M5-T5 — LOAM Architecture

Learning objectives

After this lesson, you should be able to:

• Describe the two modules of LOAM, their rates, inputs and outputs.
• Write the edge and planar correspondence residuals of the odometry cost.
• Explain how motion distortion is handled inside the optimisation.
• Say why the split makes the pipeline real-time and what it gives up.

The problem and the intuition

Walking in the dark with a torch, you do two things at once. Every step, you glance at the wall next to you and keep your direction without thinking: quick and a little sloppy. Every few seconds you stop and compare what you see with the plan of the building in your head: slow and careful, and it corrects the drift of the quick checks. LOAM is that: odometry is the quick glance at the previous scan; mapping is the careful comparison with the whole map.

Registration against a big map is accurate but slow; against the previous scan it is fast but drifts. Do both, and use the slow one to correct the fast one.

The algorithm: odometry, mapping and integration

Problem. Estimate the sensor’s pose at the scan rate (10 Hz) and build a map, from a continuous stream of 3D LiDAR sweeps, in real time on a modest computer, with low drift and without an IMU.

Assumptions. The sweeps have scan-line structure (a spinning sensor); the scene has enough planar and edge structure; the motion during a sweep is smooth (constant velocity); the environment is static.

Lidar odometry (10 Hz)

1. Features. From the new sweep P_k+1 extract edge and planar points (M5-T3).
2. Correspondences with the previous sweep. For an edge point, find the two closest edge points on different scan lines in the previous sweep: they define a line. For a planar point, find three closest planar points: a plane.
3. Distances. The point-to-line distance d_e and the point-to-plane distance d_p.
4. Motion during the sweep. The pose of a point at time t within the sweep is linearly interpolated between the start (identity) and the end pose T: the unknown 6-DOF transform T_k+1 enters each point’s distance through its own interpolated fraction s of the motion.
5. Minimise Σ d_e² + Σ d_p² over the 6 parameters with Levenberg-Marquardt, with robust (bisquare) weights that down-weight large residuals; iterate, re-computing correspondences, until convergence.

s_i

the fraction of the sweep elapsed when point i was measured (0 to 1).

d_e, d_p

point-to-line and point-to-plane distances (M5-T4) of the interpolated point to its matched target.

T

the sensor’s motion over one sweep (6 degrees of freedom).

Lidar mapping (1 Hz)

The odometry output (the undistorted sweep and its pose) is accumulated and, once per second, matched to the map: all points seen so far, stored in cubes that are loaded around the sensor. Many more features (about ten times as many) are used, the correspondences are found in a local region of the map through fitting a line or plane to the 5 nearest neighbours, and the same cost is minimised for a more accurate pose. The matched sweep is added to the map.

Transform integration

Mapping gives an accurate but delayed transform (once per second); odometry gives a transform every 0.1 s that drifts. The output pose at 10 Hz is the mapping pose composed with the odometry motion since the last mapping update. Each mapping result corrects the accumulated odometry drift.

Strengths, limitations, failure modes

A worked example: how the two rates cooperate

A robot drives at 2 m/s for 10 s (20 m). Odometry (scan-to-scan) accumulates a drift of 1.0% of the path length; the mapping corrects the pose once per second.

1. Odometry alone: after 20 m the position is wrong by 0.20 m. Each sweep adds 2 cm of the error in the direction of travel (20 cm over 100 sweeps).
2. Mapping at 1 Hz. After each second the pose is matched against the accumulated map, whose geometry was built with earlier corrections. Its error relative to the map is about the registration accuracy, say 3 cm, not growing with distance travelled within the mapped area. The correction removes the drift accumulated in that second (about 2 cm per metre × 2 m = 4 cm) before it can accumulate.
3. Pose at 10 Hz. At t = 7.35 s the latest mapping pose is from t = 7.0 s. The output is T_map(7.0 s) ⊕ [odometry motion from 7.0 to 7.35 s]. The odometry motion over 0.35 s (70 cm) has error of about 0.7 cm (1%), so the output stays within a few centimetres.
4. Time budget. Odometry has about 100 ms per sweep and uses a few hundred features and small neighbourhoods: typically 10–30 ms. Mapping uses ten times as many features and a k-d tree over the map: 200–800 ms, which fits a 1 Hz budget. If mapping were run at 10 Hz it would never keep up; if odometry were skipped the initial guess for mapping would be poor, and mapping would converge slowly or to a wrong match.
5. What the result looks like. On the KITTI odometry benchmark LOAM reported about 0.9% relative translation error at the time of publication: roughly 1 m per 100 m. That is the drift rate left after mapping; without loop closure it remains.

The lab asks you to implement a simplified version of both: scan-to-scan odometry on 2D features, then a refinement against a map.

Engineering implications and limitations

Engineering implications

• Respect the interface. Odometry gives the mapping stage its initial guess and its deskewed cloud; a mistake in either (wrong timestamps, wrong rate) shows up in the map, not in the odometry.
• Manage the map. Store points in cubes or voxels, keep only those near the sensor in the search structure and downsample (5–20 cm). Without this the cost grows with the length of the run.
• Skip frames in mapping, not in odometry. Mapping every second or every few frames is the intended mode; dropping scans in odometry increases the motion between them and breaks the constant-velocity deskewing.
• Add what LOAM lacks. Fuse an IMU for deskewing and for initial guess (M5-T8, M5-T9), add loop closure with a pose graph (M5-T10), and a degeneracy check.
• Diagnose by module. If the odometry path looks smooth but the map is blurred, mapping is at fault; if the path is jumpy, odometry or feature extraction is.

Limits

LOAM treats odometry and mapping as separate estimators with no uncertainty propagation and no global consistency: errors that mapping cannot see (a drift in the unmapped direction, a degenerate corridor) pass into the result silently. It is also tied to features and to the sensor geometry. These are the weaknesses later systems address: LeGO-LOAM (ground and speed), LIO-SAM and FAST-LIO (inertial fusion and consistent estimation), and direct methods.

Common misconceptions

Engineering takeaways

1. LOAM splits LiDAR SLAM into odometry (scan-to-scan, 10 Hz) and mapping (scan-to-map, 1 Hz) and fuses their outputs.
2. Odometry matches edge points to lines and planar points to planes, with the pose interpolated across the sweep to handle distortion.
3. Mapping uses more features against the accumulated map and corrects the drift of the odometry.
4. The split makes real-time operation possible; the lack of an IMU and of loop closure limits robustness and long-term accuracy.
5. Diagnose a LOAM-like system by module: smooth path and blurred map point at mapping; jumpy path points at odometry or features.

References and further reading

Primary papers and authoritative documentation.

• Zhang, J. and Singh, S. (2014). LOAM: Lidar odometry and mapping in real-time. Robotics: Science and Systems.
• Zhang, J. and Singh, S. (2017). Low-drift and real-time lidar odometry and mapping. Autonomous Robots 41(2).
• Geiger, A., Lenz, P. and Urtasun, R. (2012). Are we ready for autonomous driving? The KITTI vision benchmark suite. IEEE CVPR: the odometry benchmark.
• Shan, T. and Englot, B. (2018). LeGO-LOAM. IEEE/RSJ IROS.

Knowledge check

Three conceptual questions. Write an answer, then reveal the explanation. Your answers stay in your browser.

What’s next

Next: A-LOAM, a compact open implementation of this pipeline, and how to read it module by module.

Related content

Nothing below is required to finish this lesson.

• Before
  M5-T4: Scan Registration

  Compare point-to-point, point-to-plane and NDT/GICP-style registration and their failure modes.

• Lab
  M5-L3: Understand the LOAM Pipeline

  Implement smoothness-based edge and plane selection and a two-rate odometry step on 2D scans.