M2DGR： a Multi-modal and Multi-scenario SLAM Dataset for Ground Robots

First Author: Jie Yin

Figure 1. Sample Images

Updates

2022.12.13 Our new work coming soon! Codes and datasets will be available in https://github.com/sjtuyinjie/M2DGR-plus and https://github.com/SJTU-ViSYS/Ground-Fusion upon paper acceptance.

2022.9.13 welcome to follow and star our new work: Ground-Challenge at https://github.com/sjtuyinjie/Ground-Challenge. Feel free to propose issues if needed.

2022.06.20 Thanks Jialin Liu (Fudan University) for his work to test LVI-SAM on M2DGR. Following is the link of their modified LVI-SAM version link. Refer Link for detailed information. And the configuration files for LVI-SAM on M2DGR are given in launch file,camera file and lidar file. Feel free to test the demo on your machine!

LVI-SAM on M2DGR

2022.02.18 We have upload a brand new SLAM dataset with GNSS, vision and IMU information. Here is our link SJTU-GVI. Different from M2DGR, new data is captured on a real car and it records GNSS raw measurements with a Ublox ZED-F9P device to facilitate GNSS-SLAM. Give us a star and folk the project if you like it.

2022.02.01 Our work has been accepted by ICRA2022!

NOTICE

We strongly recommend that the newly proposed SLAM algorithm be tested on our data, because our data has following features:

1. A rich pool of sensory information including vision, lidar, IMU, GNSS,event, thermal-infrared images and so on
2. Various scenarios in real-world environments including lifts, streets, rooms, halls and so on.
3. Our dataset brings great challenge to existing SLAM algorithms including LIO-SAM and ORB-SLAM3. If your proposed algorihm outperforms SOTA systems on M2DGR, your paper will be much more convincing and valuable.

ABSTRACT:

We introduce M2DGR: a novel large-scale dataset collected by a ground robot with a full sensor-suite including six fish-eye and one sky-pointing RGB cameras, an infrared camera, an event camera, a Visual-Inertial Sensor (VI-sensor), an inertial measurement unit (IMU), a LiDAR, a consumer-grade Global Navigation Satellite System (GNSS) receiver and a GNSS-IMU navigation system with real-time kinematic (RTK) signals. All those sensors were well-calibrated and synchronized, and their data were recorded simultaneously. The ground truth trajectories were obtained by the motion capture device, a laser 3D tracker, and an RTK receiver. The dataset comprises 36 sequences (about 1TB) captured in diverse scenarios including both indoor and outdoor environments. We evaluate state-of-the-art SLAM algorithms on M2DGR. Results show that existing solutions perform poorly in some scenarios. For the benefit of the research community, we make the dataset and tools public.

Keywords:Dataset, Multi-model, Multi-scenario,Ground Robot

MAIN CONTRIBUTIONS:

• We collected long-term challenging sequences for ground robots both indoors and outdoors with a complete sensor suite, which includes six surround-view fish-eye cameras, a sky-pointing fish-eye camera, a perspective color camera, an event camera, an infrared camera, a 32-beam LIDAR, two GNSS receivers, and two IMUs. To our knowledge, this is the first SLAM dataset focusing on ground robot navigation with such rich sensory information.
• We recorded trajectories in a few challenging scenarios like lifts, complete darkness, which can easily fail existing localization solutions. These situations are commonly faced in ground robot applications, while they are seldom discussed in previous datasets.
• We launched a comprehensive benchmark for ground robot navigation. On this benchmark, we evaluated existing state-of-the-art SLAM algorithms of various designs and analyzed their characteristics and defects individually.

Video

For Chinese users, try

1.LICENSE

This work is licensed under MIT license. International License and is provided for academic purpose. If you are interested in our project for commercial purposes, please contact us on [email protected] for further communication.

If you face any problem when using this dataset, feel free to propose an issue. And if you find our dataset helpful in your research, simply give this project a .

The paper has been accepted by both RA-L and ICRA 2022. A preprint version of the paper in Arxiv and IEEE RA-L.If you use M2DGR in an academic work, please cite:

@ARTICLE{9664374,
  author={Yin, Jie and Li, Ang and Li, Tao and Yu, Wenxian and Zou, Danping},
  journal={IEEE Robotics and Automation Letters}, 
  title={M2DGR: A Multi-sensor and Multi-scenario SLAM Dataset for Ground Robots}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/LRA.2021.3138527}}

2.SENSOR SETUP

2.1 Acquisition Platform

Physical drawings and schematics of the ground robot is given below. The unit of the figures is centimeter.

Figure 2. The GAEA Ground Robot Equipped with a Full Sensor Suite.The directions of the sensors are marked in different colors,red for X,green for Y and blue for Z.

2.2 Sensor parameters

All the sensors and track devices and their most important parameters are listed as below:

• LIDAR Velodyne VLP-32C, 360 Horizontal Field of View (FOV),-30 to +10 vertical FOV,10Hz,Max Range 200 m,Range Resolution 3 cm, Horizontal Angular Resolution 0.2°.

• RGB Camera FLIR Pointgrey CM3-U3-13Y3C-CS,fish-eye lens,1280*1024,190 HFOV,190 V-FOV, 15 Hz

• GNSS Ublox M8T, GPS/BeiDou, 1Hz

• Infrared Camera,PLUG 617,640*512,90.2 H-FOV,70.6 V-FOV,25Hz;

• V-I Sensor,Realsense d435i,RGB/Depth 640*480,69H-FOV,42.5V-FOV,15Hz;IMU 6-axix, 200Hz

• Event Camera Inivation DVXplorer, 640*480,15Hz;

• IMU,Handsfree A9,9-axis,150Hz;

• GNSS-IMU Xsens Mti 680G. GNSS-RTK,localization precision 2cm,100Hz;IMU 9-axis,100 Hz;

• Laser Scanner Leica MS60, localization 1mm+1.5ppm

• Motion-capture System Vicon Vero 2.2, localization accuracy 1mm, 50 Hz;

The rostopics of our rosbag sequences are listed as follows:

• LIDAR: /velodyne_points

• RGB Camera: /camera/left/image_raw/compressed ,
  /camera/right/image_raw/compressed ,
  /camera/third/image_raw/compressed ,
  /camera/fourth/image_raw/compressed ,
  /camera/fifth/image_raw/compressed ,
  /camera/sixth/image_raw/compressed ,
  /camera/head/image_raw/compressed

• GNSS Ublox M8T:
  /ublox/aidalm ,
  /ublox/aideph ,
  /ublox/fix ,
  /ublox/fix_velocity ,
  /ublox/monhw ,
  /ublox/navclock ,
  /ublox/navpvt ,
  /ublox/navsat ,
  /ublox/navstatus ,
  /ublox/rxmraw

• Infrared Camera:/thermal_image_raw

• V-I Sensor:
  /camera/color/image_raw/compressed ,
  /camera/imu

• Event Camera:
  /dvs/events,
  /dvs_rendering/compressed

• IMU: /handsfree/imu

3.DATASET SEQUENCES

We make public ALL THE SEQUENCES with their GT now.

Figure 3. A sample video with fish-eye image(both forward-looking and sky-pointing),perspective image,thermal-infrared image,event image and lidar odometry

An overview of M2DGR is given in the table below:

Scenario	Street	Circle	Gate	Walk	Hall	Door	Lift	Room	Roomdark	TOTAL
Number	10	2	3	1	5	2	4	3	6	36
Size/GB	590.7	50.6	65.9	21.5	117.4	46.0	112.1	45.3	171.1	1220.6
Duration/s	7958	478	782	291	1226	588	1224	275	866	13688
Dist/m	7727.72	618.03	248.40	263.17	845.15	200.14	266.27	144.13	395.66	10708.67
Ground Truth	RTK/INS	RTK/INS	RTK/INS	RTK/INS	Leica	Leica	Leica	Mocap	Mocap	---

3.1 Outdoors

Figure 4. Outdoor Sequences:all trajectories are mapped in different colors.

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
gate_01	2021-07-31	16.4g	172s	dark,around gate	Rosbag	GT
gate_02	2021-07-31	27.3g	327s	dark,loop back	Rosbag	GT
gate_03	2021-08-04	21.9g	283s	day	Rosbag	GT

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
Circle_01	2021-08-03	23.3g	234s	Circle	Rosbag	GT
Circle_02	2021-08-07	27.3g	244s	Circle	Rosbag	GT

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
street_01	2021-08-06	75.8g	1028s	street and buildings,night,zigzag,long-term	Rosbag	GT
street_02	2021-08-03	83.2g	1227s	day,long-term	Rosbag	GT
street_03	2021-08-06	21.3g	354s	night,back and fourth,full speed	Rosbag	GT
street_04	2021-08-03	48.7g	858s	night,around lawn,loop back	Rosbag	GT
street_05	2021-08-04	27.4g	469s	night,staight line	Rosbag	GT
street_06	2021-08-04	35.0g	494s	night,one turn	Rosbag	GT
street_07	2021-08-06	77.2g	929s	dawn,zigzag,sharp turns	Rosbag	GT
street_08	2021-08-06	31.2g	491s	night,loop back,zigzag	Rosbag	GT
street_09	2021-08-07	83.2g	907s	day,zigzag	Rosbag	GT
street_010	2021-08-07	86.2g	910s	day,zigzag	Rosbag	GT
walk_01	2021-08-04	21.5g	291s	day,back and fourth	Rosbag	GT

3.2 Indoors

Figure 5. Lift Sequences:The robot hang around a hall on the first floor and then went to the second floor by lift.A laser scanner track the trajectory outside the lift

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
lift_01	2021-08-04	18.4g	225s	lift	Rosbag	GT
lift_02	2021-08-04	43.6g	488s	lift	Rosbag	GT
lift_03	2021-08-15	22.3g	252s	lift	Rosbag	GT
lift_04	2021-08-15	27.8g	299s	lift	Rosbag	GT

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
hall_01	2021-08-01	29.1g	351s	randon walk	Rosbag	GT
hall_02	2021-08-08	15.0g	128s	randon walk	Rosbag	GT
hall_03	2021-08-08	20.5g	164s	randon walk	Rosbag	GT
hall_04	2021-08-15	17.7g	181s	randon walk	Rosbag	GT
hall_05	2021-08-15	35.1g	402s	circle	Rosbag	GT

Figure 6. Room Sequences:under a Motion-capture system with twelve cameras.

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
room_01	2021-07-30	14.0g	72s	room,bright	Rosbag	GT
room_02	2021-07-30	15.2g	75s	room,bright	Rosbag	GT
room_03	2021-07-30	26.1g	128s	room,bright	Rosbag	GT
room_dark_01	2021-07-30	20.2g	111s	room,dark	Rosbag	GT
room_dark_02	2021-07-30	30.3g	165s	room,dark	Rosbag	GT
room_dark_03	2021-07-30	22.7g	116s	room,dark	Rosbag	GT
room_dark_04	2021-08-15	29.3g	143s	room,dark	Rosbag	GT
room_dark_05	2021-08-15	33.0g	159s	room,dark	Rosbag	GT
room_dark_06	2021-08-15	35.6g	172s	room,dark	Rosbag	GT

3.3 alternative indoors and outdoors

Figure 7. Door Sequences:A laser scanner track the robot through a door from indoors to outdoors.

Sequence Name	Collection Date	Total Size	Duration	Features	Rosbag	GT
door_01	2021-08-04	35.5g	461s	outdoor to indoor to outdoor,long-term	Rosbag	GT
door_02	2021-08-04	10.5g	127s	outdoor to indoor,short-term	Rosbag	GT

4. CONFIGURERATION FILES

For convenience of evaluation, we provide configuration files of some well-known SLAM systems as below:

A-LOAM,

LeGO-LOAM,

LINS,

LIO-SAM,

VINS-MONO,

ORB-Pinhole,

ORB-Fisheye,

ORB-Thermal,

CUBMAPSLAM

5.DEVELOPMENT TOOLKITS

5.1 Extracting Images

• For rosbag users, first make image view

roscd image_view
rosmake image_view
sudo apt-get install mjpegtools

open a terminal,type roscore.And then open another,type

rosrun image_transport republish compressed in:=/camera/color/image_raw raw out:=/camera/color/image_raw respawn="true"

• For non-rosbag users,just take advantage of following script export_tum,export_euroc and get_csv to get data in formats of Tum or EuRoC.

5.2 Evaluation

We use open-source tool evo for evalutation. To install evo,type

pip install evo --upgrade --no-binary evo

To evaluate monocular visual SLAM,type

evo_ape tum street_07.txt your_result.txt -vaps

To evaluate LIDAR SLAM,type

evo_ape tum street_07.txt your_result.txt -vap

To test GNSS based methods,type

evo_ape tum street_07.txt your_result.txt -vp

5.3 Calibration

For camera intrinsics,visit Ocamcalib for omnidirectional model. visit Vins-Fusion for pinhole and MEI model. use Opencv for Kannala Brandt model

For IMU intrinsics,visit Imu_utils

For extrinsics between cameras and IMU,visit Kalibr For extrinsics between Lidar and IMU,visit Lidar_IMU_Calib For extrinsics between cameras and Lidar, visit Autoware

5.4 Getting RINEX files

For GNSS based methods like RTKLIB,we usually need to get data in the format of RINEX. To make use of GNSS raw measurements, we use Link toolkit.

5.5 ROS drivers for UVC cameras

We write a ROS driver for UVC cameras to record our thermal-infrared image. UVC ROS driver

6.FUTURE PLANS

In the future, we plan to update and extend our project from time to time, striving to build a comprehensive SLAM benchmark similar to the KITTI dataset for ground robots.

If you have any suggestions or questions, do not hesitate to propose an issue. And if you find our dataset helpful in your research, a simple star is the best affirmation for us.

7.ACKNOWLEGEMENT

This work is supported by NSFC(62073214). Authors from SJTU hereby express our appreciation.