- A very minimal but working Python only implementation for loosely-coupled LiDAR-Inertial odometry for education purpose.
- Recently, Pypose made a great example code for the imu integration with an easy-to-follow python-based code. see https://github.com/pypose/pypose/tree/main/examples/module/imu. This is called inertial odometry.
- In that example, the ground-truth rotation from the expensive GPS+INS system is injected into the propagtaion step.
- As you will see below, without the external rotation source, the imu itself quickly diverges.
- I simply replaced it by scan-to-scan registration from LiDAR sensor (with Open3D ICP). Then, the corrected PVA (position, velocity, and attitude) is set as the estimator's newer state. I think this can be said loosely coupled lidar-inertial odometry.
- The point is: LiDAR and IMU help each other.
- A LiDAR prevents an IMU diverges.
- The IMU provides a good initial guess for the LiDAR scan matching.
- and then, the LiDAR again better prevents an IMU diverges... and go on and on ...
- The point is: LiDAR and IMU help each other.
- This simple project, which only used Python and a few files, is intended to be educational.
- I expect, after playing it, a reader could be able to answer why IMU and LiDAR should be fusioned.
- Download the KITTI data from KITTI dataset.
- Prepare the dataset directory structure like this.
- Build an image:
cd docker && docker build . - change the host-side volume path as yours in the docker_run.sh
- Just
sh docker_run.sh. Enjoy!
- This video-based(TBA) guide.
- Exhaustive comparisons
- IMU only (with external GT rotation) vs. mini-pyllio
- Try them yourself! (the below list was summarized and provided by Y. Cho)
- 00: 2011_10_03_drive_0027 (000000 to 004540)
- 02: 2011_10_03_drive_0034 (000000 to 004660)
- 05: 2011_09_30_drive_0018 (000000 to 002760)
- 06: 2011_09_30_drive_0020 (000000 to 001100)
- 07: 2011_09_30_drive_0027 (000000 to 001100)
- 08: 2011_09_30_drive_0028 (001100 to 005170)
- 09: 2011_09_30_drive_0033 (000000 to 001590)
- 10: 2011_09_30_drive_0034 (000000 to 001200)
- A set of additional efforts could enhance the speed/accruacy.
- For example,
- I used scan-to-scan matching. Scan-to-submap matching would definitely helpful.
- Not explicit replacement of the state, but trying weighting the propagated state and the correction such as Kalman filter.
- Fusion with a cheap GPS.
- For example,
- But, this is a minimal codebase for an educational purpose, thus I left them for you. Do it yourself!
- Giseop Kim,
[email protected]