mini-pyllio

• A very minimal but working Python only implementation for loosely-coupled LiDAR-Inertial odometry for education purpose.

What's this?

• 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.
    1. A LiDAR prevents an IMU diverges.
    2. 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 ...
• 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.

How to use?

The steps

1. Download the KITTI data from KITTI dataset.
2. Prepare the dataset directory structure like this.
3. Build an image: cd docker && docker build .
4. change the host-side volume path as yours in the docker_run.sh
5. Just sh docker_run.sh. Enjoy!

You may like ...

• This video-based(TBA) guide.

Experiments on KITTI odometry dataset

05 (2011_09_30_0018)

1. Exhaustive comparisons

• 

2. IMU only (with external GT rotation) vs. mini-pyllio

• 

Others

• 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)

Future work

• A set of additional efforts could enhance the speed/accruacy.
  • For example,
    1. I used scan-to-scan matching. Scan-to-submap matching would definitely helpful.
    2. Not explicit replacement of the state, but trying weighting the propagated state and the correction such as Kalman filter.
    3. Fusion with a cheap GPS.
• But, this is a minimal codebase for an educational purpose, thus I left them for you. Do it yourself!

Contact

• Giseop Kim, [email protected]