A ground segmentation method and system based on multi-radar fusion

Abstract

The application discloses a ground segmentation method and system based on multi-radar fusion, which can ensure the continuity of point calculation between adjacent regions and avoid the problem of point loss. The ground segmentation method based on multi-radar fusion comprises the following steps: calculating the index range of each point scanned by the radar on the vehicle body and the center region of the vehicle body, and calculating the first final center normal vector and the first final ground center point of the center region of the vehicle body; finally, the points in the range of [-∞, v n +thres ground ] are ground points, and the remaining points are non-ground points; extracting a point set P unseg in an adjacent region which has not been segmented as a ground point and calculating the second final center normal vector and the second final ground center point; finally, the points in the range of [-∞, v l +thres ground ] are ground points, and the remaining points are non-ground points. The application can eliminate the points that have been calculated, only calculate the ground points that have not been segmented, avoid the problem of repeated calculation, and improve the ground point segmentation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of segmentation algorithm, and particularly relates to a ground segmentation method and system based on multi-radar fusion. BACKGROUND

[0002] In the field of outdoor unmanned driving, laser radar is usually used for obstacle detection, and ground segmentation is a basic task in obstacle detection using laser radar. Through ground segmentation, the point cloud of the ground is removed, and the remaining point cloud is the obstacle point cloud.

[0003] In the actual environment, the ground environment is complex, and the main task of ground segmentation is to extract the area that the vehicle can pass through to ensure the normal operation of the vehicle. In the prior art, the simplest way to perform ground segmentation is to filter by height, and the point cloud below a certain height is directly segmented as ground. However, if the height is set too low, many gently raised ground may be identified as obstacles, which will cause the unmanned vehicle to be unable to drive normally. If the height is set too high, some relatively short obstacles may be missegmented as ground, resulting in missed detection of obstacles. In addition, there are some drivable areas on slopes in outdoor environments, such as the transition from a motorway to a pedestrian walkway. If the ground in this area cannot be correctly extracted, the unmanned vehicle will not be able to pass through these two areas.

[0004] In addition, some algorithms in the prior art use local ground segmentation to achieve good segmentation results, but still cannot be applied to some special scenarios. In these algorithms, only one radar is considered. In actual use, multiple radars are usually used for blind filling. After the radars are calibrated, there will be certain errors. These errors will not affect ground segmentation in the near distance, but when the distance reaches a certain point, the ground point cloud will appear layered, and the farther the position, the more obvious the phenomenon. Local ground segmentation will extract the lowest layer of point cloud as ground, and the remaining ground point cloud scanned by the radar above may be segmented as obstacles. In this case, even an open ground will have obstacle point clouds. In other scenarios, for point cloud height mutations and missing ground point clouds, because the calculation is based on a local area, the continuity of the ground is not considered, which will also cause segmentation errors.

[0005] In addition, these algorithms need to set parameters according to the scanning angle of the radar, which is dependent on the type and installation position of the device, and is not conducive to the update and maintenance of actual application.

[0006] Therefore, the prior art has defects and needs to be improved. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art, provide a ground segmentation method and system based on multi-radar fusion, solve the problem of poor continuity in point cloud segmentation in the prior art, which leads to segmentation errors; solve the problem that local ground segmentation will extract the lowest layer of point cloud as the ground, which will cause the remaining ground point cloud scanned by the radar above to be segmented as obstacles.

[0008] The technical scheme of the present application is as follows: a ground segmentation method based on multi-radar fusion, the method comprising the following steps:

[0009] Calculate the index range of each point scanned by each radar belonging to the region; and calculate the index range of the center region of the vehicle body, to obtain the center region point set P center .

[0010] Calculate the first final center normal vector and the first final ground center point of the center region of the vehicle body, and take the points in the range of [-∞, v n +thres ground ] as ground points, and the remaining points as non-ground points.

[0011] The thres ground is a ground height threshold; the v n is the result of projecting the first final ground point onto the first final center normal vector.

[0012] Select the region with the highest evaluation score from the regions that have been calculated for ground segmentation and have adjacent regions that have not been calculated for ground segmentation as the propagation region, and obtain its adjacent regions that have not been calculated for ground segmentation; after this step, the second final center normal vector and the second final ground center point of the ground of the remaining regions can be calculated in order according to the optimal propagation path.

[0013] Extract the point set P unseg that has not been segmented as a ground point in the adjacent region that has not been calculated for ground segmentation, and calculate the second final center normal vector and the second final ground center point corresponding to the adjacent region that has not been calculated for ground segmentation, and project P unseg onto the second final center normal vector, take the points in the range of [-∞, v l +thres ground ] as ground points, and the remaining points as non-ground points. The v l is the result of projecting the second final ground point onto the second final center normal vector.

[0014] Preferably, the calculation step of the index range of each point scanned by each radar belonging to the region is:

[0015] Define the starting position of the scene in the x, y plane as p st (x st , y st );

[0016] Locating point p i The region encompassed by (x, y) in the x and y directions:

[0017]

[0018] step x =ll o (2);

[0019] step y =ww o (3);

[0020] Where l is the side length of the region in the x-direction, w is the side length of the region in the y-direction, and l o w is the overlap length in the x-direction. o This represents the overlap length in the y-direction; it should be noted that... All values ​​are integers; furthermore, "and" in the above formula represent rounding up and rounding down, respectively. The index range of the region to which each point scanned by each radar belongs is calculated through the above steps, thus calculating the index range corresponding to all points.

[0021] Preferably, the formula for calculating the index range of the vehicle body center region is:

[0022]

[0023] The point cloud set of the central region of the vehicle body is extracted to obtain the central region point set P. center Based on the above calculation results, the point cloud set belonging to the central region, i.e., P, can be extracted. center .

[0024] Preferably, the steps for calculating the first final center normal vector and the first final ground center point of the vehicle body center region are as follows:

[0025] Projecting the points in the vehicle's central region and the initial ground center point onto the initial normal vector, we obtain the set V. center and v0;

[0026]

[0027] in, Let p be the initial normal vector. init This is the initial ground center point.

[0028] From the set V center The selection range is [v0-thres] down v0+thres upThe plane normal vector is calculated for the set of points corresponding to the inner boundary; where thres down and thres up These are the lower and upper thresholds, which are set according to the actual scenario.

[0029] Take the eigenvector corresponding to the smallest eigenvalue in the calculated plane normal vector as the new normal vector, and take the centroid of the point set as the new ground center point. Then, take the point set P of the central region... center And the new ground center point is projected onto the new normal vector, from the set V center The selection range is [v1-thres] down v1+thres up The point set corresponding to the area within the boundary is used to calculate the plane normal vector again. This process is repeated n times to obtain the first final center normal vector and the first final ground center point. n is greater than or equal to 0, and the specific number of repetitions depends on the actual needs. Obviously, v0, v1, ..., v n All are obtained through the above calculation formula (7), only the data input is different. Specifically, when calculating v0, p is the initial normal vector of the central region of the vehicle body. init The initial ground center point is used, and when calculating v1, p is the new normal vector obtained during the first calculation of the plane normal vector. init For point set P center The centroid, similarly, can be substituted into the result of the nth plane normal vector calculation for v. n The calculation.

[0030] Because there is overlap between adjacent regions, some points may have already been segmented into ground points during the calculation of the previous region. These points do not need to be recalculated to avoid duplication; therefore, these points do not belong to the point set P. unseg .

[0031] Preferably, the calculation steps for the second final center normal vector and the second final ground center point corresponding to the adjacent regions without ground segmentation are as follows:

[0032] Point set P unseg By projecting the initial center point onto the initial normal vector, we can calculate the set V. reg and v l-1 and V reg [v] l-1 -thres down v l-1 +thres up The new point set formed by the points within ] and the point set P ground Merge into point set P cal ;

[0033] perform plane normal vector calculation on the point set P cal , take the eigenvalue corresponding to the minimum eigenvalue in the calculation result as the second final center normal vector, and take the centroid of the point set P cal as a new second final ground center point. After the above calculation is completed, the point set P unseg and the second final ground center point can be projected onto the second final center normal vector, and the points in the range of [-∞, v l + thres ground ] are further taken as ground points, and the remaining points are non-ground points, so as to realize the segmentation of the ground points and the non-ground points in the region. In addition, it should be noted that the calculation formula of v l-1 and v l is the above calculation formula (7), and the calculation principle is the same as above.

[0034] Preferably, the calculation formula of the evaluation score of the region which has been calculated ground segmentation and has adjacent regions which have not been calculated ground segmentation is:

[0035]

[0036] wherein Count g is the number of points in the calculation region which are segmented as ground points, and Count all is the number of points of all points in the scene. The number of points of all points in the scene is the total number of points obtained by each radar scan.

[0037] Preferably, the plane normal vector calculation adopts the least square method or the SVD method.

[0038] The application also provides a ground segmentation system based on multi-radar fusion, comprising the above-mentioned ground segmentation method based on multi-radar fusion.

[0039] By adopting the above scheme, the ground segmentation method and system based on multi-radar fusion have the following beneficial effects:

[0040] 1. Adjacent regions have overlapping regions and overlapping points, and when ground segmentation is performed, the calculation of points between adjacent regions is ensured to be continuous, the continuity of ground segmentation is ensured, and the problem of point loss is avoided;

[0041] 2. By calculating the index range of each point belonging to a region, the region to which each point belongs can be calculated, so that when the unsegmented region is calculated, the points which have been calculated can be eliminated, and only the unsegmented ground points are further calculated, so that the problem of repeated calculation is avoided, and the calculation efficiency is improved;

[0042] 3, the first final central normal vector, the first final ground center point, the second final ground center point and the second final central normal vector are data after multiple iteration calculations, and the accuracy in the range of [-∞, v n +thres ground ] and [-∞, v l +thres ground ] is high, and the ground point segmentation precision is improved.

[0043] 4, the method and the system are only related to the installation height and the area of the radar on the vehicle body, and are irrelevant to the line scanning angle of the radar, so that it is not necessary to adjust the related parameters according to the line scanning angle of the radar, convenience is strong, and different types of radars can be compatible, and the compatibility is strong. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 it is a schematic diagram of the adjacent area x direction overlap state of the application;

[0045] Fig. 2 it is a schematic diagram of the adjacent area y direction overlap state of the application;

[0046] Fig. 3 it is a schematic diagram of the point projection to the initial normal vector of the application. DETAILED DESCRIPTION

[0047] The application will be described in detail below in combination with the drawings and specific embodiments.

[0048] Please refer to Figs. 1-3 , the application provides a ground segmentation method based on multi-radar fusion, which comprises the following steps:

[0049] Step S1: calculate the index range of each point scanned by each radar on the vehicle body belonging to each area; and calculate the index range of the center area of the vehicle body, to obtain the center area point set P center .

[0050] S10: calculate the step length of the adjacent area moving in the x direction and the y direction;

[0051] step x =l-l o (2);

[0052] step y =w-w o (3);

[0053] Wherein, l is the edge length of the area in the x direction, w is the edge length of the area in the y direction, l o is the overlap length in the x direction, and w o is the overlap length in the y direction. In this embodiment, l, l o, w, w o The values of step x , step y are 8m, 4m, 6m, 3m respectively, thus, step st = 3m, step st = 2m.

[0054] In addition, it should be noted that in the present embodiment, the size of any one region is 8m x 6m, and the overlapping length of adjacent regions in the x direction is 4m and the overlapping length in the y direction is 3m; of course, in the actual use process, it can also be adjusted according to the actual needs.

[0055] S11: define the starting position of the scene in the x, y plane as p st (x st , y st ).

[0056] In the scene, the coordinates of the center of the vehicle body are generally selected as the original coordinates, the minimum position of the vehicle body center in the x direction and the y direction can be selected as the starting position, and in order to make the center of the vehicle body coincide with the center of the region, the selection of the starting position needs to meet the following conditions: Both are integers.

[0057] In the present embodiment, the coordinates of the point of the minimum position of the vehicle body center in the x direction and the y direction are (-19.54, -9.63), and p st (x st , y st ) is (-21, -10) through the above constraint relationship (integer).

[0058] S12: calculate the index range of each point belonging to the region respectively; specifically calculated by the following formula:

[0059]

[0060] In the above formula, "and" respectively represents the rounding up and rounding down.

[0061] The index range of the point (-19.54, -9.63) is calculated, and the index range of the point (-19.54, -9.63) in the x direction is -1 to 0, thus, the index starts from 0, and the region index in the x direction is 0, and similarly, the region index in the y direction is 0.

[0062] Specifically, in the present embodiment, the index range of the region to which the point (26.165, -3.65) belongs is calculated by the above formula (1), and the result is that the index range thereof in the x direction is 10-11, and the index range thereof in the y direction is 3-4, thus the point (26.165, -3.65) is simultaneously in the four regions with indexes (10, 3), (11, 3), (10, 4), and (11, 4). Since there is an overlap between adjacent regions, after the result of calculating the index range of the region to which a point belongs is obtained, a point can belong to multiple regions simultaneously, which is to make the adjacent regions have an overlap, so that when the ground segmentation is calculated, the adjacent regions are not independent of each other, and the problems of missing of some ground points or sudden change of height are avoided, and the ground segmentation failure can be avoided when the ground segmentation is performed later.

[0063] S13: Calculate the index range of the vehicle center region; specifically, calculate by the following formula:

[0064]

[0065] In the present embodiment, the coordinates of the vehicle center region are (0, 0), and it is known from the previous calculation that step x = 3m, step y = 2m, p st (x st , y st ) is (-21, -10), thus row center = 6, col center = 4 can be calculated by formula (4), that is, the index of the center region in the x and y directions is (6, 4).

[0066] According to the results of the above steps S10-S13, the point cloud set of the vehicle center region is extracted to obtain the center region point set P center .

[0067] Step S2: Calculate the first final center normal vector and the first final ground center point of the vehicle center region, and take the points in the range of [-∞, v n + thres ground ] as ground points, and the rest of the points as non-ground points.

[0068] S20: Project the points of the vehicle center region and the initial ground center point onto the initial normal vector to calculate the set V center and v0;

[0069]

[0070] Obviously for v0, is the initial normal vector of the vehicle center region, which is set to (0, 0, 1) in the present embodiment, and pinit The initial ground center point is set to (0, 0, -1.9) in this embodiment. The process of projecting points onto relevant normal vectors (initial normal vector, first final center normal vector, second final center normal vector, etc.) involved in this invention can be referred to... Fig. 3 .

[0071] The initial ground center point and initial normal vector are set according to the actual situation. Generally, the initial normal vector is chosen as (0, 0, 1), that is, along the Z-axis of the coordinate system. The initial ground center point needs to be set according to the height of the coordinate system center point above the ground. The height of the coordinate system center point above the ground is measured. hcenter =1.9m, therefore the initial ground center point can be set to (0, 0, -1.9).

[0072] S21: From the set V center The selection range is [v0-thres] down v0+thres up The point set corresponding to the value within the range is used for SVD calculation; where the thresholds are... down and thres up These are the lower and upper thresholds, respectively, set according to the actual scenario. In this embodiment, thres down thres up The values ​​are 0.1 and 0.2 respectively.

[0073] S22: Take the eigenvector corresponding to the smallest eigenvalue in the calculated plane normal vector as the new normal vector, and take the centroid of the point set as the new ground center point, and set the point set P of the central region... center The new ground center point is projected onto the new normal vector, and from the set V center The selection range is [v1-thres] down v1+thres up The point set corresponding to the value within ] is then subjected to SVD calculation again; for v1, p is the new normal vector obtained during the first calculation of the plane normal vector. init For point set P center Center of mass.

[0074] S23: Repeat step S22 n times to obtain the first final center normal vector and the first final ground center point. In this embodiment, repeating step S22 once is sufficient. This process is for iterative optimization; generally, once the value of n is set, the iteration will automatically exit after reaching the corresponding number of iterations. During this process, v n The calculation process is as follows: Obviously, in this process, The eigenvector corresponding to the minimum eigenvalue calculated after repeating step S22 once, The center region point set P center calculated after repeating step S22 once,

[0075] S24: Project the point set P center and the first final ground center point onto the first final center normal vector, and take the points in the range of [-∞, v n +thres ground ] as ground points, and the rest as non-ground points. Wherein thres ground is a ground height threshold, which is 0.2m in this embodiment.

[0076] Obviously, through the above steps S2 (S20-S24), it can be judged whether the points in the point set P center belong to ground points or non-ground points; specifically, when judging whether the points in the point set P center belong to ground points or non-ground points, it is necessary to project them onto the first final center normal vector, in combination with the result of projecting the first final ground center point onto the first final center normal vector, to make the judgment. Taking the point (2.35, 1.86, -1.8) in the point set P center as an example, the first final center normal vector is (0.01, 0.062, 0.998), and the first final ground center point is (0, 0, -1.9). The result v n of projecting it onto the first final center normal vector is -1.89, and the value of the point (2.35, 1.86, -1.8) after being projected onto the first final center normal vector is -1.658, v n +thres ground is -1.69, and -1.658 is greater than -1.69, so it does not belong to the range of [-∞, -1.69], and therefore it is a non-ground point.

[0077] S3: Select the region with the highest evaluation score from the regions whose ground segmentation has been calculated and which have adjacent regions whose ground segmentation has not been calculated as the propagation region, and obtain the adjacent regions whose ground segmentation has not been calculated.

[0078] Specifically, taking the normal vector and the ground center point of the propagation region as the initial values, the ground segmentation of the adjacent regions and the evaluation score of the region are calculated:

[0079]

[0080] Wherein, Count g is the number of points in the calculation region that are segmented as ground points, and Count all is the number of points in all points in the scene. In the test scene of this embodiment, the total number of points Countall is 32561, Count g The evaluation score of the central region is 0.047, which can be calculated by the above formula (5).

[0081] The step S3 can be repeated to calculate the evaluation scores of the adjacent regions of all the regions "whose ground segmentation has been calculated and which have adjacent regions whose ground segmentation has not been calculated". The point set P unseg Because there is an overlapping part between the regions, some points have been calculated in the previous region and segmented as ground points, so these points do not need to be calculated again.

[0082] S4: After this step, the second final central normal vector and the second final ground central point of the ground of the remaining regions can be calculated in order according to the optimal propagation path.

[0083] S40: Project the point set P unseg and the initial central point onto the initial normal vector to calculate the set V reg and v l-1 .

[0084]

[0085] wherein, is the initial normal vector, which is set to (0, 0, 1) in this embodiment, and p init is the initial ground central point, which is set to (0, 0, -1.9) in this embodiment. The same as in step S2.

[0086] S41: Select a new point set formed by the points belonging to the range [v reg -thres l-1 , v down +thres l-1 ] from the set V up and merge it with the point set P ground to form the point set P cal (to be calculated). Wherein, thres down and thres up are the lower threshold and the upper threshold, respectively, and in this embodiment, the values of thres down , thres up are 0.1 and 0.2, respectively.

[0087] S42: Perform SVD calculation on the point set P cal , and take the feature corresponding to the smallest eigenvalue in the calculation result as the second final central normal vector, and take the centroid of the point set P cal as the new second final ground central point.

[0088] S5: projecting P unseg and the second final ground center point onto the second final center normal vector, taking the points in the range [-∞, v l +thres ground as ground points, and the rest as non-ground points. The value of thres ground is consistent with the value of thres ground calculated in step S2.

[0089] Specifically, in the present embodiment, the application also provides a ground segmentation system based on multi-radar fusion, comprising the above-mentioned ground segmentation method based on multi-radar fusion. After the ground segmentation system of the application adopts the above-mentioned ground segmentation method based on multi-radar fusion, it can realize the rapid segmentation of ground points and non-ground points, ensure the continuity of data, and has high accuracy.

[0090] In summary, the ground segmentation method and system based on multi-radar fusion in the application can calculate the region to which each point belongs by calculating the index range of each point, and calculate the index of the center region of the vehicle body. By combining the above calculation process, the point set belonging to the center region can be extracted, the first final ground center point and the first final center normal vector of the ground in the center region of the vehicle body are calculated, and the value range [-∞, v n +thres ground ] of the ground points in the center region of the vehicle body is calculated. After projecting the points in the point set P center onto the first final center normal vector, it is determined whether the result belongs to [-∞, v n +thres ground ], that is, whether the points in the point set P center are ground points or non-ground points. For points in non-center regions of the vehicle body, the adjacent principle can be used to select the region with the highest evaluation score from the regions that have been calculated for ground segmentation and have adjacent regions that have not been calculated for ground segmentation as the propagation region, to obtain the adjacent regions that have not been calculated for ground segmentation, further extract these adjacent regions that have not been segmented for calculation, and eliminate the points that have been calculated, only further calculating the ground points that have not been segmented, which can avoid the problem of repeated calculation, and ensure the continuity of the calculation of points between adjacent regions and the continuity of ground segmentation. Further calculation of the point set P unsegWhether the point inside belongs to the ground point, all collected points are calculated and judged to belong to the ground point or non-ground point in sequence, all points are calculated and judged to prevent the problem of point missing. In addition, since the ground center point and the normal vector are continuously optimized in the calculation process, the segmentation accuracy is guaranteed to be high, the problem that the local ground segmentation in the prior art extracts the lowest point cloud as the ground, and the remaining ground point cloud scanned by the radar above may be segmented as an obstacle is solved, and the accuracy of the ground point segmentation is guaranteed to be high.

[0091] The above are only preferred embodiments of the present application and are not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-radar fusion based ground segmentation method, characterized in that, The method comprises the following steps: Calculate the index range of the region to which each point scanned by each radar on the vehicle body belongs; and calculate the index range of the central region of the vehicle body to obtain the central region point set P center ; Compute the first final center normal vector and the first final ground center point of the center region of the vehicle body, take the points in the range of [-∞, v n +thres ground ] as ground points, and the rest as non-ground points; the thres ground is a ground height threshold; the v n is a result of projecting the first final ground point onto a first final central normal vector; The step of calculating the first final center normal vector of the vehicle body center region and the first final ground center point is: Projecting the point of the center region of the vehicle body and the initial ground center point onto the initial normal vector, a set V is calculated center and v0; selecting a point set corresponding to a range of [v0-thres down , v0+thres up ] from the set V center , and performing plane normal vector calculation; thres down and thres up are lower and upper threshold values, respectively; Taking the eigenvector corresponding to the minimum eigenvalue in the plane normal vector calculation result as a new normal vector, taking the centroid of the point set as a new ground center point, projecting the center area point set P center and the new ground center point onto the new normal vector, selecting the point set corresponding to the range [v1-thres down , v1+thres up ] from the set V center , and performing plane normal vector calculation again, the process is repeated n times to obtain a first final center normal vector and a first final ground center point; The region with the highest evaluation score is selected as the propagation region from the regions with calculated ground segmentation and the regions with adjacent regions without calculated ground segmentation, and the adjacent regions without calculated ground segmentation are obtained; extracting a point set P in the adjacent region of the non-computed ground segmentation which is not segmented as ground points unseg and computing the second final center normal vector and the second final ground center point corresponding to the adjacent region of the non-computed ground segmentation, and projecting P unseg to the second final center normal vector, taking the points in the range of [-∞, v l +thres ground ] as ground points, and the rest as non-ground points; the v l is the result of the projection of the second final ground point onto the second final central normal vector; The step of calculating the second final center normal vector and the second final ground center point corresponding to the adjacent regions without calculated ground segmentation is: Project the point set P unseg and the initial center point onto the initial normal vector to obtain the set V reg and v l-1 , and the set V reg is obtained by merging the point set P l-1 and the new point set formed by the points in V down that belong to [v l-1 -thres up , v ground +thres cal ]; The point set P cal The plane normal vector is calculated, and the eigenvalue corresponding to the minimum eigenvalue in the calculation result is taken as the second final center normal vector, and the centroid of the point set P cal is taken as the new second final ground center point.

2. The method according to claim 1, wherein, The step of calculating the index range of the region to which each point scanned by each radar belongs is: Define the starting position of the scene in the x, y plane as p st (x st , y st ) Positioning point p i (x, y) Range of area to which it belongs in x, y directions: step x = 1 - l o (2); step y = w - w o (3); wherein "and" represents upward rounding and downward rounding, respectively, l is the side length in the x direction of the region, w is the side length in the y direction of the region, l o is the overlap length in the x direction, w o is the overlap length in the y direction.

3. The method of claim 2, wherein, The calculation formula of the index range of the vehicle body center region is: extracting a point cloud set of a center region of the vehicle body to obtain a center region point set P center .

4. The method of claim 1, wherein, The calculation formula of the evaluation score of the region with calculated ground segmentation and the adjacent regions without calculated ground segmentation is: where Count g is the number of points in the computed region that are segmented as ground points, Count all is the number of points in the scene.

5. The method of claim 1, wherein, The method used for calculating the plane normal vector is the least square method or the SVD method.

6. A multi-radar fusion based ground segmentation system, characterized in that, A ground segmentation method based on multi-radar fusion comprising any one of claims 1-5.

Images