Skip to content

ADR 0001 — Planar-Neighbourhood ICP for Boresight Estimation

Status: Accepted
Date: 2026-04-26
Deciders: Daud Tasleem

Context

Airborne and UAV LiDAR surveys are collected in multiple overlapping flight strips. Small errors in the boresight angles (misalignment between the IMU and the LiDAR scanner head) and lever-arm offsets produce visible seams in the merged point cloud: adjacent strips are offset by 5–15 cm vertically and show step-like artefacts on flat surfaces and building rooftops.

No maintained open-source Python library exists to correct this automatically. The only relevant OSS project (cdfbdex/StripAdjustment, 4 ★, last commit July 2022) uses Coherent Point Drift (CPD) — a method designed for non-rigid registration that is overkill and slow for this problem, and ships without a public API, tests, or CI.

Commercial tools (BayesStripAlign, LP360, LiDAR360) solve this well but are proprietary and cost thousands of dollars per seat.

Decision

Use planar-neighbourhood point-to-plane ICP minimised by Levenberg–Marquardt over a 6-DOF rigid body model (3 boresight angles + 3 lever-arm components).

Algorithm sketch

  1. Feature extraction. For every point, compute PCA on its k-NN neighbourhood. Retain patches whose planarity (Weinmann eigenvalue ratio) ≥ threshold. Apply normal-space sampling to ensure angular diversity.

  2. Correspondence matching. Build a KD-tree on reference patch centroids. For each target patch, accept the nearest reference patch if the centroid distance ≤ max_dist AND the inter-normal angle ≤ max_angle.

  3. Objective. Minimise the sum of squared point-to-plane distances: F(p) = Σ_k [ n_k · (R(ω,φ,κ) q_k + t − c_k) ]² where n_k, c_k are the reference normal and centroid for patch k.

  4. Solver. scipy.optimize.least_squares with method="lm" (Levenberg– Marquardt), ftol/xtol/gtol = 1e-9.

  5. Apply and report. Transform the target strip by the estimated correction. Report per-strip RMSE before/after, number of correspondences, convergence.

Rejected Alternatives

A. Coherent Point Drift (CPD)

CPD (Myronenko & Song 2010) models registration as a probability density estimation problem and supports non-rigid deformations. For boresight calibration, the deformation is strictly rigid (small angle + translation); CPD's extra flexibility buys nothing and adds 3–10× computational cost. It also requires choosing a regularisation hyper-parameter that is unintuitive for practitioners. Rejected: unnecessarily general, slow.

B. Vanilla ICP (point-to-point)

Point-to-point ICP converges to a local minimum on planar surfaces because it does not exploit surface normals. On a flat rooftop, it cannot distinguish vertical from horizontal errors. Rejected: incorrect convergence behaviour on the dominant geometry type in airborne surveys.

C. DEM-based strip adjustment (raster differencing)

Bin points to a raster DEM, subtract, and fit a correction surface. Simple and fast, but loses sub-pixel precision and cannot handle dense vegetation or steep slopes where a DEM surface is ill-defined. (See MDPI Sensors 21 2782, 2021 for a UAV-DEM approach.) Rejected: restricted to surveys with clear ground returns.

D. RANSAC-based global registration (e.g. Open3D FPFH + RANSAC)

Feature-based global registration (FPFH descriptors + RANSAC) is designed for scenes with no prior pose estimate. Boresight errors are small (< 0.1°), so global registration is unnecessary and introduces a factor-of-100 overhead. Rejected: solves the wrong problem.

Consequences

  • Correct convergence on flat-terrain surveys (standard airborne and UAV mapping).
  • Performance: ~5 s per strip pair at 200k points (see benchmark).
  • Limitation: requires some flat or planar surface patches in the overlap zone. Surveys over pure dense forest with no ground penetration may yield insufficient planar features. Documented in README Limitations section.
  • Extension path: add IMU trajectory refinement in v0.2 (see ROADMAP).