KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way
Author: Vizzo
Year: 2023
- odometry estimation based on point to point ICP, adaptive thresholding for matching, motion compensation based on constant velocity model and point cloud downsampling
- Requires very few parameter tuning
- No loop closure
- No Pose Graph Optimization
- 5 steps for current pose estimation and map update:
- Motion prediction and deskewing
- subsample current scan
- Find point to point correspondence between localmap and PC
- registration to find relative motion
- update the local map with subsample cloud
- Deskewing: noting $\mathbf{p}_i^$ the deskewed point and knowing the angular velocity $\omega_t$, linear velocity $v_t$ and the time delay $s_i$ between the point ts and the first point of the scan ts: $$ \boldsymbol{p}_i^=\operatorname{Exp}\left(s_i \boldsymbol{\omega}_t\right) \boldsymbol{p}_i+s_i \boldsymbol{v}_t, $$
- Downsample is done by using a voxel grid and keeping one point per voxel
- Found adventageous to keep a point in the cloud as the coord of the voxel
- Performs frame to local map registration
- Compute a ICP threshold based on a 3 sigma bound computed on all the motions of the local map
- Perform ICP on the Voxel grid with a robust kernel on the cost function