PFilter
-
This work is the official implementation of "PFilter: Building Persistent Maps through Feature Filtering for Fast and Accurate LiDAR-based SLAM", which saves 20.9% processing time per frame and improves the accuracy by 9.4% than FLOAM.
-
This code is modified from FLOAM.
[Demo Video] [Preprint Paper]
Modifier: Yifan Duan, University of Science and Technology of China, China
1. Demo
-
Edge map and surface map on KITTI dataset:
-
Map in USTC with 32-lines LiDAR:
2. Prerequisites and build
- It's similar to FLOAM, please see FLOAM for the build details.
3. Usage
-
To test on KITTI dataset and get the same ATEs as the shown in the Sec. 4:
roslaunch pfilter pfilter_kitti.launch bag_filename:=/YOUR/BAG/PATH
-
To test on other data:
roslaunch pfilter pfilter.launch bag_filename:=/YOUR/BAG/PATH k_new:=xx theta_p:=xx theta_max:=xx topic:=xx- The xx should be replaced with suitable value.
- "k_new " is Int $\in [0,\infty)$, but [0,5] is recommanded.
- "theta_p " is Float $\in [0,\infty)$, but [0,3] is recommanded.
- "theta_max " is Int $\in [0,255]$.
- "topic" is the ros topic of LIDAR points.
-
Some notes:
- The parameters "k_new", "theta_p" and "theta_max" can control how hard PFilter works, which means how many noise points will be removed from the map.
- k_new $\downarrow$, theta_p $\uparrow$, theta_max $\uparrow$ means more points are removed.
- Generally speaking, if the robot is moving slow, I usually let PFilter remove more points to keep the map tiny. For example, I use a low-speed unmanned vehicle on campus, equipped with a 32-lines LiDAR for mapping, and the parameters I use are "0, 1, 200". In comparison, the parameters used on KITTI dataset are "0, 0.4, 75".
- To get the performance of origin FLOAM, you can set the parameters to "0,0,0".
4. Evluation
| KITTI sequence | FLOAM | PFilter | KITTI sequence | FLOAM | PFilter |
|---|---|---|---|---|---|
00 |
0.7007% | 0.6208% | 06 |
0.5435% | 0.4906% |
01 |
1.9504% | 1.8055% | 07 |
0.4159% | 0.3740% |
02 |
0.9549% | 0.8042% | 08 |
0.9349% | 0.9326% |
03 |
0.9549% | 0.8941% | 09 |
0.7031% | 0.6242% |
04 |
0.6875% | 0.6420% | 10 |
1.0257% | 0.9206% |
05 |
0.4910% | 0.5085% | all |
0.8511% | 0.7833% |
- The tool for evaluation is KITTI odometry evaluation tool.
5.Acknowledgements
6. Citation
If you use this work for your research, you may want to cite
@inproceedings{duan2022pfilter,
title={PFilter: Building Persistent Maps through Feature Filtering for Fast and Accurate LiDAR-based SLAM},
author={Duan, Yifan and Peng, Jie and Zhang, Yu and Ji, Jianmin and Zhang, Yanyong},
booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
pages={11087--11093},
year={2022},
organization={IEEE}
}