M2dgr slam data set of multi-source and multi scene ground robot

M2DGR: Multi source and multi scene Ground robots SLAM Data sets

Paper：https://arxiv.org/pdf/2112.13659.pdf

Source ：ICRA2022 & RAL2021

Project address ：https://github.com/SJTU-ViSYS/M2DGR

Speaker ：M2DGR One work was handed over to master Yinjie , Tutor Professor zoudanping ;

ICRA Official share ：https://www.bilibili.com/video/BV1q3411G7iF

Paper Literature Review： Temporary vacancy ;

Notes

The outline ：

• 1. Multi-source SLAM Current situation of the development of
• 2. Main stream SLAM Data set review
• 3.M2DGR Acquisition platform and environment
• 4. Experimental evaluation and result analysis
• 5. Data set usage guide

1 Multi-source SLAM Current situation of the development of

The main contents include ： Vision SLAM、 laser SLAM、 Multi source fusion ;

among ： Multi source fusion SLAM Method Can effectively improve SLAM Accuracy and robustness , It is one of the hottest fields at present ;

VSLAM It is mainly divided into the above four modules ; It is mainly divided into traditional methods and learning based methods ; At present, academia will think ORB-SLAM3 Performance ratio of VINS-Mono Better , This is TUM-VI and EuRoC Conclusions drawn on two data sets , But the speaker found through experimental verification ： On cars and real cars ,VINS-Mono Performance will be more stable , It would be better ,ORB-SLAM3 It is easy to track failures on such a platform ; If it is on a small scale or by a large number of loops ,ORB-SLAM3 It can reach decimeter level , Even centimeter accuracy .SVO2 Compared with the SVO, Added multi-purpose support 、 Back end optimization and loopback detection , Comparable performance ORB-SLAM3;DM-VIO Is the latest job . It is a system of delayed marginalization , Performance can also be comparable ORB-SLAM3; The above algorithms have their own advantages and disadvantages on different data sets , Can be regarded as SOTA Algorithm ;

Based on learning SLAM In the system ,Droid-slam It's all SLAM The positioning accuracy is the highest , Its presence TUM-RGBDI and EuRoC The performance of is far better than ORB-SLAM3, It can achieve centimeter accuracy in all sequences , Its disadvantage is that it consumes a lot of computing resources , I need two 3090 To run in real time ;NICE-SLAM It is the orientation of Zhejiang University SLAM Dynamic neural implicit scalable coding , Neural implicit is more suitable for rendering dense geometry （ such as IMAP, however Imap It's hard to deal with large scenes ,NICE-SLAM Can handle large scenes ）;TANDEM It shows good real-time 3D reconstruction performance .

summary ： Traditional methods consume less resources , But the constructed map is relatively sparse ; The accuracy of the learning based method will be better , The map is also much denser , But the generalization ability is poor .

Vision SLAM It is sensitive to illumination and texture information , But laser SLAM It is unaffected ; General situation , laser SLAM Accuracy and robustness ratio in large scenes VSLAM Higher ; laser SLAM The disadvantage is , Difficult to relocate （ Cause drift ）、 Build a sparse map 、 The high cost ;Cartographer In the paper, the test accuracy can reach 5cm even to the extent that 3cm; mechanical 3D Lidar can 360° rotate ; solid state 3D Laser radar is used most livox, But solid 3D Lidar has a fixed FOV;

Vision and laser fusion ： Earlier work V-LOAM（2014, Non open source ）;

GNSS The fusion ： In the outdoor scene, the global coordinates can be obtained to eliminate the drift ;

Event camera fusion ： It can solve the problems of motion blur and overexposure of traditional cameras ;

Wheel speedometer fusion ：VINS on wheels It solves the degradation problem in some scenes , At present, the scene has been very perfect .

The speaker believes that the most promising thing at present is GNSS and SLAM Fusion , because GNSS and SLAM Are very complex algorithms , And not many people do this .

2 Main stream SLAM Data set review

KITTI: It's old and there are few sensors ; Vision SLAM Generally, this data set is not used , And use TUM( hold RGBD) and EUROC( Unmanned aerial vehicle (uav) );

NCLT: Ground robot acquisition , Camera acquisition frequency is low , Fewer sensors ;

OpenLORIS: Ground robot acquisition , Camera internal and external parameters are not published ;gt The acquisition is not rigorous , By laser SLAM Ran out of ;

3 M2DGR Acquisition platform and environment

Effect drawing of data collection ;

Acquisition platform , The specific equipment is as above ;.GNSS-IMU Integrated navigation is mainly used to collect outdoor RTK Signal as outdoor track gt;lidar It is connected to the laptop through the Internet port . IMU It adopts domestic consumer grade equipment (500-800rmb)

(a) The outdoor sequence is handed in for collection , These sequences are relatively long , The existing SLAM The effect of the algorithm running on it is not ideal , It belongs to a very challenging data set ;

(b)Roomdark It's in a completely dark scene , Used to compare ordinary cameras Performance of infrared camera and event camera ;

The purpose is to test indoor and outdoor alternating SLAM Algorithm performance ;

(d) This sequence is also very challenging ;

4 Experimental evaluation and result analysis

It's on it 7 A representative sequence ;

Visualization of positioning results ; The conclusion is : Spooky: ! The drift is very obvious .

5 Data set usage guide

The information of several data sets is listed above .

Students or teachers can connect to the campus network , You can achieve 10M/S .

VSLAM There is no absolute scale for a single purpose , Need to add - s (scale). If your algorithm can be more stable than the above SOTA Algorithm ( For example, your algorithm is better than orb-slam3 Higher than 2-3 rice ), congratulations , You can also send an article ICRA 了

The above is the process of using data sets . The author will keep the data set updated !

Q&A

1. Camera and IMU Calibration use calib , Good effect and simple calibration ;

2. Nine axis IMU Than six axis IMU It can collect more three-dimensional information .

3.KITTI Data set of IMU There may be a problem with the dataset , Not recommended KITTI Running vision SLAM The system of .

4.NUC 3000-5000 The performance of the left and right sides is not very good , It is recommended to use the top equipped NUC.