gisbi-kim/PyICP-SLAM

Full-python LiDAR SLAM using ICP and Scan Context

PyICP SLAM

Full-python LiDAR SLAM.

• If you find a C++ version of this repo, go to SC-LeGO-LOAM or SC-A-LOAM.

Purpose

• Full-python LiDAR SLAM
  • Easy to exchange or connect with any Python-based components (e.g., DL front-ends such as Deep Odometry)
    • Here, ICP, which is a very basic option for LiDAR, and Scan Context (IROS 18) are used for odometry and loop detection, respectively.
• Hands-on LiDAR SLAM
  • Easy to understand (could be used for educational purpose)
• The practical use case of miniSAM
  • The miniSAM is easy to use at Python

What is SLAM?

• In this repository, SLAM (Simultaneous localization and mapping) is considered as
  • SLAM = Odometry + Loop closure
• In this repository, the state to be optimized only has robot poses; that is pose-graph SLAM.

Overview of the PyICP SLAM

• The pipeline of the PyICP SLAM is composed of three parts
  1. Odometry: ICP (iterative closest point)
     • In here, Point-to-point and frame-to-frame (i.e., no local mapping)
  2. Loop detection: Scan Context (IROS 18)
     • Reverse loop detection is supported.
  3. Back-end (graph optimizer): miniSAM
     • Python API

Features

• Thanks to the Scan Context, reverse loops can be successfully closed.

  • E.g., see KITTI 14 at Results section below.

• Time costs

  • (Here, no accelerated and naive) ICP gets 7-10 Hz for randomly downsampled points (7000 points)
  • (Here, no accelerated and naive) Scan Context gets 1-2 Hz (when 10 ringkey candidates).
    • if you need the more faster one, consider the C++ version of Scan Context (e.g., https://github.com/irapkaist/SC-LeGO-LOAM)
  • miniSAM is enough fast.

How to use

Just run

$ python3 main_icp_slam.py

The details of parameters are eaily found in the argparser in that .py file.

Results (KITTI dataset)

Those results are produced under the same parameter conditions:

• ICP used random downsampling, 7000 points.
• Scan Context's parameters:
  • Ring: 20, Sector: 60
  • The number of ringkey candidates: 30
  • Correct Loop threshold: 0.17 for 09, 0.15 for 14, and 0.11 for all others

Results (left to right):

• 00 (loop), 01, 02 (loop), 03

• 04, 05 (loop), 06 (loop), 09 (loop)

• 10, 11, 12, 13 (loop)

• 14 (loop), 15 (loop), 16 (loop), 17

• 18 (loop), 20

Some of the results are good, and some of them are not enough. Those results are for the study to understand when is the algorithm works or not.

Findings

• The Scan Context does not find loops well when there is a lane level change (i.e., KITTI 08, as below figures).

• If the loop threshold is too low (0.07 in the below figure), no loops are detected and thus the odometry errors cannot be reduced.

  • 

• If the loop threshold is high (0.20 in the below figure), false loops are detected and thus the graph optimization failed.

  • 
  • but using this non-conservative threshold with a robust kernel would be a solution.
    • see https://github.com/irapkaist/SC-LeGO-LOAM/blob/3170ba1de7bb0d3f8de5192742e0bbe267415ab4/SC-LeGO-LOAM/LeGO-LOAM/src/mapOptmization.cpp#L994

Author

  Giseop Kim (paulgkim@kaist.ac.kr)

Contirbutors

  @JustWon
    - Supports Pangolin-based point cloud visualization along the SLAM poses.
    - Go to https://github.com/JustWon/PyICP-SLAM