Introduction

Build

requirements

Basic

## basic depencencies
sudo apt install cmake \
  libboost-dev \
  libeigen3-dev \
  libpng-dev \
  libgoogle-glog-dev \
  libatlas-base-dev

# SuiteSparse and CXSparse (optional)
# - If you want to build Ceres as a *static* library (the default)
#   you can use the SuiteSparse package in the main Ubuntu package
#   repository:
sudo apt-get install libsuitesparse-dev
# - However, if you want to build Ceres as a *shared* library, you must
#   add the following PPA:
sudo add-apt-repository ppa:bzindovic/suitesparse-bugfix-1319687
sudo apt update
sudo apt install libsuitesparse-dev 

ROS

You can refer to http://wiki.ros.org/kinetic/Installation/Ubuntu for more information for installing ROS kinetic or higher version. This repo has been tested in kinetic and melodic.

PCL

## tested in pcl-1.7 (ubuntu16.04) and pcl-1.8 (ubuntu18.04)
## pcl-1.8 or higher version is strongly recommended
sudo apt install libpcl-dev

GTSAM

## Go to your wordspace lisk /home/user/3rd_parties
## GTSAM(4.0 or higher is needed)
git clone https://bitbucket.org/gtborg/gtsam.git
cd gtsam 
mkdir build && cd build
ccmake ..
### important
# set GTSAM_USE_SYSTEM_EIGEN to ON
# then generate makefile
make -j8
sudo make install 

Ceres Solver

you can just install with apt

sudo apt install libceres-dev

or compile with source code

## Ceres Solver(1.12 or higher)
git clone https://github.com/ceres-solver/ceres-solver.git
cd ceres-solver
git checkout tags/1.14.0
mkdir build && cd build
cmake ..
make -j8
sudo make install

libnabo

git clone https://github.com/ethz-asl/libnabo.git
cd libnabo
### checkout to the latest release version
git checkout tags/1.0.7
mkdir build && cd build
cmake ..
make -j8
sudo make install

libpointmatcher

git clone https://github.com/ethz-asl/libpointmatcher.git
cd libpointmatcher
mkdir build && cd build
cmake ..
make -j8
sudo make install

Optional libs

• CUDA: We have made some attempts in fasting the kdtree in ICP by creating the kdtree on GPU, but the GPU Kdtree is not fast enough(just 1.5~2 times faster than libnabo). Notice that if you use CUDA, your g++ version should be lower than 6.0 because the nvcc does not support the 6.0 or high version g++.
• cuda_utils:
• TBB: We have used concurrency containers in TBB for many multi-thread situations, if you turn off this options, the process will use stl containers such as std::vector instead and many multi-thread algorithm will degenerate into single-thread. So, Using TBB is strongly recommended;
• OpenCV: All the matrices in code is in Eigen way, Opencv is only for generating the jpg file of pose gragh. It is a debug function so you can change this option as you wish.

compiling

mkdir build && cd build
cmake ..
make -j8
sudo make install

BUG

• DO NOT use g++ 7.x, unknow bug with Eigen

How to use?

step1 run the mapping process

mkdir pcd
mkdir -p pkgs/test
./mapping.sh

before that, you should kown what is in the script:

## usally, you can leave this config file just like this, it will work fine
CONFIG_PATH=./config/static_mapping_default.xml
## the follow 2 items must be set!!!
## the topic name of your pointcloud msg (ros)
POINT_CLOUD_TOPIC=/fused_point_cloud
## the frame id of your pointcloud msg (ros)
POINT_CLOUD_FRAME_ID=base_link

## the following items are optional
## if you do not have a imu or gps or odom
## just remove the line started with
## imu: -imu -imu_frame_id
## odom: -odom -odom_frame_id
## gps: -gps -gps_frame_id
## and If you got one of these topics
## you MUST provide the tf connection between the one to pointcloud frame
IMU_TOPIC=/imu/data
IMU_FRAME_ID=novatel_imu

ODOM_TOPIC=/navsat/odom
ODOM_FRAME_ID=novatel_odom

GPS_TOPIC=/navsat/fix
GPS_FRAME_ID=novatel_imu

./build/static_mapping_node \
  -cfg ${CONFIG_PATH} \
  -pc ${POINT_CLOUD_TOPIC} \
  -pc_frame_id ${POINT_CLOUD_FRAME_ID} \
  -imu ${IMU_TOPIC} \
  -imu_frame_id ${IMU_FRAME_ID} \
  -odom ${ODOM_TOPIC} \
  -odom_frame_id ${ODOM_FRAME_ID} \
  -gps ${GPS_TOPIC} \
  -gps_frame_id ${GPS_FRAME_ID} 

exit 0 

step2

play bag that includes pointcloud msgs or run the lidar driver

step3

when finished, just press 'CTRL+C' to terminate the mapping process. NOTICE that the mapping process will not end right after you 'CTRL+C', it has many more calculations to do, so just wait.
Finally, you will get a static map like this:

part of it:

Document

You can use doxygen Doxyfile to generate your docs, they are in the doc folder.

References

1. M2DP: A Novel 3D Point Cloud Descriptor and Its Application in Loop,Li He, Xiaolong Wang and Hong Zhang,IEEE International Conference on Intelligent Robots and Systems (2016) 2016-November 231-237
2. ISAM2: Incremental smoothing and mapping using the Bayes tree,Michael Kaess, Hordur Johannsson, Richard Roberts, Viorela Ila, John Leonard, and Frank Dellaert Abstract, International Journal of Robotics Research (2012) 31(2) 216-235
3. Comparing ICP Variants on Real-World Data Sets, Autonomous Robots 2013
4. Fast Segmentation of 3D Pointcloud for Ground Vehicles, M. Himmelsbach and Felix v. Hundelshausen and H.-J. Wuensche, IEEE Intelligent Vehicles Symposium, Proceedings, 2010

TODO

• filtering the cars directly at the input of the pointclouds
• add imu integrating factor in backend
• improve motion filter
• culling data structures like ImuMsg, NavSatMsg, etc.
• add tests
• fix motion compensation when motion filtering happens
• lidar motion compensation after optimization
• lidar motion compensation inside ICP
• ICP using GPU
• stand-alone ICP without libpointmatcher
• parallel PointCloudLib
• finish multi-trajectory map builder
• improve MRVM
• use ground detection to label the pointcloud
• use some machine learning or deep learning method to add semantic labels
• use a docker to release the environment
• get odom message from a cheap GPS and IMU intergration
• fix bug in imu pre-integration (now the imu is just for INS but not normal IMU)
• add support for different kind of GPS (INS&RTK&cheap gps)
• read GPS noise(and other sensors' noise) from config files
• add support for different kind of IMU and ODOM
• add support for more kind of pointclouds
• add a Pose3d struct for simple operation of matrix4f or just use gtsam::pose3d
• using Eigen::Vector instead of sensors::Vector3
• use double instead of float (eigen::matrix)
• using kdtree for loop detection (m2dp search and matching)
• mrvm output to a special format data file and can be transformed to pcd independently
• mrvm output center of voxels
• add more config files and .sh files for other sensor-sets
• using ecef or some other coordinate system instead of utm
• gravity alignment (in pose extrapolator)