A radar calibration method for unmanned vehicles based on improved point cloud registration method

Abstract

The present invention provides a radar calibration method for an unmanned vehicle based on an improved point cloud registration method. The main radar and the calibration radar calculate their own ground equations respectively, and then generate virtual ground points to add to their own point clouds respectively, and then use the point clouds after data augmentation for registration. This overcomes the problem of unbalanced calibration results caused by missing ground points, and can reuse the existing point cloud registration algorithm interface without the need to make internal changes to the point cloud registration algorithm.

Description

Technical Field

[0001] The present invention relates to the field of radar positioning technology, and in particular to an unmanned vehicle radar calibration method based on an improved point cloud registration method. Background Art

[0002] Autonomous vehicles, or unmanned vehicles as they are commonly known, are equipped with a variety of sensors to sense the surrounding environment and objects, and then adopt different strategies such as moving forward, waiting, detouring and avoiding. As an active detection and perception system, LiDAR has become an indispensable sensor in the realization of autonomous driving because of its advantages such as accurate acquisition of three-dimensional information of the target, high resolution, strong anti-interference ability, wide detection range, and near all-weather operation.

[0003] Due to the large size of unmanned vehicles, and the fact that a single laser radar will have a detection blind spot no matter where it is installed on the unmanned vehicle, unmanned vehicles are often equipped with a main laser radar and multiple blind spot radars to achieve the purpose of information complementarity. Figure 1 A common radar installation method for unmanned vehicles is given. A main radar is installed on the top of the unmanned vehicle, and two blind spot radars are installed on the left and right front. Both the main radar and the blind spot radar are traditional mechanical rotating laser radars that can sense objects around 360 degrees.

[0004] Each laser radar gives the coordinates of the detected point cloud in its own coordinate system. To make comprehensive use of all laser radars, calibration is indispensable. The so-called calibration refers to determining the mutual conversion relationship between different coordinate systems, which is the prerequisite for sensor data fusion. Taking unmanned vehicles as an example, it is necessary to calibrate the blind radar and the main radar, determine a three-dimensional rotation matrix and translation vector, and project the point cloud data in the blind radar coordinate system to the main radar coordinate system to achieve point cloud data fusion.

[0005] The common lidar calibration method is to use point cloud registration method, such as the closest point iteration algorithm ICP, GICP, NDT algorithm, etc.

[0006] In such Figure 1 In the radar configuration, due to the different installation positions of the main radar and the blind spot radar, the point cloud coverage of the main radar and the blind spot radar is also different: the closest ground point that the main radar can detect ( Figure 1 The midpoint A) will be closer than the nearest ground point that the blind spot radar can detect ( Figure 1 The midpoint B) is farther away, the former is about 8m, and the latter is about 2m. In the calibration scene, the point cloud within 15m in front is generally collected for matching. At this time, the main radar will have more points that hit objects above the ground, and few points that can hit the ground.

[0007] Existing lidar calibration methods use point cloud registration methods similar to ICP. In this type of method, all point clouds are assigned the same weight, and the sum of point-to-point matching errors is used as the optimization target to iterate the algorithm. However, since there are fewer ground points than non-ground points in the calibration scene, the calibration results will fall into a local minimum. Specifically, non-ground points can be matched well, but there will be a large deviation between the ground points of the main radar and the blind radar. Figure 2 The result after calibration imbalance is shown. The Py area represents the point cloud of the top radar on the wall, the Pg area represents the point cloud of the blind radar on the wall, and Pyg represents the overlapping area of ​​the top main radar and the front blind radar. In the horizontal direction, the two can be well aligned, but due to the scarcity of ground points, they cannot match in the vertical direction.

[0008] Aiming at the calibration characteristics of unmanned vehicles, the present invention proposes a radar calibration method based on an improved point cloud registration method, which can solve the problem of inaccurate calibration caused by too few ground point clouds in traditional calibration algorithms. Summary of the invention

[0009] The technical problem to be solved by the present invention is: in order to overcome the shortcomings in the prior art, the present invention provides an unmanned vehicle radar calibration method based on an improved point cloud registration method, wherein the main radar and the blind radar respectively calculate their own ground equations, and then respectively generate virtual ground points and add them to their respective point clouds, and then use the point clouds after data augmentation for registration.

[0010] The technical solution to be adopted by the present invention to solve the technical problem is: a radar calibration method for an unmanned vehicle based on an improved point cloud registration method, including an unmanned vehicle radar system, the unmanned vehicle radar system including a main radar and at least one blind spot radar, and further including the following steps:

[0011] S1: Generate point cloud data

[0012] Determine an area with obvious point cloud features, so that the main radar and at least one blind spot radar can cover the area with obvious point cloud features, and then use the main radar and the blind spot radar to collect point cloud data of the area with obvious point cloud features;

[0013] S2: Extracting overlapping area point cloud data

[0014] According to the installation positions of the main radar and the blind radar, the point cloud data in the overlapping area are extracted from the point cloud data of the area with obvious point cloud features collected by the main radar and the blind radar, respectively. The point cloud data of the main radar is recorded as P1, and the point cloud data of the blind radar is recorded as P2.

[0015] S3: Calculate ground equations

[0016] According to the point cloud data P1 and P2 of the main radar and the blind radar, the respective ground equations G1 and G2 are calculated;

[0017] S4: Augmented Point Cloud

[0018] According to the ground equation G1 of the main radar, several virtual ground points are randomly generated and added to the point cloud data P1 of the main radar. The point cloud after adding the virtual ground points is recorded as Q1; according to the ground equation G2 of the blind radar, several virtual ground points are randomly generated and added to the point cloud data P2 of the blind radar. The point cloud after adding the virtual ground points is recorded as Q2;

[0019] S5: Fine calibration

[0020] The point cloud registration algorithm is used on point cloud Q1 and point cloud Q2 to obtain the coordinate change relationship between point cloud Q1 and point cloud Q2, and the calibration result of the blind radar to the main radar can be obtained.

[0021] Specifically, the ground equation calculation process in step S3 includes the following steps:

[0022] For a given point cloud P, the process of obtaining the ground equation G is:

[0023] S3.1: Select the N points with the lowest height in the point cloud P and calculate their average height h;

[0024] S3.2: Keep the point cloud P whose height does not exceed h+a, denoted as P G , where a represents half of the height range of the ground point cloud;

[0025] S3.3: Using the random sampling consensus algorithm (RANSAC), the point cloud P G A plane is fitted out from the ground, and the equation of the plane is the ground equation G.

[0026] The process of obtaining the ground equation G in the present invention includes but is not limited to the method of steps S3.1 to S3.3. This method is a common ground equation calculation method. Other algorithms that can realize the ground equation are also applicable to the present invention.

[0027] Preferably, the value range of N is: 500≤N≤3000.

[0028] Preferably, the value range of a is: 0.05~0.3m.

[0029] Preferably, the area with obvious point cloud features includes but is not limited to any one of intersecting walls, large trucks, and several large trees.

[0030] Preferably, the point cloud registration algorithm in step S6 adopts any one of ICP, GICP and NDT.

[0031] Preferably, the main radar and the blind spot radar are both laser radars.

[0032] The beneficial effect of the present invention is that the present invention provides an unmanned vehicle radar calibration method based on an improved point cloud registration method, which overcomes the problem of imbalanced calibration results caused by missing ground points, and can reuse the existing point cloud registration algorithm interface without the need to make internal changes to the point cloud registration algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0034] Figure 1 This is a schematic diagram of the radar installation of an unmanned delivery vehicle.

[0035] Figure 2 This is a schematic diagram showing that too few ground points lead to unbalanced calibration results.

[0036] Figure 3 It is a schematic flow chart of the calibration method of the present invention. DETAILED DESCRIPTION

[0037] The present invention will now be described in detail with reference to the accompanying drawings. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore only shows the components related to the present invention.

[0038] The present invention discloses an unmanned vehicle radar calibration method based on an improved point cloud registration method, including an unmanned vehicle radar system, wherein the unmanned vehicle radar system includes a main radar and at least one blind spot radar. In this embodiment, the main radar is located on the top of the vehicle head, and the blind spot radar is located directly in front of the vehicle head. The area with obvious point cloud features adopts an intersecting wall surface, and the vehicle head is facing the intersection of the wall surface. The overlapping area that can be obtained by the main radar and the blind spot radar is located in the area in front of the vehicle head. Therefore, the main radar and the blind spot radar collect point cloud data of the area in front of the vehicle head as raw data, such as Figure 3 As shown, the specific steps include:

[0039] S1: Generate point cloud data

[0040] The intersecting wall surface is selected as the area with obvious point cloud features, so that both the main radar and the blind radar can cover the intersecting wall surface, and then the main radar and the blind radar are used to collect point cloud data of the intersecting wall surface.

[0041] S2: Extracting overlapping area point cloud data

[0042] According to the installation positions of the main radar and the blind spot radar, the point cloud data located in the overlapping area is extracted from the point cloud data of the intersecting walls collected by the main radar and the blind spot radar, that is, the point cloud data located in front of the vehicle is extracted, among which the point cloud data of the main radar is recorded as P1, and the point cloud data of the blind spot radar is recorded as P2.

[0043] S3: Calculate ground equations

[0044] According to the point cloud data P1 and P2 of the main radar and the blind spot radar, the respective ground equations G1 and G2 are calculated respectively; there are many algorithms for calculating the ground equations of the point cloud, and only one is shown in this embodiment. Other ground algorithms can also achieve the same effect.

[0045] For the point cloud P1 of the main radar, the process of obtaining the ground equation G1 is:

[0046] S3.1-a: Select the N points with the lowest height in the point cloud P1 and calculate their average height h1, where the lowest height refers to the smallest value on the z-axis in the three-dimensional coordinate system; the value range of N is: 500≤N≤3000, and N=1000 is preferred.

[0047] S3.2-a: Keep the point cloud in P1 whose height does not exceed h1+a1, denoted as P G1 In this embodiment, the reference value of a1 is 0.1;

[0048] S3.3-a: Using the random sampling consensus algorithm (RANSAC), the point cloud P G1 A plane is fitted out from the ground, and the equation of the plane is the ground equation G1.

[0049] For the point cloud P2 of the blind radar, the process of obtaining the ground equation G2 is:

[0050] S3.1-b: Select the N points with the lowest height in the point cloud P2 and calculate their average height h2; the value range of N is: 500≤N≤3000, and N=1000 is preferred.

[0051] S3.2-b: Keep the point cloud P2 whose height does not exceed h2+a2, denoted as P G2 In this embodiment, the reference value of a2 is 0.1;

[0052] S3.3-b: Using the random sampling consensus algorithm (RANSAC), P G2 A plane is fitted in the , and the equation of the plane is the ground equation G2.

[0053] S4: Augmented Point Cloud

[0054] According to the ground equation G1 of the main radar, K1 virtual ground points are randomly generated and added to the point cloud data P1 of the main radar. The point cloud after adding the virtual ground points is recorded as the augmented point cloud Q1; according to the ground equation G2 of the main radar, K2 virtual ground points are randomly generated and added to the point cloud data P2 of the main radar. The point cloud after adding the virtual ground points is recorded as the augmented point cloud Q2; preferably, K1 and K2 take the same value. In this embodiment, K1 and K2 take the values ​​of 1000. The randomly generated virtual ground points are generally within a certain range. For example, when the range of the selected point clouds of interest P1 and P2 is 15m, you can choose to randomly generate virtual ground points within 10m. The generated virtual ground points solve the problem of too few ground points in the original point cloud, can balance the number of point clouds on the wall and the ground, and avoid the problem of inaccurate calibration caused by too few ground point clouds when using traditional registration algorithms to register the original point cloud (see Figure 2 ).

[0055] S5: Fine calibration

[0056] The point cloud registration algorithm is used on the point cloud Q1 and the point cloud Q2 to obtain the coordinate change relationship between the point cloud Q1 and the point cloud Q2, and the calibration result of the blind radar to the main radar can be obtained. Among them, the point cloud registration algorithm includes but is not limited to ICP, GICP, NDT, etc.

[0057] The radar calibration method of the present invention first estimates the ground equation of each radar, adds virtual ground points, and then performs point cloud registration, which overcomes the problem of unbalanced calibration results caused by missing ground points. The existing point cloud registration algorithm interface can be reused without making internal changes to the point cloud registration algorithm.

[0058] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the scope of the present invention through the above description. The technical scope of this invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A radar calibration method for an unmanned vehicle based on an improved point cloud registration method, characterized in that: The invention comprises an unmanned vehicle radar system, wherein the unmanned vehicle radar system comprises a main radar and at least one blind spot radar, and further comprises the following steps: S1: Generate point cloud data Determine an area with obvious point cloud features, so that the main radar and at least one blind spot radar can cover the area with obvious point cloud features, and then use the main radar and the blind spot radar to collect point cloud data of the area with obvious point cloud features; S2: Extracting overlapping area point cloud data According to the installation positions of the main radar and the blind radar, the point cloud data in the overlapping area are extracted from the point cloud data of the area with obvious point cloud features collected by the main radar and the blind radar, respectively. The point cloud data of the main radar is recorded as P1, and the point cloud data of the blind radar is recorded as P2. S3: Calculate ground equations According to the point cloud data P1 and P2 of the main radar and the blind radar, the respective ground equations G1 and G2 are calculated; S4: Augmented Point Cloud According to the ground equation G1 of the main radar, several virtual ground points are randomly generated and added to the point cloud data P1 of the main radar. The point cloud after adding the virtual ground points is recorded as Q1; according to the ground equation G2 of the blind radar, several virtual ground points are randomly generated and added to the point cloud data P2 of the blind radar. The point cloud after adding the virtual ground points is recorded as Q2; S5: Fine calibration The point cloud registration algorithm is used on point cloud Q1 and point cloud Q2 to obtain the coordinate change relationship between point cloud Q1 and point cloud Q2, and the calibration result of the blind radar to the main radar can be obtained.

2. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 1, characterized in that: The ground equation calculation process in step S3 includes the following steps: For a given point cloud P, the process of obtaining the ground equation G is: S3.1: Select the N points with the lowest height in the point cloud P and calculate their average height h; S3.2: Keep the point cloud P whose height does not exceed h+a, denoted as P G , where a represents half of the height range of the ground point cloud; S3.3: Using the random sampling consensus algorithm, the point cloud P G A plane is fitted out from the ground, and the equation of the plane is the ground equation G.

3. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 2, characterized in that: The value range of N is: 500≤N≤3000.

4. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 2, characterized in that: The value range of a is: 0.05~0.3m.

5. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 1, characterized in that: The area with obvious point cloud features is any one of intersecting walls, large trucks, and several large trees.

6. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 1, characterized in that: In step S5, the point cloud registration algorithm adopts any one of ICP, GICP and NDT.

7. The unmanned vehicle radar calibration method based on the improved point cloud registration method according to claim 1, characterized in that: The main radar and the blind spot radar are both laser radars.

Images