PaperlayerView exact revision on arXiv

Source-linked AI summary

P$^2$Calib: Utilizing Pattern Priors for LiDAR-Camera Extrinsic Calibration

Xiangcheng Hu

arXiv:2609.07516v2cs.RO

TL;DR

Sparse angular coverage and mixed-pixel returns limit LiDAR hole-center extraction in four-hole calibration. P$^2$Calib fits with CAD-specified radius and layout priors, and on two real solid-state LiDAR datasets reduces joint residual and held-out reprojection error over the raster-circle baseline.

  • Problem

    Sparse LiDAR sampling and mixed-pixel returns make hole-center extraction a limiting accuracy factor in four-hole calibration.

    Source: A. Motivation and Challenges · pp. 1–2

  • Method

    P$^2$Calib fixes the hole radius to its CAD value and enforces the four-hole rectangular layout as a global rigidity constraint.

    Source: A. Motivation and Challenges · p. 2, §VIII · p. 11

  • Results

    On two real solid-state LiDAR datasets, P2Calib lowers the joint residual by 90% and 82% and the held-out reprojection error by 96% and 77% over the raster-circle baseline.

    Source: §VIII · p. 11

  • Takeaways & Limitations

    The two priors are integrated into an interactive tool that provides a complete extrinsic calibration pipeline.

    Source: §VIII · p. 11

    The current implementation is tailored to a four-hole calibration board with a rectangular layout; extension to other target-based problems remains open.

    Source: D. Limitations · p. 11

Abstract

from arXiv · show

Target-based LiDAR-camera extrinsic calibration is a prerequisite for multi-sensor fusion in robotics. However, in the widely adopted four-hole pipeline, \rt{calibration accuracy is limited by hole-center extraction in the LiDAR side, where sparse angular coverage and mixed-pixel returns displace the estimated centers}. This paper presents P$^2$Calib, which exploits \textit{pattern priors}, geometric constraints specified by the CAD model of the target board, to improve calibration accuracy. First, we incorporate the known hole radius as a fitting constraint to prevent center estimates from degrading under sparse angular coverage. Building on the improved hole estimates, we further enforce the rigid rectangular layout of the four holes as a global consistency constraint to correct residual errors across holes. Both priors are integrated into an \rt{interactive tool that runs the pipeline from target detection to the final extrinsic}. Experiments on simulated and real datasets show that P$^2$Calib reduces the joint registration residual by 90\% and 82\% and the held-out reprojection error by 96\% and 77\% over the baseline. \rt{We will release the code and data\codelink{} to facilitate future research.

I. INTRODUCTION

Sparse LiDAR sampling and mixed-pixel returns make hole-center extraction the accuracy bottleneck in four-hole calibration. P$^2$Calib uses the board’s known radius and rectangular layout as fitting constraints, while the related-work discussion situates this approach among target-based and other calibration methods.

II. RELATED WORK

  • Four-hole boards with ArUco markers enable four 3D–3D correspondences from one view, while prior uses of known geometry generally apply radius bounds or layout checks after unconstrained fitting.The paper positions its priors as constraints within estimation rather than post-fit checks or template alignment.Source: A. Motivation and Challenges · pp. 1–2

A. Motivation and Challenges

  • Sparse, uneven hole-boundary sampling makes LiDAR center extraction the calibration bottleneck; free-circle fitting is degenerate, and mixed-pixel returns bias the boundary inward along the same direction.The CAD-specified radius and rectangular arrangement are the two pattern priors used to address these challenges.Source: A. Motivation and Challenges · p. 2

B. Contributions

  • The radius prior fixes each fitting radius to its CAD value, while the layout prior enforces global rigidity across the four centers, reducing their degrees of freedom from twelve to four.Both priors are integrated into an interactive calibration tool and evaluated on simulated and real data.Source: B. Contributions · p. 2

C. Organization

  • The paper reviews targetless and motion-based calibration before examining how target-based methods use calibration-board geometry.This organizational overview frames the work’s position in target-based calibration.Source: §II · p. 2

A. Targetless and Motion-Based Methods

  • Targetless methods rely on scene signals, motion-based methods require trajectories exciting all rotational axes, and learning-based methods depend on training sensor configurations and scene statistics.Target-based calibration avoids reliance on environmental structure, motion, or appearance and remains standard in laboratory and production settings,.Source: A. Targetless and Motion-Based Methods · p. 2

B. Target-Based Calibration

The four-hole pipeline obtains camera-side centers from the board pose and LiDAR-side centers from the point cloud, then registers their correspondences. P$^2$Calib distinguishes its priors from earlier approaches by imposing radius and layout constraints during estimation.

B. Target-Based Calibration

  • Earlier four-hole methods use the known radius as a tolerance while leaving it free, and apply the rectangular layout after estimation, so neither step removes fitting degeneracy within the fit.The cited alternatives include combinatorial checks, optional rigid-template alignment, and independent center-normal-radius estimation,,,,.Source: B. Target-Based Calibration · p. 2
  • P$^2$Calib fixes the radius to its CAD value with a shared mixed-pixel bias, then applies the rectangular rigidity constraint to radius-constrained centers when its residual agrees with observations.The layout constraint is applied after the per-hole radius-constrained estimates.Source: B. Target-Based Calibration · p. 3

A. Four-Hole Calibration Pipeline

  • The camera branch recovers board pose from ArUco markers and obtains camera-frame hole centers from the CAD layout; the LiDAR branch extracts the board plane and estimates hole centers.The board has four circular holes with known nominal radius and rectangular layout.Source: A. Four-Hole Calibration Pipeline · p. 3
  • The two branches provide four 3D–3D correspondences, and a closed-form absolute-orientation solution estimates the extrinsic transform using SVD.The camera-side centers follow from the board pose and CAD layout without image circle fitting.Source: A. Four-Hole Calibration Pipeline · p. 3

B. Accuracy Analysis

The calibration pipeline’s error is governed by independently estimated LiDAR hole centers, because other stages either introduce no error or are well conditioned. The radius prior addresses sparse-arc fitting ambiguity, while the layout prior couples the four centers.

B. Accuracy Analysis

  • LiDAR hole-center estimates govern calibration accuracy: they are independently recovered from boundary points, so their uncorrelated errors enter registration as correspondence noise.The closed-form registration only propagates input error, while plane segmentation is well conditioned because noise averages across board points.Source: B. Accuracy Analysis · p. 3

IV. PATTERN-PRIOR EXTRINSIC CALIBRATION

  • The two pattern priors address fitting degeneracy and the lack of coupling among the four independently estimated holes.The radius prior constrains each hole, while the rigid CAD layout couples the four centers.Source: §IV · p. 3, Figure 3 · p. 4

A. Free-Circle Ambiguity

  • A short measured arc permits center shifts compensated by radius changes, leaving the center weakly determined even with accurate ranges.The residual differences can fall below observed range noise, and mixed-pixel returns are absorbed as apparent radius reduction; fixing the CAD radius removes this ambiguity without modeling cost.Source: A. Free-Circle Ambiguity · p. 4

B. Radius Prior

The radius prior fits each hole center against an effective radius that accounts for shared inward shrinkage from mixed pixels. Sector-based sampling and robust fitting support the per-hole updates.

B. Radius Prior

  • Mixed pixels shrink observed hole radii by a comparable amount, modeled as one shared bias δ, so each center is fit against effective radius r−δ.The shared bias is bounded by δmax = 30 mm.Source: B. Radius Prior · p. 4
  • The radius prior reduces each hole’s unknowns from three to two and adds one scalar for the whole board.It fits centers using the effective radius rather than estimating a free radius for every hole.Source: B. Radius Prior · p. 4
  • Each hole draws annulus candidates and keeps the innermost surviving point per azimuth sector, updating only with coverage across at least ten of 24 sectors and no empty run longer than six.The center fit uses a Huber loss with a 12 mm threshold, and the shared bias is refreshed from weighted observed radii.Source: B. Radius Prior · p. 4

C. Layout Prior

The layout prior couples the four radius-constrained hole centers by fitting them to the CAD rectangle. A conditional acceptance test guards against applying the rigid projection when the individual fits or their agreement are unreliable.

C. Layout Prior

  • The layout prior constrains four independently estimated centers to a rigid CAD rectangle, reducing eight center coordinates to three pose parameters.The fit also assigns detected centers to CAD corners.Source: C. Layout Prior · p. 4
  • For each corner assignment, the rigid placement is a planar Procrustes problem; the solver enumerates all 24 assignments.Rectangle relabeling symmetry resolves the mirror ambiguity implicitly.Source: C. Layout Prior · p. 4
  • A rigid fit can spread one gross center error across all four centers, motivating the acceptance test before snapping centers to the rectangle.The algorithm snaps all four centers only after the acceptance test passes.Source: C. Layout Prior · pp. 4–5
  • The pipeline applies the projection only when all holes pass the coverage test and projected centers satisfy the dmax = 35 mm acceptance condition.Otherwise, it retains the individual hole fits.Source: C. Layout Prior · p. 5

D. Estimation Model and Solver

The solver estimates the two pattern priors through a relaxed fit and conditional projection onto the rigid layout. The analysis explains how the priors reduce unresolved parameters and improve conditioning under sparse angular coverage.

D. Estimation Model and Solver

  • The solver minimizes a nine-parameter relaxation with four free centers and a shared bias, then conditionally projects the result onto the rigid rectangle.The relaxation avoids reselecting sector samples during joint minimization.Source: D. Estimation Model and Solver · p. 5

A. Degrees of Freedom

  • The priors reduce the four-hole model from 12 to 4 continuous parameters while keeping all boundary measurements, increasing the measurement-to-parameter ratio from 4S/12 to 4S/4.The radius prior replaces four radii with a known radius and shared bias; the layout prior reduces eight center coordinates to the rectangle pose.Source: A. Degrees of Freedom · p. 5

B. Conditioning under the Radius Prior

  • As angular span Φk narrows, the radius prior reduces the rate of conditioning loss from O(Φk^-4).Source: B. Conditioning under the Radius Prior · p. 5
  • A coverage test must constrain both sector count and the longest empty run because diametrically opposed samples can yield a singular center-fit matrix despite the span measure.The second-order circular resultant detects this directional-concentration failure mode.Source: B. Conditioning under the Radius Prior · p. 5

C. Error Projection under the Layout Prior

The layout prior projects noisy center estimates onto the rigid rectangle’s three-dimensional pose manifold. Under the stated isotropic-error model and correct assignment, this projection reduces expected center error, though the model is an idealization.

C. Error Projection under the Layout Prior

  • For a fixed corner assignment, rigid placements form a three-dimensional manifold in the eight center coordinates, and the layout prior projects estimates to its closest point.The closed-form solution realizes the orthogonal projection for the selected assignment.Source: C. Error Projection under the Layout… · p. 6
  • Under independent isotropic center errors and a correct assignment, projection onto the rigid layout retains 3/8 of expected squared center error, or about 0.61 in root-mean-square.The projection keeps the component in the rectangle’s three-dimensional tangent space.Source: C. Error Projection under the Layout… · p. 6
  • The 3/8 error-retention result is a limiting value because isotropy is idealized; the discarded five-dimensional deformation component is thresholded by the acceptance test.A component large enough to fail the test indicates unreliable hole estimation.Source: C. Error Projection under the Layout… · p. 6

D. Complementarity of the Two Priors

The radius and layout priors depend on one another: radius constraints bound errors before layout projection, while the rigid rectangle helps identify the shared bias. The interactive implementation reports intermediate diagnostics and supports several LiDAR types.

D. Complementarity of the Two Priors

  • The radius prior bounds degenerate center errors before layout projection, while the four distinct boundary orientations supplied by the rectangle make the shared bias identifiable.A single hole observed on a narrow arc can leave the bias indistinguishable from translation along that arc’s direction.Source: D. Complementarity of the Two Priors · p. 6

VI. SYSTEM IMPLEMENTATION

  • Figure 4 pairs camera images with range-shaded LiDAR scans for the simulator, Avia, Mid-360, and area-array FS sensor.The holes appear as white voids in the scans.Source: Figure 4 · p. 6
  • The interactive tool reports each pipeline stage, layout disagreement, and registration residual so practitioners can distinguish extraction from correspondence quality and exclude views before a multi-scene solve.The tool integrates both priors into the FAST-Calib codebase.Source: §VI · p. 6

VII. EXPERIMENTS

The experiments evaluate the two pattern priors on simulated boards and real data from multiple LiDAR types. The full pipeline improves hole-center estimates and held-out reprojection error over FAST-Calib, with the priors contributing in complementary regimes.

A. Experimental Setup

  • Simulation uses a CAD-matched four-hole board with range noise, quantization, and mixed-pixel returns across six standoffs, five placements, and two sampling densities.The protocol uses 60 frames; each frame is rendered as a single frame (SF) and an accumulated cloud (AC).Source: A. Experimental Setup · pp. 6–7
  • Real evaluation includes 18 FS-B and 20 FS-C scenes, plus five Avia and seven Mid-360 scenes; FAST-Calib is the primary baseline, with velo2cam also compared.The FS datasets use an area-array solid-state LiDAR; the additional FAST-Calib datasets cover non-repetitive-scan and hybrid solid-state scanners.Source: A. Experimental Setup · p. 7
  • Simulation scores hole-center and extrinsic errors against ground truth, while real-data joint residual and LOO error use detected ellipse centers.The LOO reprojection error fits the extrinsic on remaining frames and compares projected held-out centers with their true pixel positions in simulation.Source: A. Experimental Setup · p. 7

1) Overall results analysis:

  • Against FAST-Calib, leave-one-out (LOO) reprojection error falls from 2.61 to 0.40 px on SF frames and from 1.51 to 0.23 px on AC frames.On the 40 SF frames both methods detect, the ordering is unchanged at 0.36 against 2.61 px.Source: 1) Overall results analysis · pp. 7–8
  • Hole-center error drops from 7.0–14.7 mm to 2.0–2.1 mm on SF and from 6.8–7.4 mm to 1.6–1.9 mm on AC frames.Table I reports this metric by standoff group.Source: 1) Overall results analysis · p. 7, Table I · p. 8
  • The radius prior alone achieves 91–96% of the full improvement; the layout prior lowers far-SF error from 5.1 to 3.9 mm.In the real-scene ablation, the radius-only centers form a skewed quadrilateral, while both priors produce the rectangular layout.Source: Figure 5 · p. 7, 1) Overall results analysis · p. 8

2) Coverage, standoff and noise analysis:

The full pipeline stays more accurate as rim coverage and standoff worsen, while stage-wise tests show how the radius and layout priors contribute and where the radius prior has a limit.

2) Coverage, standoff and noise analysis:

  • As rim coverage falls from 100% to 10%, full-pipeline error rises from 1.7 to 3.2 mm, versus 4.3 to 15.0 mm for the baseline.The baseline is non-monotonic; the full pipeline remains below 4.5 mm as coverage drops.Source: 2) Coverage, standoff and noise analysis · p. 8
  • With one hole at 25% coverage, layout projection cuts its radius-only error from 6.2 to 2.9 mm, at a 0.7 mm cost on fully covered holes.With full rims, the radius prior alone matches the full pipeline; as coverage falls, the layout prior contributes an increasing share of improvement.Source: 2) Coverage, standoff and noise analysis · p. 8
  • At 5.0 m, the full pipeline reaches 5.0 mm while the SF baseline recovers no board at 4.0 or 5.0 m, extending usable range by more than 1.5×.The full pipeline remains within 1.8–2.8 mm through 4.0 m.Source: 2) Coverage, standoff and noise analysis · p. 8
  • Across lateral placements, SF full-pipeline error rises from 2.4 to 3.7 mm while baseline error stays between 9.1 and 12.0 mm; under range noise, the full pipeline remains below baseline on both densities.On SF frames, it still recovers the board at σ = 50 mm, while the baseline fails from σ = 35 mm.Source: 2) Coverage, standoff and noise analysis · p. 8

3) Stage-wise error analysis:

  • Sector sampling reduces error relative to raster initialization by 63% and 90%, and layout projection removes a further 43 to 52%.The fixed radius does not increase error at full coverage; the measured squared-error ratio after and before projection is 0.25 and 0.31, rising to 0.37 and 0.40 under 25% and 50% coverage.Source: 3) Stage-wise error analysis · p. 9
  • On one-sided arcs of 10–12 sectors, the free fit is more accurate than the radius prior, at 7.1 against 9.9 mm, because the shared δ cannot represent the radial bias.Layout projection recovers the advantage in each such case.Source: 3) Stage-wise error analysis · p. 9

C. Real-World Experiments

On the common detected scenes, the full pipeline substantially lowers joint residual and held-out reprojection error on both FS datasets. The results are conditional on successful extraction, which trades off against detection rate.

1) Area-array LiDAR calibration:

  • On FS-B, the full pipeline reduces joint residual from 217.7 to 22.3 mm and LOO error from 67.5 to 2.82 px; on FS-C, it reduces joint residual from 191.5 to 35.1 mm and LOO error from 43.3 to 9.91 px.These are reductions of 90% and 96% on FS-B and 82% and 77% on FS-C; paired errors use 15 FS-B and 10 FS-C common scenes.Source: Table III · p. 9, 1) Area-array LiDAR calibration · p. 9
  • On five FS-C scenes at 1 m, every variant including the baseline exceeds 13 px LOO error; excluding them, the full pipeline reaches 4.67 px on the five scenes at 2–4 m.The authors identify the consistent ≈15 mm image-plane offset as a range-dependent sensor inconsistency.Source: 1) Area-array LiDAR calibration · p. 9
  • On FS-B, FAST-Calib, RP-only, LP-only, and the full pipeline detect four centers in 85, 90, 80, and 80 of 90 attempts; on FS-C, they do so in 65, 85, 60, and 60 of 100 attempts.The full pipeline and LP-only both have fewer detected scenes than the baseline on these datasets.Source: 1) Area-array LiDAR calibration · p. 9
  • Detection rate is traded for extraction quality: the layout screening rejects five FS-C scenes at the 50 mm layout check, and three others have fewer than four candidates.The paired results therefore support improved calibration consistency conditional on successful extraction.Source: 1) Area-array LiDAR calibration · p. 10

2) Cross-sensor generality:

The priors also improve several scanning-LiDAR results, with smaller gains than on the FS sensor and no uniform improvement in held-out error. The paper reports modest feature-extraction overhead and identifies the current board layout as an implementation boundary.

2) Cross-sensor generality:

  • The full pipeline lowers joint residual from 13.2 to 12.4 mm on Avia, 6.1 to 5.3 mm on Mid-360 A, and 5.4 to 4.8 mm on Mid-360 B.LOO error decreases from 4.99 to 3.83 px on Avia and 1.54 to 1.15 px on Mid-360 A, but increases from 2.03 to 2.41 px on Mid-360 B.Source: 2) Cross-sensor generality · p. 10

3) Runtime:

  • Feature extraction takes 16.4 and 18.6 ms on FS-B and FS-C, a 38% and 51% increase over baseline totals of 11.9 and 12.3 ms.Layout projection accounts for less than 0.5 ms and radius refinement for 4.3 to 5.0 ms; on Avia and Mid-360, plane extraction dominates and the overhead is negligible.Source: Table IV · p. 10, 3) Runtime · pp. 10–11

2) Cross-sensor generality:

  • The priors' improvements are smaller on scanning LiDARs than on the FS sensor, and lower joint residual does not guarantee lower held-out error in every case.The authors relate the smaller marginal contribution to denser rim coverage on scanning LiDARs.Source: 2) Cross-sensor generality · p. 10

D. Limitations

  • The current implementation is tailored to four-hole boards with a rectangular layout; whether the construction helps other target-based estimation problems remains open.The paper notes that the radius prior applies to known-dimension circular features and the layout prior to known rigid patterns.Source: D. Limitations · p. 11

VIII. CONCLUSION

  • On two real solid-state LiDAR datasets, P2Calib lowers joint residual by 90% and 82% and held-out reprojection error by 96% and 77% over the raster-circle baseline.The conclusion identifies multi-circle and non-rectangular board layouts as future extensions.Source: §VIII · p. 11

Related papers

Loading 2609.07516v2…