SLAM with Velodyne Lidar and Jackal UGV

Individual Project, Winter 2019, Northwestern University

Chenge Yang

Demo

SLAM in Simulation (Gazebo)

SLAM in MSR Lab

Pedestrian Detection

Introduction

The goal of this project is to build a mobile robot system in both simulation and real-world environments to achieve SLAM, autonomous navigation and pedestrian detection with Velodyne VLP-16 Lidar sensor and Clearpath Jackal UGV.

Hardware

Clearpath Jackal UGV with Ubuntu 14.04 / ROS Indigo
Velodyne VLP-16 Lidar
PC with Ubuntu 18.04 / ROS Melodic

Simulation

Simulation is established for the Jackal and Velodyne VLP-16 Lidar in Gazebo and rviz.

Implementation

Pipeline Flow

State Estimation

The package robot_localization is used to fuse Odometry with IMU data through Extended Kalman Filter. It estimates the robot state and provides the transformation between tf frames /odom and /base_link.

PointCloud Processing

The raw Lidar PointCloud is processed in C++ using PCL library. The following steps are implemented:

Find the floor plane and remove it from the raw PointCloud
Remove all points that are higher than the robot, lower than the ground and too close to the robot
Get rid of noisy points that have no neighbour points within a particular radius
Downsample to obtain a less dense PointCloud

SLAM

I used the 2D SLAM package gmapping, which implements the Rao-Blackwellized Particle Filter to generate a 2D grip costmap and the the localizaion of the robot in the map (transformation between tf frames /map and /odom).

Navigation

I used package move_base to achieve autonomous navigation of Jackal. It reads the map and localization from gmapping and plans a local and global path for the robot.

Pedestrian Detection

I used package hdl_people_tracking to achieve pedestrian detection and tracking with Velodyne Lidar. It generate semantic segmentation of the human clusters in the PointCloud.

Dependencies and Installation

The Jackal packages are released in ROS Indigo and Kinetic. To use them in ROS Melodic, the following compiling processes are implemented:

Build from apt-get:

Install in Terminal with sudo apt-get install:

ros-melodic-velodyne
ros-melodic-velodyne-description
ros-melodic-velodyne-simulator
ros-melodic-geographic-info
ros-melodic-robot-localization
ros-melodic-twist-mux
ros-melodic-pointcloud-to-laserscan

Build from source: (version: kinetic)

Git clone the original Github repo to local catkin workspace, and run catkin_make.

Usage

Launch in simulation:

roslaunch winter_project simulation.launch

Launch in Jackal:

SSH into Jackal, and run:

roslaunch winter_project real_jackal.launch

Open a new Terminal, and run:

export ROS_MASTER_URI=http://CPR-J100-0076.local:11311
export ROS_HOSTNAME=robostation.local
roslaunch winter_project real_pc.launch

Use rosbag to record data in Jackal

rosbag record
scp -r rosbag.bag [email protected]:~/Downloads

Jackal Setup

Router connection

Connect to JACKALROUTER24; Password: jackbenimble

nmcli connection up JACKALROUTER24

SSH into Jackal

ssh [email protected]
Password: clearpath

Direct Ethernet connection

nmcli connection up Jackal
ssh [email protected]

PS3 Joystick

Plug into the laptop will erase the memory. Reset up required.
Connect the joystick to the Jackal through USB
SSH into jackal
Run

sudo sixpair
sudo sixad --boot-yes

Unplug the joystick, long hold the playstation button

Files on Jackal

catkin_ws workspace for Nate
jackal_ws workspace for Michael
chenge_ws workspace for me
/etc/ros/setup.bash change catkin_make path
/etc/ros/indigo/ros.d/ certain launch file to run when pushing the red button

Copy and remove file to/from Jackal

sudo rm -r winter_project/

scp -r /home/ethan/jackal_ws/src/winter_project/  [email protected]:~/chenge_ws/src

Main issues solved

Z drifting in rviz

The problem is caused by the default settings of robot_localization package. It can be solved by changing the parameters for EKF in the following files:

/jackal_control/config/control.yaml
/jackal_control/config/robot_localization.yaml
/robot_localization/params/ekf_template.yaml

Yaw drifting in rviz

The problem is caused by the noise measurements of IMU. The roll and pitch measurements have gravity as the absolute reference (measured by accelerometer), while the yaw measurement does not. Thus, the magnetometer is enabled for package imu_filter_madgwick to provide the absolute yaw reference. The launch file is located at:

/etc/ros/indigo/ros.d/base.launch

Velodyne Lidar issues

VLP-16.urdf.xacro configuration file for the Velodyne VLP16 Lidar in simulation. Change the param samples at the beginning of the file from 1875 to 200. This will significantly improve the FPS of the package in rviz.
~/jackal_ws/src/velodyne/velodyne_pointcloud/launch/VLP16_points.launch configuration file in Jackal for the velodyne_pointcloud package that transforms the raw Lidar data to ROS message sensor_msgs::PointCloud2. Change the param rpm from 600 to 300. This reduces the publish frequency of topic /velodyne_points, thus allows the Velodyne VLP16 to sample the surrounding environment for all 360 degrees. This setting solves the issue in rviz that the pointcloud is blinking all the time while each frame contains incomplete surrounding information.

PCL library issue in Jackal:

The PCL used in Jackal is version 1.7. Running the node written with PCL will cause a segmentation fault (core dumped). This is because PCL 1.7 does not support C++11. Therefore, I commented the line in CMakeLists.txt that specifying to compile using C++11. Also, the following commands are used in the debugging process:

sudo apt update
sudo apt install gdb
rosrun --prefix 'gdb --args' winter_project floor_removal

michaelchi08 / slam-with-velodyne-lidar-and-jackal-ugv Goto Github PK