# LiDAR Odometry

> Satyajit Ghana — Head of Engineering @ Inkers Technology
> canonical: https://ai.thesatyajit.com/notes/lidar-odometry
> date: 2026-05-28

LiDAR odometry estimates a sensor's motion by aligning consecutive point-cloud
scans. Each new scan is matched against a reference — the previous scan or a local
map — and the rigid transform that best aligns them is the frame-to-frame motion.
Integrate those transforms and you have a trajectory.

The matching step is exactly [[point-cloud-registration]]: the odometry front-end is a
registration problem solved fast enough to run online. Modern systems fuse an IMU to
constrain the optimization (LiDAR-inertial odometry), which keeps the registration
well-conditioned during fast motion and feature-poor scenes.

## Why it is hard

- Scans are sparse and unstructured, so naive nearest-neighbour matching is costly.
- Drift accumulates: small per-frame errors compound over a long trajectory.
- Degenerate geometry (long corridors, open fields) leaves the alignment
  under-constrained along some axes.

Most of the engineering effort goes into making [[point-cloud-registration]] both
robust and cheap enough to close the loop every frame.