FAST-Calib: LiDAR–Camera Calibration

• FAST-Calib is a target-based calibration pipeline that estimates the rigid-body transformation between LiDAR and camera frames using a custom 3D target and ArUco markers.
• It employs pattern-agnostic LiDAR edge extraction and ellipse-based hole recovery to robustly construct cross-modal 3D correspondences across multiple scenes.
• The method achieves point-to-point residuals below 6.5 mm in under 0.7 s on a CPU, supporting both mechanical and solid-state LiDARs for fast deployment.

Searching arXiv for FAST-Calib and closely related calibration work to ground the article in current papers. FAST-Calib is a target-based method for LiDAR–camera extrinsic calibration that estimates the rigid-body transformation $T_{CL}\in SE(3)$ mapping points from the LiDAR frame into the camera frame. It is presented as a fast and user-friendly tool built around a custom-made 3D target, a pattern-agnostic LiDAR edge extraction procedure, ellipse-based compensation for LiDAR spot spread, and joint optimization across multiple scenes. The reported system supports both mechanical and solid-state LiDARs, operates in under $0.7$ s on CPU, and achieves point-to-point registration residuals consistently below $6.5$ mm on the tested platforms (Zheng et al., 23 Jul 2025).

1. Problem definition and estimation objective

FAST-Calib addresses LiDAR–camera extrinsic calibration by estimating the rigid-body transform $T_{CL}$ that maps a LiDAR point $p_L$ into the camera frame as

$p_C = R_{CL}\,p_L + t_{CL},$

where $R_{CL}\in SO(3)$ and $t_{CL}\in \mathbb{R}^3$. Given two sets of 3D correspondences $\{p_i^C\}_{i=1}^M$ from the camera and $\{p_i^L\}_{i=1}^M$ from the LiDAR, the extrinsic parameters minimize the least-squares objective

$0.7$0

For a single capture, FAST-Calib produces four exact correspondences; for $0.7$1 captures, the correspondence count is $0.7$2. The multi-scene registration problem is written as

$0.7$3

This formulation places FAST-Calib in the class of 3D–3D rigid registration methods. Its distinctive contribution lies not in altering the registration criterion itself, but in constructing cross-modal correspondences quickly and in a way that is stated to be agnostic to LiDAR scan patterns (Zheng et al., 23 Jul 2025).

2. Custom 3D target and cross-modal correspondence construction

The physical target is a flat board with four circular holes of known radius $0.7$4 and known center locations in the board frame $0.7$5, together with four ArUco markers at the board corners. The camera-side processing begins by detecting the four ArUco markers; with camera intrinsics $0.7$6 and lens distortion, a PnP solver recovers the board-to-camera pose $0.7$7. The known offsets of the hole centers in the board frame then yield their 3D coordinates in the camera frame according to

$0.7$8

On the LiDAR side, the observed structure consists of the board’s planar surface and the circular-hole edges. After plane fitting, FAST-Calib extracts the hole-edge contours and recovers their centers, producing the corresponding points $0.7$9.

The target design is central to the method’s geometry. Four hole centers define a compact set of correspondences per capture, while the ArUco markers provide a direct image-based route to the board pose. A plausible implication is that FAST-Calib reduces the calibration problem to repeated recovery of a small, fixed correspondence set rather than dependence on scene texture, environmental semantics, or LiDAR-specific scan topology (Zheng et al., 23 Jul 2025).

3. Pattern-agnostic LiDAR processing and ellipse-based hole recovery

The LiDAR pipeline begins by merging $6.5$0 sequential scans into a dense point cloud. For very sparse mechanical LiDARs, a small “wobble” motion is used to aggregate additional beams. FAST-Calib then applies axis-aligned pass-through filters in $6.5$1, $6.5$2, and $6.5$3 to isolate the board region, fits the largest plane with RANSAC using a distance threshold of $6.5$4 m, and retains the inliers. Voxel-grid downsampling at $6.5$5 mm resolution reduces the point count, after which the dominant plane is rotated to coincide with $6.5$6 via its normal (Zheng et al., 23 Jul 2025).

Edge extraction is performed in the aligned plane. FAST-Calib builds a 2D KD-tree of the aligned points and, for each point $6.5$7, finds all neighbors $6.5$8 within radius $6.5$9 m. For each neighbor, it computes the polar angle

$T_{CL}$0

After sorting the angles in ascending order, the method forms circular angular gaps

$T_{CL}$1

and

$T_{CL}$2

If $T_{CL}$3 with $T_{CL}$4, the point is classified as an edge point because it has a “wedge” in empty space. The paper states that this test depends only on local neighbor geometry and therefore works for any LiDAR scan pattern.

The extracted edge points are clustered by Euclidean distance, with an example threshold of $T_{CL}$5 mm, yielding up to four clusters per capture. Each cluster $T_{CL}$6 is then fitted with a general conic

$T_{CL}$7

via the direct least-squares method of Fitzgibbon et al. (1996), under the ellipse constraint $T_{CL}$8. From the fitted parameters $T_{CL}$9, the center is recovered as

$p_L$0

Ellipse validity is checked by requiring $p_L$1 cm and eccentricity $p_L$2. Valid clusters yield hole centers $p_L$3 in the aligned plane, which are then transformed back into the LiDAR’s original frame. This ellipse-fitting stage is described as compensating for edge dilation artifacts caused by LiDAR spot spread (Zheng et al., 23 Jul 2025).

4. Camera processing, registration, and multi-scene optimization

The camera-side processing is deliberately minimal. FAST-Calib detects ArUco markers, uses known marker size $p_L$4 together with camera intrinsics to solve PnP for $p_L$5, and computes the hole-center positions in the camera frame by applying $p_L$6 to the known offsets in the board frame $p_L$7. The resulting set $p_L$8 is paired with the LiDAR-derived $p_L$9 (Zheng et al., 23 Jul 2025).

For registration, FAST-Calib uses a closed-form Kabsch–SVD solution to obtain $p_C = R_{CL}\,p_L + t_{CL},$0 and $p_C = R_{CL}\,p_L + t_{CL},$1 immediately. The method can optionally refine all parameters across all scenes with a Gauss–Newton iteration on the same least-squares cost. The paper emphasizes that the entire pipeline remains fast because edge detection and ellipse fitting operate in 2D on a downsampled planar cloud and because the multi-scene steps run in parallel.

The operational workflow is correspondingly simple: the target is placed arbitrarily in front of the sensors, $p_C = R_{CL}\,p_L + t_{CL},$2–$p_C = R_{CL}\,p_L + t_{CL},$3 viewpoints are captured, and FAST-Calib is launched. No manual point-cloud cropping or parameter tuning is required. This suggests that the method is engineered less as a research-only prototype than as a deployable calibration routine for robotics systems that already have calibrated camera intrinsics and can present a target to both sensors (Zheng et al., 23 Jul 2025).

5. Empirical performance and reported operating envelope

FAST-Calib was tested on three LiDAR models—Ouster OS1-128, Livox Avia, and Livox Mid360—each paired with a wide-angle camera. The reported evaluation uses five runs per sensor: four runs based on $p_C = R_{CL}\,p_L + t_{CL},$4 subsets of captures and one run using all four captures. The point-to-point registration residuals in the registration objective are stated to be always below $p_C = R_{CL}\,p_L + t_{CL},$5 cm ($p_C = R_{CL}\,p_L + t_{CL},$6 mm) (Zheng et al., 23 Jul 2025).

Against Velo2Cam, FAST-Calib is reported to outperform on solid-state LiDARs: Velo2Cam fails on Mid360 and yields $p_C = R_{CL}\,p_L + t_{CL},$7 cm errors on Avia. For the Ouster, the two methods are described as on-par, with FAST-Calib at $p_C = R_{CL}\,p_L + t_{CL},$8–$p_C = R_{CL}\,p_L + t_{CL},$9 mm and Velo2Cam at $R_{CL}\in SO(3)$0–$R_{CL}\in SO(3)$1 mm.

The CPU-only runtime measurements on an Intel i7-10700K are as follows:

Sensor	LiDAR processing	Total
Avia	0.633 s	0.66 s
Mid360	0.675 s	0.69 s
Ouster	0.647 s	0.68 s

The same evaluation reports camera processing times of $R_{CL}\in SO(3)$2 s for Avia, $R_{CL}\in SO(3)$3 s for Mid360, and $R_{CL}\in SO(3)$4 s for Ouster, while registration costs approximately $R_{CL}\in SO(3)$5s per capture and is described as negligible.

The implementation guidance is specific. The recommended voxel size is $R_{CL}\in SO(3)$6 mm for general use, the target should be kept $R_{CL}\in SO(3)$7–$R_{CL}\in SO(3)$8 m away for best edge resolution, and ambient lighting should permit clear ArUco detection. For sparse scanners, a gentle handheld pitch oscillation is recommended to ensure sufficient edge coverage. These operating conditions delimit the method’s practical envelope: the reported accuracy and timing are tied to a target-based acquisition procedure in which the board is visible to both modalities and LiDAR observations contain enough edge structure for valid ellipse fitting (Zheng et al., 23 Jul 2025).

6. Practical scope, assumptions, and common misconceptions

FAST-Calib is explicitly a target-based automatic calibration pipeline. Its automation concerns correspondence extraction, registration, and runtime behavior after the target has been placed and the captures collected; it does not eliminate the need for a custom 3D target or for controlled views of that target (Zheng et al., 23 Jul 2025).

A second common misconception is that support for different LiDAR types must rely on scan-pattern-specific logic. The paper states the opposite: FAST-Calib supports both multi-line mechanical and non-repetitive solid-state LiDARs without any change in parameters because its edge extraction relies purely on local neighbor geometry rather than scan “rings.” This is a central design claim, and the Avia and Mid360 experiments are used to substantiate it.

A third point concerns what the method returns. FAST-Calib estimates the LiDAR–camera extrinsic $R_{CL}\in SO(3)$9 through 3D point correspondences derived from the target geometry; it is not posed as an online self-calibration system that continually updates extrinsics during operation, nor as a targetless scene-structure method. The fact that the pipeline completes in under $t_{CL}\in \mathbb{R}^3$0 s does not change this methodological scope. A plausible implication is that FAST-Calib is most naturally suited to fast setup, re-calibration, and validation cycles rather than continuous background adaptation (Zheng et al., 23 Jul 2025).

7. Related calibration methods and terminological ambiguity

FAST-Calib occupies a distinct position within a broader calibration literature. FastCal, for example, is an online sensor-extrinsic self-calibration method that uses information-theoretic segment selection, observability-aware updates through a rank-revealing decomposition of the Fisher information matrix, and time-decay of segments. It is framed as loosely coupled self-calibration from independent sensor odometry streams and reports constant-time updates with average runtime $t_{CL}\in \mathbb{R}^3$1 ms per key-frame on a single core (Nobre et al., 2019). That problem setting differs from FAST-Calib’s target-based LiDAR–camera correspondence construction.

LiMo-Calib addresses a different extrinsic problem again: on-site calibration of the LiDAR–motor transformation for a quadruped robot-based panoramic 3D sensing system. It estimates only roll $t_{CL}\in \mathbb{R}^3$2, pitch $t_{CL}\in \mathbb{R}^3$3, and horizontal translations $t_{CL}\in \mathbb{R}^3$4, uses point-to-plane residuals on planar features, applies normal homogenization for feature selection, and solves the resulting optimization in Ceres with a Huber loss (Li et al., 18 Feb 2025). In contrast, FAST-Calib calibrates a LiDAR with respect to a camera through a board target with circular holes and ArUco markers.

The name “FAST-Calib” is also ambiguous outside robotics geometry. “Fast Calibrated Explanations” introduces FAST-Calib, also referred to in that paper as FCE, as a local, post-hoc explanation method for machine learning models that produces calibrated predictions and feature-wise importance weights with uncertainty intervals, using Venn-Abers for classification and Conformal Predictive Systems for regression (Löfström et al., 2024). That usage is unrelated to LiDAR–camera extrinsic estimation.

Taken together, these papers indicate that FAST-Calib, in the strict robotics sense of (Zheng et al., 23 Jul 2025), denotes a one-second, target-based LiDAR–camera extrinsic calibration pipeline built on pattern-agnostic edge extraction, ellipse-based hole-center recovery, and multi-scene SVD registration, rather than a targetless calibration method, an online self-calibrator, or an explainability framework.