe distance, m; S v The standard value for parking sight distance is in meters (m). Along the driving trajectory, the three-dimensional parking sight distance verification value for each viewpoint corresponding to the road segment is calculated sequentially: (1) Input the key feature point cloud and viewpoint coordinates within the field of view; (2) Connect the viewpoint to each feature point in the field of view and calculate the Euclidean distance; (3) Determine in turn whether there is an environmental point cloud that falls on the line connecting the viewpoint and the feature point: a) If yes, the line of sight is blocked, output the occluded point cloud with category label, and record the minimum line of sight distance as the parking sight distance test value; b) If no, the line of sight is unobstructed, and record the maximum sight distance value as the parking sight distance test value. (4) Loop to the next viewpoint and repeat the above steps until the entire road section test is completed; The driving speed and friction coefficient in the parking sight distance are changed according to dry, wet, muddy and icy conditions, and the corresponding parking sight distance values ​​are calculated.

Images