Laser radar target heading angle fitting method and device

Abstract

The present invention belongs to the field of autonomous driving technology and provides a laser radar target heading angle fitting method and device. The fitting method includes: obtaining a point cloud cluster of a target vehicle and traversing all point cloud clusters; using an outlier filtering algorithm based on grid neighborhood density statistics to perform outlier filtering on the point cloud cluster; using a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm to perform heading angle fitting on the point cloud cluster after outlier filtering to obtain the heading angle of the target vehicle. The present application can not only improve fitting efficiency but also improve the accuracy of heading angle fitting results by performing outlier filtering on the point cloud cluster of the target vehicle and by cooperating with the search-based L-shape fitting algorithm and the principal component analysis-based fitting algorithm.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving technology, and in particular relates to a laser radar target heading angle fitting method and device. Background Art

[0002] In a vehicle's assisted driving sensor system, multi-beam lidar can scan and generate point clouds with extremely high angular resolution. Single-frame perception processing using traditional machine learning algorithms typically involves four major algorithmic modules: ground filtering, ROI selection, multi-viewpoint clustering, and target heading angle fitting. The target heading angle fitting module is the final step in outputting single-frame perception observations. If the fitting result deviates from the target's true orientation, it will affect the accuracy of calculations such as the target's lateral and longitudinal position, length and width, and motion trends. Therefore, target heading angle fitting is crucial.

[0003] A Chinese patent (patent number CN201510574374.9) discloses a method for calculating the heading angle of a target vehicle based on 3D lidar. This method classifies the target vehicle's distributed point cloud into two categories based on clustering results, and then performs linear fitting on these two point clouds to obtain the heading angle. This method does not account for the presence of noise in the point cloud cluster, incomplete target point clouds, or point cloud distribution that does not conform to an "L" shape, resulting in reduced accuracy of the heading angle fitting result. Summary of the Invention

[0004] The purpose of the present invention is to provide a laser radar target heading angle fitting method and device to solve the problem pointed out in the background technology that the heading angle calculation method in the prior art does not take into account the existence of noise points in the point cloud cluster, resulting in reduced accuracy of the heading angle fitting result.

[0005] In order to achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0006] In a first aspect, a laser radar target heading angle fitting method is provided, the fitting method comprising:

[0007] Get the point cloud cluster of the target vehicle and traverse all point cloud clusters;

[0008] An outlier filtering algorithm based on grid neighborhood density statistics is used to perform outlier filtering on the point cloud cluster;

[0009] A search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm are used to perform heading angle fitting processing on the point cloud cluster after outlier filtering processing to obtain the heading angle of the target vehicle.

[0010] In combination with the first aspect, the outlier filtering algorithm based on grid neighborhood density statistics is used to perform outlier filtering on the point cloud cluster, including:

[0011] Projecting the point cloud cluster in three-dimensional form onto a plane in a top-down view;

[0012] Drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein;

[0013] Traversing each grid in the grid, and calculating the comprehensive point cloud density of the corresponding central grid with a grid as the center;

[0014] Sort all the results of the comprehensive point cloud density from large to small;

[0015] The point clouds whose comprehensive point cloud density results are within the last 10% of the grid are filtered out.

[0016] In combination with the first aspect, the comprehensive point cloud density is obtained by weighting the point cloud density of the central grid and the point cloud density of all grids adjacent to the central grid. The calculation formula of the comprehensive point cloud density is as follows:

[0017] D=C*0.4+N*0.4+P*0.2,

[0018] Among them, D represents the comprehensive point cloud density, C represents the number of point clouds in the central grid, N represents the number of grids adjacent to the central grid, and P represents the number of all point clouds in the neighborhood of the central grid.

[0019] In combination with the first aspect, the search-based L-shape fitting algorithm and the principal component analysis-based fitting algorithm are used to perform heading angle fitting processing on the point cloud cluster after outlier filtering to obtain the heading angle of the target vehicle, including:

[0020] Using a search-based L-shape fitting algorithm, the heading angle of the point cloud cluster after outlier filtering is fitted, and the first heading angle of the target vehicle is obtained by minimum loss direction fitting. The size and position of the target vehicle are then calculated to obtain the target frame of the target vehicle.

[0021] Determine whether the target frame is within the cut-in area of ​​the vehicle. If not, directly use the first heading angle as the heading angle of the target vehicle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain the second heading angle of the target vehicle and related information, and then proceed to the next step.

[0022] Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

[0023] In combination with the first aspect, in the step of performing heading angle fitting processing on the point cloud cluster after outlier filtering using a search-based L-shape fitting algorithm, the search angle range is -45° to 45°, and the search angle step is 2.5°.

[0024] In combination with the first aspect, the fitting method further includes:

[0025] The associated information of the heading angle of the target vehicle is verified based on the experience level. If the verification fails, the heading angle of the target vehicle is assumed to be 0°.

[0026] In a second aspect, a laser radar target heading angle fitting device is provided, the fitting device comprising:

[0027] The acquisition module is used to obtain the point cloud cluster of the target vehicle and traverse all point cloud clusters;

[0028] A first processing module is configured to perform outlier filtering on the point cloud cluster using an outlier filtering algorithm based on grid neighborhood density statistics;

[0029] The second processing module is used to use a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm to perform heading angle fitting processing on the point cloud cluster after outlier filtering processing to obtain the heading angle of the target vehicle.

[0030] In conjunction with the second aspect, the first processing module is further configured to:

[0031] Projecting the point cloud cluster in three-dimensional form onto a plane in a top-down view;

[0032] Drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein;

[0033] Traversing each grid in the grid, and calculating the comprehensive point cloud density of the corresponding central grid with a grid as the center;

[0034] Sort all the results of the comprehensive point cloud density from large to small;

[0035] The point clouds within the last 10% of the grid of the comprehensive point cloud density result are filtered out.

[0036] In conjunction with the second aspect, the second processing module is further configured to:

[0037] Using a search-based L-shape fitting algorithm, the heading angle of the point cloud cluster after outlier filtering is fitted, and the first heading angle of the target vehicle is obtained by minimum loss direction fitting. The size and position of the target vehicle are then calculated to obtain the target frame of the target vehicle.

[0038] Determine whether the target frame is within the vehicle's cut-in area. If not, directly use the first heading angle as the target vehicle's heading angle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain a second heading angle of the target vehicle, and proceed to the next step.

[0039] Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

[0040] In combination with the second aspect, the fitting device further includes:

[0041] The verification module is used to verify the associated information of the heading angle of the target vehicle based on the experience level. If the verification fails, the heading angle of the target vehicle is defaulted to 0°.

[0042] The invention adopting the above technical solution has the following advantages:

[0043] 1. Through the fitting method of this application, the point cloud cluster of the target vehicle is first obtained, and then the outlier filtering algorithm is used to filter out the outliers in the point cloud cluster. Finally, the search-based L-shape fitting algorithm and the principal component analysis-based fitting algorithm are used to cooperate with each other to fit the heading angle of the target vehicle. Through this design, not only the fitting efficiency can be improved, but also the influence of noise on the heading angle fitting result can be reduced, thereby improving the accuracy of the heading angle fitting result.

[0044] 2. By combining the search-based L-shape fitting algorithm with the principal component analysis-based fitting algorithm, this method can fit the heading angle suitable for different scenarios, regardless of whether the point cloud of the target vehicle has an "L"-shaped distribution or an "I"-shaped distribution, thereby improving the accuracy of the heading angle fitting results. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention can be further illustrated by the non-limiting examples given in the accompanying drawings;

[0046] Figure 1 Schematic diagram of the flow of the laser radar target heading angle fitting method in an embodiment of the present invention;

[0047] Figure 2 This is the point cloud before outlier filtering in an embodiment of the present invention;

[0048] Figure 3 This is the point cloud after outlier filtering in an embodiment of the present invention;

[0049] Figure 4 This is the fitting result of the search-based L-shape fitting algorithm used in the embodiment of the present invention;

[0050] Figure 5 This is the fitting result of the fitting algorithm based on principal component analysis in the embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that in the drawings or descriptions, similar or identical parts are numbered the same. Implementations not shown or described in the drawings are forms known to those of ordinary skill in the art. In addition, directional terms mentioned in the embodiments, such as "upper," "lower," "top," "bottom," "left," "right," "front," and "back," are merely references to the directions in the drawings and are not intended to limit the scope of protection of the present invention.

[0052] An embodiment of the present application provides a laser radar target heading angle fitting method, which is mainly used for vehicles equipped with laser radar. The laser radar can collect point cloud data of the target vehicle, and the fitting method is mainly to fit the heading angle of the target vehicle more accurately, thereby assisting the vehicle in obtaining related information of the target vehicle during the automatic driving process. The related information mainly includes information such as the size, position and target frame of the target vehicle.

[0053] like Figure 1 Specifically, the fitting method may include the following steps:

[0054] S100, obtaining the point cloud cluster of the target vehicle and traversing all point cloud clusters;

[0055] S200, using an outlier filtering algorithm based on grid neighborhood density statistics to perform outlier filtering on the point cloud cluster;

[0056] S300 , using a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm to perform heading angle fitting processing on the point cloud cluster after outlier filtering processing, so as to obtain the heading angle of the target vehicle.

[0057] Through the above implementation method, the point cloud cluster of the target vehicle is first obtained, and then the outlier filtering algorithm is used to filter out the outliers in the point cloud cluster. Finally, the search-based L-shape fitting algorithm and the principal component analysis-based fitting algorithm are used to fit the heading angle of the target vehicle. The cooperation of these three algorithms can not only improve the fitting efficiency, but also reduce the influence of noise on the heading angle fitting results, thereby improving the accuracy of the heading angle fitting results.

[0058] The following is a detailed description of each implementation step:

[0059] In step S100, the method for obtaining a point cloud cluster for the target vehicle is as follows: first, all point cloud data of the target vehicle is acquired using a LiDAR radar. Then, operations such as ground filtering, ROI selection, and clustering are performed on the point cloud data to obtain a point cloud cluster for the target vehicle. There are typically multiple point cloud clusters, each of which can be fitted to the heading angle of the target vehicle.

[0060] In step S200, considering that most point cloud clusters contain noise during the actual fitting process, an outlier filtering algorithm is used to remove outliers from the point cloud clusters before fitting the heading angle. This reduces the impact of noise and incomplete target point clouds, extracts better contour features for subsequent heading angle fitting, and thus improves the accuracy of heading angle fitting. This algorithm also operates very efficiently, which can improve the efficiency of heading angle fitting.

[0061] In step S300, the existing technique for fitting heading angles abstracts the target vehicle as a rectangle. For a target vehicle in an adjacent lane and in front of the host vehicle, the point cloud acquired by the LiDAR can exhibit an approximately L-shaped distribution. In this case, a search-based L-shape fitting algorithm is more appropriate for fitting the heading angle. Furthermore, this algorithm adds a regularization term to the fitting loss. This regularization term uses the fitting heading angle as a parameter. The larger the heading angle, the greater the loss added by the regularization term. Therefore, this regularization term prevents individual noise points from affecting the accuracy of the heading angle.

[0062] For targets in adjacent lanes and traveling parallel to the vehicle, the LiDAR can only observe one side of a rectangle. Therefore, the target vehicle's point cloud exhibits an I-shaped distribution. In this case, a fitting algorithm based on principal component analysis is more appropriate for fitting the heading angle. This method combines prior knowledge of viewing angles and applies these two algorithms across different perception regions. This not only accurately fits the heading angle, but also facilitates the vehicle's autonomous vehicle cut-in operation, reducing the risk of traffic accidents.

[0063] As an optional implementation, the outlier filtering algorithm based on grid neighborhood density statistics is used to perform outlier filtering on the point cloud cluster, including the following steps:

[0064] S210, projecting the three-dimensional point cloud cluster onto a plane in a top view;

[0065] S220, drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein;

[0066] S230, traversing each grid in the grid, and calculating the comprehensive point cloud density corresponding to the center grid with one grid as the center;

[0067] S240, sorting all the comprehensive point cloud density results from largest to smallest;

[0068] S250 , filtering out point clouds whose comprehensive point cloud density results are within the last 10% of the grids.

[0069] Through the above implementation, outliers in the point cloud cluster can be quickly and effectively filtered out, thereby improving the accuracy of the subsequent fitting heading angle.

[0070] As an optional implementation, the comprehensive point cloud density is obtained by weighting the point cloud density of the central grid and the point cloud density of all grids adjacent to the central grid. The calculation formula of the comprehensive point cloud density is as follows:

[0071] D=C*0.4+N*0.4+P*0.2,

[0072] Among them, D represents the comprehensive point cloud density, C represents the number of point clouds in the central grid, N represents the number of grids adjacent to the central grid, and P represents the number of all point clouds in the neighborhood of the central grid.

[0073] As an optional implementation, the search-based L-shape fitting algorithm and the principal component analysis-based fitting algorithm are used to perform heading angle fitting on the point cloud cluster after outlier filtering to obtain the heading angle of the target vehicle, including:

[0074] S310: Using a search-based L-shape fitting algorithm, perform heading angle fitting on the point cloud cluster after outlier filtering, obtain a first heading angle of the target vehicle through minimum loss direction fitting, and calculate the size and position of the target vehicle to obtain a target frame of the target vehicle;

[0075] S320: Determine whether the target frame is in the cut-in area of ​​the vehicle. If not, directly use the first heading angle as the heading angle of the target vehicle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain a second heading angle of the target vehicle and related information, and then proceed to the next step.

[0076] S330. Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

[0077] Through the above implementation, regardless of whether the point cloud of the target vehicle is distributed in an "L" shape or an "I" shape, this method can fit the heading angle suitable for the scene according to different scenarios, thereby improving the accuracy of the heading angle fitting result.

[0078] As an optional implementation, in the step of performing heading angle fitting processing on the point cloud cluster after outlier filtering using a search-based L-shape fitting algorithm, the search angle range is -45° to 45°, and the search angle step is 2.5°.

[0079] As an optional implementation, the fitting method further includes:

[0080] S400: Verify the associated information of the heading angle of the target vehicle based on experience. If the verification fails, the heading angle of the target vehicle is assumed to be 0°.

[0081] It should be noted that during actual driving, if the target vehicle is far away from the host vehicle and / or the target vehicle is too small, the fitted heading angle is unstable and cannot be used as a reference during autonomous driving. Therefore, based on actual driving experience, specific position distance thresholds and size thresholds can be set. When the target vehicle's position is greater than the position distance threshold and / or the target vehicle's size is less than the size threshold in the associated information obtained based on the target vehicle's heading angle, the target vehicle's heading angle can be defaulted to 0°. At this time, the host vehicle is informed to track the target vehicle through multi-frame tracking or other methods.

[0082] The present application also provides a laser radar target heading angle fitting device, the fitting device comprising:

[0083] The acquisition module is used to obtain the point cloud cluster of the target vehicle and traverse all point cloud clusters;

[0084] A first processing module is configured to perform outlier filtering on the point cloud cluster using an outlier filtering algorithm based on grid neighborhood density statistics;

[0085] The second processing module is used to use a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm to perform heading angle fitting processing on the point cloud cluster after outlier filtering processing to obtain the heading angle of the target vehicle.

[0086] Optionally, the first processing module is further configured to:

[0087] Projecting the point cloud cluster in three-dimensional form onto a plane in a top-down view;

[0088] Drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein;

[0089] Traversing each grid in the grid, and calculating the comprehensive point cloud density of the corresponding central grid with a grid as the center;

[0090] Sort all the results of the comprehensive point cloud density from large to small;

[0091] The point clouds within the last 10% of the grid of the comprehensive point cloud density result are filtered out.

[0092] Optionally, the second processing module is further configured to:

[0093] Using a search-based L-shape fitting algorithm, the heading angle of the point cloud cluster after outlier filtering is fitted, and the first heading angle of the target vehicle is obtained by minimum loss direction fitting. The size and position of the target vehicle are then calculated to obtain the target frame of the target vehicle.

[0094] Determine whether the target frame is within the vehicle's cut-in area. If not, directly use the first heading angle as the target vehicle's heading angle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain a second heading angle of the target vehicle, and proceed to the next step.

[0095] Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

[0096] Optionally, the fitting device further comprises:

[0097] The verification module is used to verify the associated information of the heading angle of the target vehicle based on the experience level. If the verification fails, the heading angle of the target vehicle is defaulted to 0°.

[0098] The above describes in detail the lidar target heading angle fitting method and apparatus provided by the present invention. The description of the specific embodiments is intended only to facilitate understanding of the method and core concepts of the present invention. It should be noted that those skilled in the art will be able to make various improvements and modifications to the present invention without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the claims of the present invention.

Claims

1. A laser radar target heading angle fitting method, characterized in that: The fitting method includes: Get the point cloud cluster of the target vehicle and traverse all point cloud clusters; An outlier filtering algorithm based on grid neighborhood density statistics is used to perform outlier filtering on the point cloud cluster; Using a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm, the heading angle fitting process is performed on the point cloud cluster after outlier filtering to obtain the heading angle of the target vehicle; The method adopts a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm to perform heading angle fitting processing on the point cloud cluster after outlier filtering processing to obtain the heading angle of the target vehicle, including: Using a search-based L-shape fitting algorithm, the heading angle of the point cloud cluster after outlier filtering is fitted, and the first heading angle of the target vehicle is obtained by minimum loss direction fitting. The size and position of the target vehicle are then calculated to obtain the target frame of the target vehicle. Determine whether the target frame is within the cut-in area of ​​the vehicle. If not, directly use the first heading angle as the heading angle of the target vehicle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain the second heading angle of the target vehicle and related information, and then proceed to the next step. Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

2. The fitting method according to claim 1, wherein The outlier filtering algorithm based on grid neighborhood density statistics is used to perform outlier filtering on the point cloud cluster, including: Projecting the point cloud cluster in three-dimensional form onto a plane in a top-down view; Drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein; Traversing each grid in the grid, and calculating the comprehensive point cloud density of the corresponding central grid with a grid as the center; Sort all the results of the comprehensive point cloud density from large to small; The point clouds whose comprehensive point cloud density results are within the last 10% of the grids are filtered out.

3. The fitting method according to claim 2, wherein: The comprehensive point cloud density is obtained by weighting the point cloud density of the central grid and the point cloud density of all grids adjacent to the central grid. The calculation formula of the comprehensive point cloud density is as follows: , Among them, D represents the comprehensive point cloud density, C represents the number of point clouds in the central grid, N represents the number of grids adjacent to the central grid, and P represents the number of all point clouds in the neighborhood of the central grid.

4. The fitting method according to claim 1, wherein: In the step of performing heading angle fitting processing on the point cloud cluster after outlier filtering processing using a search-based L-shape fitting algorithm, the search angle range is -45° to 45°, and the search angle step size is 2.5°.

5. The fitting method according to any one of claims 1 to 4, characterized in that: The fitting method further comprises: The associated information of the heading angle of the target vehicle is verified based on the experience level. If the verification fails, the heading angle of the target vehicle is assumed to be 0°.

6. A laser radar target heading angle fitting device, characterized in that: The fitting device comprises: The acquisition module is used to obtain the point cloud cluster of the target vehicle and traverse all point cloud clusters; A first processing module is configured to perform outlier filtering on the point cloud cluster using an outlier filtering algorithm based on grid neighborhood density statistics; a second processing module, configured to perform heading angle fitting processing on the point cloud cluster after outlier filtering using a search-based L-shape fitting algorithm and a principal component analysis-based fitting algorithm, so as to obtain the heading angle of the target vehicle; Using a search-based L-shape fitting algorithm, the heading angle of the point cloud cluster after outlier filtering is fitted, and the first heading angle of the target vehicle is obtained by minimum loss direction fitting. The size and position of the target vehicle are then calculated to obtain the target frame of the target vehicle. Determine whether the target frame is within the vehicle's cut-in area. If not, directly use the first heading angle as the target vehicle's heading angle. If so, use a fitting algorithm based on principal component analysis to perform heading angle fitting on the point cloud cluster again to obtain a second heading angle of the target vehicle, and proceed to the next step. Determine whether the difference between the first heading angle and the second heading angle is within a threshold range. If so, weight the first heading angle and the second heading angle to obtain the heading angle of the target vehicle. If not, directly use the second heading angle as the heading angle of the target vehicle.

7. The fitting device according to claim 6, characterized in that The first processing module is further configured to: Projecting the point cloud cluster in three-dimensional form onto a plane in a top-down view; Drawing a grid of adaptive size on the plane, and performing two-dimensional rasterization processing on the point cloud cluster, wherein the grid is a 10*10 grid, and each grid in the grid is represented by a point cloud projected therein; Traversing each grid in the grid, and calculating the comprehensive point cloud density of the corresponding central grid with a grid as the center; Sort all the results of the comprehensive point cloud density from large to small; The point clouds whose result of the comprehensive point cloud density is within the last 10% of the grid are filtered out.

8. The fitting device according to claim 6, characterized in that The fitting device further comprises: The verification module is used to verify the associated information of the heading angle of the target vehicle based on the experience level. If the verification fails, the heading angle of the target vehicle is defaulted to 0°.

Images