3D laser slam: Interpretation of logo-loam paper - Introduction

The title of the paper is ：LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain

• The application scenario given in the title is Variable terrain
• The key is Lightweight and utilize Ground optimization
• The essence is still a Lidar odometer and mapping

Introduction part

Technical background ：

Map building and state estimation are very important functions in intelligent robots

Many people have made great efforts to this , In two ways ：

• Vision based
• Based on lidar

Vision SLAM The advantage of can be very good loop detection , But it is sensitive to the change of light and angle of view

laser SLAM Its advantage is that it can still be used at night , And can get high-precision measurement

Therefore, the paper uses 3d Lidar performs real-time SLAM

The traditional way to solve the pose of two adjacent frames is to iterate the nearest neighbor （ICP）, When the scene is very large , Contains many points , that ICP The method can be very time-consuming . in the light of ICP There are several ways to upgrade , Match the point with the local plane , Face to map matching , And use parallel computing , Improve efficiency .

Later, we will continue to introduce some point cloud registration algorithms

LOAM The advantages of

LOAM It is a low drift and real-time lidar odometer and mapping method

LOAM It is to extract corners and face points to establish constraints , To find the pose transformation between frames . Feature points are judged by calculating the curvature of points

LOAM The real-time performance is achieved by dividing the estimation problem into two independent algorithms , An algorithm runs at high frequencies , Low accuracy estimation of sensor motion . The other algorithm runs less frequently , But it returns high-precision motion estimation .

stay KITTI On dataset , Only through lidar estimation ,LOAM The accuracy of is the best .

LOAM The problem of

Pointed out that LOAM Problems in this working scenario

Description of work scenario ：
The working scene is to install a 3D Laser radar of , To get real-time and reliable 6 Degree of freedom pose estimation . And deploy the algorithm to small-scale embedded systems .

problem ：

• 1 The problem of computation
  Many unmanned aerial vehicles cannot be equipped with powerful computing units

• 2 Problems during intense sports
  When the car runs in a variety of scenes , Due to turbulence , The motion is not very smooth , It causes the data to have motion distortion .（LOAM It is the distortion removal through the uniform velocity model , This is no longer applicable ） Strong motion will also cause abnormal matching of feature points connecting two frames .
  in addition , A large number of laser point clouds are difficult to achieve real-time performance for low-power embedded platforms

When put LOAM Directly use it in the above scenario , When UGV The movement is relatively stable , And the characteristics are stable , When computing resources are sufficient , Low drift motion estimation can be achieved

But resources are limited ,LOAM Your performance will deteriorate . When there are more point clouds in the scene , It is time-consuming to calculate the curvature of each point , Computing on a platform with low computing power can't keep up

• 3 Noise problem
  UGV There is a lot of noise in the operating environment . If the radar is close to the ground , Then the noise on the ground will also affect LOAM The performance of the
  For example, running on the grass , May extract the grass into corners , It's hard to find a match again . Are leaves stable feature points

Put forward LeGO-LOAM Solutions for

Therefore, a lightweight , Optimized by the ground LOAM（LeGO-LOAM）

The classification of point clouds is through ground segmentation first and then , Then remove the unreliable feature points （ Solve the noise problem ）

Due to the optimization based on the ground ,LeGO-LOAM Through two-step optimization to estimate the pose .（ Solve the problem of lightweight ）
First step , By extracting face points from ground points , Then proceed z roll pitch Estimation （ Not estimated x y yaw）
The second step , Through corners x y yaw Estimation

Loop optimization is integrated to correct pose drift .