A detailed explanation of the laser slam framework logo-loam

The authors introduce ：Zach, Mobile robot practitioners , Love the mobile robot industry , Aspire to science and technology to help a better life .

LOAM The problem is

LeGO-LOAM Its full name is ：Lightweight and Groud-Optimized Lidar Odometry and Mapping on Variable Terrain, It can be seen from the title that LeGO-LOAM Coping with Variable ground Ground optimization has been carried out , At the same time, it ensures lightweight .

LeGO-LOAM It is specially designed for ground vehicles SLAM Algorithm , It is required to install Lidar It can be installed horizontally on the vehicle ; If the installation is inclined , It is also necessary to convert the position and posture to the vehicle . and LOAM Yes Lidar There is no requirement for the installation method of , It doesn't matter if you hold it .

The author's experimental platform is a mobile car （UGA）, There's a Velodyne VLP-16 Line lidar , It is also equipped with a low precision IMU; The selected hardware platform is Nvidia Jetson TX2（ARM Cortex-A57 CPU）; The overall load is 20Kg; Moving speed is ：2.0m/s; The test scenario is ： The ground is uneven （ It's bumpy ） Grass of .

chart 1 Hardware platform

LOAM The framework will have some problems in such hardware environments and usage scenarios ：

1. Because it is equipped with an embedded system , Computing power will be limited ,LOAM The computational needs of will be difficult to meet , As a result, real-time ;
2. If you use LOAM frame , The system calculates the processing frequency of the curvature of each point （ Large amount of data ,VLP-16 One line is 1800 A little bit ） It will be difficult to keep up with the update frequency of the sensor ;
3. UGA The road surface is non smooth and continuous （ Sports are bumpy ）, The collected data will be distorted （ Motion distortion , The uniform motion model cannot be applied to the bumpy scene ）, Use LOAM It is difficult to find a reliable feature correspondence between two frames .
4. Operate in a noisy environment UGV I will LOAM Bring some challenges , for example ： The point cloud of floating grass and swinging leaves will be wrongly extracted as corner points or face points , These characteristics are not reliable , It is difficult to obtain accurate matching between successive frames , This will cause large drift .

LeGO-LOAM Design idea

The registration of ground point cloud mainly uses the features of surface points ; In the segmented point cloud registration, the edge points and face points are mainly used . It can be seen that the number of edge points is much smaller than that of plane points , This is also the main reason for the acceleration .

LeGO_LOAM Software system input for 3D Lidar Point cloud of , Output 6 DOF Pose estimation . The whole software system is divided into 5 Parts of ：

• The first part ：Segmentation： The main operation of this part is to separate the ground point cloud ; At the same time, cluster the remaining point clouds , Filter out a small number of point cloud clusters .
• The second part ：Feature Extraction： For the segmented point cloud （ Ground point clouds have been separated ） Feature extraction of edge points and face points , This step and LOAM The operation inside is the same .
• The third part ：Lidar Odometer ： Between successive frames （ Edge points and face points ） Feature matching finds the pose transformation matrix between successive frames .
• The fourth part ：Lidar Mapping： Yes feature Further treatment , And then in the global point cloud map Registration in .
• The fifth part ：Transform Integration： Transform Integration Integrated from Lidar Odometry and Lidar Mapping Of pose estimation Output the final pose estimate.

LeGO-LOAM Algorithm details

chart 3 In a noisy environment scan Feature extraction process of

A. Segmentation

After extracting ground points , And the rest of it Distance image Clustering （ Clustering ）, The number of filtered point clouds is less than 30 Point cloud cluster of , Assign different labels to the retained point cloud clusters . Ground point cloud belongs to a special kind of point cloud cluster （ We extracted it from the beginning ）. Cluster and reprocess the point cloud , It can improve the running efficiency and extract more stable features . for example , The car runs in a noisy environment , Leaves will produce unreliable features , The same leaf is unlikely to be seen in two consecutive frame scans . chart 3(a) Is the original point cloud , Contains many vegetation point clouds ; After processing, it becomes a graph 3(b), Only the big object point cloud is left , for example ： trunk . The ground point cloud will be retained for further processing . here , Each retained point will have three properties ：（1） Label of point cloud ;（2） stay Distance image Number of rows and columns in ;（3） Distance value .

B. Feature Extraction

This step is mainly to extract features from the ground point cloud and the segmented point cloud , And LOAM Same operation in . It is mainly necessary to understand the concept of several feature point sets .

In order to extract features uniformly from all directions , We will Distance image Divide horizontally into several equal sub images , take 360° Divide evenly 6 Equal division , The precision of each bisection is 300 \times 16（ because VLP-16 Line Lidar, One scan yes 1800 A little bit ）.

Calculate the curvature value of each line of points in the subgraph , Sort the curvature values ,> Cth, It is divided into edge point features ;< Cth It is divided into facial features . Set up the following sets ( The sentences in this part of the paper are similar , Here is my personal understanding , May not be accurate )：

C. Lidar Odometry

1）Label Matching：LeGO-LOAM Cluster the point cloud , Different point cloud clusters have different Label.Label Information can be used as a constraint condition for two frame matching , Only the same label point cloud cluster can be registered between two consecutive frames . This method can improve the accuracy and efficiency of registration .

D. Lidar Mapping

LeGO-LOAM Performance of

The author is testing LeGO-LOAM Performance of , A series of experiments were designed to compare LeGO-LOAM and LOAM Performance .

The author is in small and large outdoor scenes , Respectively for LeGO-LOAM and LOAM The test platform performs intense and gentle control , To see the mapping effect and efficiency of both .

During the intense movement of the small scene ：LOAM Will pull the grass 、 Leaves are extracted as edge points （ Grass and leaves are the main sources of instability ）; and LeGOU-LOAM Will filter out these unstable features , Only in the trunk , ground , Extracting stable features from steps, etc . As shown in the figure below （ Green is the edge point , Pink is pastry ）：

stay LOAM In the frame , Violent movement is easy to cause the divergence of point cloud map , As shown in the figure below (a)LOAM, There are three tree trunks .

The author is in the urban environment of the big scene （ It's a school , The altitude error of different locations is 19m within ） A variety of tests are also carried out to verify the accuracy of the drawing , There are sidewalks , Cement road , Dirt roads and grass .

LOAM The effect of drawing on the sidewalk is not good , It may be the interference of tree leaves at one end , As shown in the figure below ：

In the other three scenarios ,LeGO-LOAM The deviation of ( The deviation of the end point from the initial position ) Perform better than LOAM.

The results of the whole test are as follows ：

• Comparison of the number of feature points ：LeGO-LOAM The overall decline of feature points exceeds ：29%,40%,68%,72%.

• Comparison of iterations ： The number of iterations of odometer is reduced 34%,48%.

• Run time comparison ： To reduce the 60%

• Posture error comparison ：LeGOLOAM Comparable or better position estimation accuracy can be achieved with less computation time .

Reference material

1. https://github.com/RobustFieldAutonomyLab/LeGO-LOAM

remarks ： The author is also us 「3D Vision goes from beginner to proficient 」 Special guests ： A super dry 3D Visual learning community

This article is only for academic sharing , If there is any infringement , Please contact to delete .