Fast Ground Segmentation Method Based on Lidar Point Cloud

A ground segmentation method based on line fitting of adjacent points was proposed for accurate and real-time segmentation of non-ground information from the LiDAR point cloud. Firstly, the point cloud is divided into several ordered regions depending upon the distribution characteristics of the LiDAR’s concentric circles. Then, the Euclidean distance between adjacent points and the spatial geometric features of ground point clouds is used for adaptive line fitting of ground point clouds. Finally, the ground points are divided by the distance between the adjacent points and the outer points of the line. The experiment was conducted using a real car and the KITTI open-source dataset. The approach presented in this research substantially enhances the accuracy of ground segmentation while ensuring real-time performance.


Introduction
At present, unmanned driving technology has been developing rapidly, among which environmental perception technology [1,2] is a key part of realizing unmanned driving.Due to its benefits of high accuracy, stability, and a broad detection range, LiDAR is used widely in robotics, unmanned driving, and other fields as environmental sensing sensors [3] .
LiDAR accurately describes its surroundings based on its dense point cloud information.However, a substantial amount of point clouds information will consume massive computing resources and affect the real-time, among which the ground point cloud information accounts for a large proportion.Whether the ground point cloud redundant information can be effectively processed is a key part of achieving the precision of environment perception.Numerous investigators are currently researching laser point cloud segmentation.Asvadi et al. used the segmented plane fitting and Random Sampling Consistency (RANSAC) algorithm for surface estimation [4] .Li et al. combined the RANSAC algorithm and normal distribution transformation unit to achieve ground segmentation of threedimensional point clouds [5] .RANSAC algorithm carries out random sampling several times and finds the plane with the best fitting effect from the sampling points.It can find the plane with a better fitting effect with high precision as the ground.However, this method has a slow execution speed, and there are cases of misidentification of walls.Zermas et al. demonstrated a method based on plane fitting to screen fitting points and accelerated the fitting speed [6] .However, the accuracy of the ground segmentation achieved using this method is limited, and it performs best only on gentle pavements.Lim et al. divided the point cloud into different perceptual areas and utilized the plane fitting method for ground plane segmentation [7] .This method showed an improvement in ground segmentation accuracy compared to the GPF algorithm but was more prone to outliers.Therefore, a PatchWork ground segmentation algorithm is proposed.Using the concentric region model to represent the point cloud, the proposed method involves plane fitting to extract the ground, followed by ground segmentation based on the direction, height, and flatness of the ground normal vector [8] .Nonetheless, this approach poses a tendency towards over-segmentation of the ground point cloud.Leng et al. enhanced a 3D point cloud ground segmentation method by integrating height and vertical constraints [9]   .However, while the method does focus on improving the performance of the algorithm in real time, it is less effective in improving its segmentation accuracy.Velas et al. proposed a method for sparse three-dimensional data coding based on convolutional neural networks [10] .Although the method can perform rapid segmentation of point clouds, it needs a complex terrain-labeled dataset and requires high computational resources, thus limiting its applicability.
In summary, most ground segmentation algorithms lack accuracy and speed and only have good applicability for gentle pavement.In this study, we demonstrated an adaptive method for segmenting ground point clouds based on the line fitting of adjacent points and the working principle of concentric scanning rings with different radii generated by LiDAR.The ground and non-ground point cloud are separated by determining the geometric distribution of point clouds regions between adjacent rings, such as the distance, angle, and slope of adjacent points.In Chapter 2, a ground point cloud segmentation method is presented that employs adjacent point line fitting.Chapter 3 focuses on the experimental validation of the proposed method.

Ground Segmentation of Point Clouds based on Adjacent Point Line Fitting
Based on the strong information distribution law of ground LiDAR point cloud, commonly used ground segmentation methods can be classified into these categories: (1) Threshold-based method: Points below the threshold are divided into ground points by setting a height threshold.Although this method is straightforward and simple to use, it lacks robustness.It is difficult to find a threshold that is suitable for all cases.
(2) Plane-fitting-based method: The information containing ground point cloud features fitted into a plane and the position relation of the point cloud outside the fitted plane is compared to determine whether the point cloud belongs to the ground.This method is suitable for flat terrain but cannot be used to detect sloped terrain.
(3) Raster-based method: The collected laser point clouds are projected onto a 2D raster to determine whether the points are part of the ground.This method can adapt to certain terrain changes but is easily affected by the accuracy of the grid.
(4) Adjacent-points-based method: Through the ordering of point clouds, line fitting of adjacency points is carried out to segment ground point clouds.This method has better segmentation accuracy and is suitable for terrain containing ramps.
Combining the vertical and horizontal field of view angles of LiDAR, an adaptive ground segmentation method based on linear fitting of adjacent points is proposed.
Step 1：The LiDAR horizontal field of view angle is evenly divided into 2/  radians of  sector.Each sector area is divided into  concentric rings with a step size of max min  arctan( ) Step 2：Considering the farthest effective information returned by ground point cloud and the segmentation efficiency, the point cloud within a certain distance is processed.To prevent the fitting of vertical plane point cloud in building walls, the ground point cloud should maintain a variation interval of elevation.Candidates are points that satisfy the following Equation (2).
where min ver L and max ver L are the minimum and maximum angle of elevation allowed.
Step 3：The initial ground height ground h is determined by referencing the distance between the Step 5：When the plane line is fitted each time, the height of the current ground point can be obtained and compared with the initial ground height.If the current height exceeds the initial ground height, the slope range should be reduced appropriately to avoid overly steep fitting lines.If the current height is lower than the initial ground height, the slope range can be appropriately increased to avoid too gentle a fitting line.Finally, adaptive adjustment of the maximum slope range is achieved as in Equation (5).Finally, the point clouds are converted to three-dimensional information suitable for publication.

Experimental Verification
To prove the practicality and effectiveness of our approach, we carried out experiments utilizing the platform and KITTI open-source dataset.The experiment was executed on a computer with an i5-7200 processor, and the Linux Robot Operating System (ROS) was used to support reproducing the methodology.The programming language utilized in the experiment's implementation was C++, while Rviz was used to visualize the data.

Scene of an Experiment Conducted on a Real Car
We utilized a vehicle-mounted LiDAR for conducting the experimental verification of our proposed approach.A curved scene located near the experimental building in the campus environment was selected, as illustrated in Figure 2(a).This environment included various natural features such as the curb, green lawn, electric pole, and wall.The purpose of this selection was to examine the effectiveness of our ground segmentation algorithm.Figure 2(b) is the information without ground segmentation processing, marked in Frame 1 as the wall, marked in Frame 2 as the ground point cloud in the road area, and marked in Frames 3 and 4 are the weeds and trees outside the road area.After processing by our algorithm, as shown in Figure 2, the pavement reflected from the road ring information and the low weed point cloud from the surrounding environment are correctly identified and classified as fragments of the location cloud.And the wall information in Frame 1 is not divided into ground point clouds.In the experimental diagram shown in Figure 2(d), the ground point cloud information in the road area (Frame 2) has been filtered.In contrast, the information on slightly higher weeds in the surrounding environment (Frames 3 and 4) has been retained, and non-ground point cloud information has finally been extracted.The experimental results demonstrate that the proposed methodology for ground segmentation based on fitting adjacent points to a line effectively performs the segmentation and filtration tasks.

Performing Experiments on Dataset
The KITTI dataset was utilized for experiments, and the proposed method was compared to RANSAC, GPF, and Patchwork ground segmentation algorithms to conduct a comprehensive evaluation from the aspects of Precision, Recall, and algorithm time, as in Equation ( 6).

Precision=
, Recall= In Figure 3(a)(b), the red information marked in the black box shows that there is a certain oversegmentation when RANSAC and GPF are used, which mistakenly classifies non-ground point clouds as ground point cloud information.This is due to the fact that the RANSAC algorithm sacrifices much information to ensure the recall rate of the ground plane segmentation but cannot guarantee the segmentation accuracy.The GPF algorithm takes the point clouds at a certain height from the ground as the ground candidate points and fits the plane through continuous iterations.This method inevitably has the phenomenon of over-segmentation.In contrast, the Patchwork algorithm has a certain level of adaptability to this part of detailed point clouds.However, the black point clouds marked in the black box in Figure 3(c) incorrectly classified ground points as non-ground points.It is important that our method has some robustness to this region.A total of 4540 KITTI datasets were extracted, and the proposed method was compared with RANSAC, GPF, and Patchwork ground segmentation algorithms.The running time of the proposed method and the other three algorithms were sampled every 10 frames, and the resulting data was used to create the running time comparison diagram shown in Figure 4.The results demonstrated that our developed algorithm has the shortest running time.We calculated the mean value of the experimental data in Table 1.In terms of Precision, Recall, and running time of the algorithm, the proposed method is superior to RANSAC and GPF.Compared with Patchwork, Precision and the time taken by the algorithm have certain advantages on the basis of a small difference in the Recall index.

Conclusion and Discussion
Combined with the geometric distribution of point cloud ground information, an adaptive ground segmentation method based on line fitting of adjacent points was proposed.We verified the superiority of our method through experiments with the KITTI dataset and a real-world vehicle.The results show that our method achieves accurately segmented ground point clouds, outperforming RANSAC, GPF, and Patchwork algorithms.Segmentation accuracy can reach 0.9689, Recall rate can reach 0.9807.In addition, under-segmentation is inevitable in the ground segmentation method based on linear fitting, which is also the main reason for the slightly low recall rate in this paper.Future work will address this problem.
ordered division of point clouds in each area, The three-dimensional LiDAR information ( , , ) x y z is converted into two-dimensional LiDAR information ( , ) dz, and the elevation angle  of LiDAR point is calculated as in Equation (1).

1 L 2 L 1  2 Figure 1 .
Figure 1.Adjacency distance diagram.Step 4：The coordinates of the new input point cloud are ( , ) dz .The z-coordinate is calculated by inputting the d-coordinate into the line fit from the previous step.The average error It iterates continuously.The ground candidate point cloud information in all regions is traversed, and the ground points are screened by the fitted straight lines.The distance error between the outer point and the adjacent line is calculated to determine whether it can be classified as the ground point.

Figure 3 .
Effect of the dataset and experimental details.

Table 1 .
Comparison table of ground plane segmentation.