Predicting Where the Laser Ring Will Fall: Keyframe-Guided Extraction of Faint Omnidirectional Structured Light

A ring-laser profiler measures the shape of a pipe or tunnel in one camera exposure. It casts a flat 360° sheet of laser light, which paints a bright ring on the surrounding wall. A wide-angle camera rigidly mounted at a known offset sees the ring, and simple triangulation turns each ring pixel into a metric 3D point. This works in pitch-dark, textureless concrete where ordinary camera-based mapping fails. Its main weakness is range. The laser's energy is spread around the whole circle, so the ring gets dimmer as walls get farther away, and beyond some distance it disappears into sensor noise. Today's systems search the entire image for the ring in every frame. We propose to stop doing that. Consecutive cross-sections of a tunnel are nearly identical, so the last reliably measured cross-section (a "keyframe"), moved forward by the rig's estimated motion and projected through the calibrated fisheye model, predicts a thin curved band where the ring must appear. Searching only inside that band, and adding up brightness along the ring's predicted direction, should detect rings far too faint to find by searching the whole image.

The underlying idea comes from 2D-guided search pruning in 3D vision. A cheap, reliable signal decides where an expensive or noisy search is allowed to look. We turn this around in time and modality. The guide is the system's own past 3D measurement, and the thing being pruned is the 2D search for a weak active signal. We expect the claim to be measurable as a longer working range, a higher fraction of usable points at a given distance, and lower downstream registration error. A simple detection-theory argument predicts how large the gain should be.

Core Claim

In a single-shot omnidirectional ring-laser profiler, using a predicted search band in the raw fisheye image, combined with intensity integration along that band, raises the farthest wall distance at which the ring can be extracted reliably. The prediction comes from a keyframe cross-section, propagated by motion. The gain should follow a predictable law, it should come from both ingredients together rather than either one alone, and it should carry through to lower mapping error against a total station.

The predicted law comes from a standard argument for detection in Gaussian noise. Suppose a ring candidate must be declared at one of $N$ image locations with false-alarm probability $P_{fa}$. A union bound then requires the integrated signal-to-noise ratio to exceed roughly $\tau = \sqrt{2\ln(N/P_{fa})}$. If the detector also sums $k$ pixels along the ring's known local direction, each with per-pixel signal-to-noise $s$, the integrated value is about $\sqrt{k}\,s$. The ring's irradiance falls roughly as $1/r$ with wall distance $r$, because the sheet's energy spreads over a circle of circumference proportional to $r$. Under that falloff, the maximum usable range scales as

$$ r_{\max} \propto \frac{\sqrt{k}}{\sqrt{2\ln(N/P_{fa})}} . $$

Pruning shrinks $N$ from a whole fisheye image (about $10^6$ pixels) to a band (about $10^4$). That gives a modest logarithmic gain. Along-band integration gives the larger $\sqrt{k}$ gain, but it is only safe when the predicted ring shape is accurate over $k$ pixels. That is exactly what the keyframe prior supplies, and what a blind search cannot assume. So the claim is specifically about the combination, and the ablations below are designed to separate the two factors.

Where This Sits in Prior Work

The sensing side builds directly on omnidirectional ring-laser profiling. Yamashita and colleagues introduced it for pipe inspection by pairing a radial laser projector with a hyperboloidal catadioptric camera (Paper). Igaue et al. extended it to mobile tunnel mapping in Line Structured Light-Based Tunnel 3D Measurement with 2D-3D Point Cloud Matching on Projected Light of Ring Laser. Their system uses the ring itself as an artificial landmark and aligns a rear observation camera's rays to the forward unit's cross-sections with Rigid Coherent Point Drift, a probabilistic point-set registration solved by expectation-maximization. It achieved an 18.36 mm mean error (8.27 mm standard deviation) over 48 m of real tunnel. This paper both motivates and constrains our work. It names the dimming of the ring with distance as the main limitation. It also gives us the downstream pipeline we must improve on end to end. It does not use any prior from earlier cross-sections when extracting the ring. Scale Optimization of Structure from Motion for Structured Light-based All-round 3D Measurement by Kawata et al. likewise uses structured-light cross-sections, there to fix metric scale for structure-from-motion. It shows how valuable each extracted cross-section is, but it does not address weak-signal extraction.

The pruning side supplies the method. Frustum PointNets for 3D Object Detection From RGB-D Data restricts 3D search to the viewing volume of a 2D detection. Any3DIS (Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking) and Go-SLAM (Grounded Object Segmentation and Localization with Gaussian Splatting SLAM) restrict work to selected keyframes. Any3DIS is roughly 10× faster than Open3DIS with higher accuracy, which is evidence that well-chosen keyframes make the search both cheaper and better. PVALane prunes dense 3D anchors using 2D objectness. All of these assume a pinhole camera, and all use pruning to save compute on a dense signal. We use it to gain sensitivity on a sparse, faint signal, in non-pinhole optics.

The closest work in spirit is Omnidirectional Multi-Object Tracking. Its FlexiTrack component feeds the previous frame's object locations to a panoramic detector as priors. That is temporal pruning in 360° imagery, but it is purely 2D and passive, with no metric geometry propagating the prior. JRMOT offers a related lesson: when active depth becomes sparse at range, beyond 15–20 m in their case, other cues must carry the estimate. Our "other cue" is the geometric continuity of the structure itself.

Research Program

We will build a ring-laser profiler on the Igaue design: a 360° laser-sheet projector, a forward fisheye measurement camera with a narrowband filter, and a pose source. The pose source is wheel or inertial odometry, plus the system's own cross-section registration. Calibration uses a generic central omnidirectional camera model on the unit sphere. The predicted band is therefore computed directly in raw fisheye pixels, avoiding the interpolation blur and seam artifacts that come from unwrapping the image into a cylinder.

The method has three parts. The first is keyframe management. A cross-section becomes a keyframe when its extraction confidence is high everywhere around the ring, for instance on close walls or after a full-image search. Between keyframes, the active prior is the most recent keyframe, moved forward by odometry and corrected by registration. The second is band prediction. The prior 3D cross-section is intersected with the current laser plane and projected into the image. Its width at each angle is set from the propagated pose covariance, so uncertain motion produces wider bands. The third is along-band detection. At each angle around the ring, we search a short segment perpendicular to the band and integrate intensity over an arc of $k$ pixels along the predicted tangent. The number $k$ adapts to how far the band's observed residual curvature departs from the prediction.

Pruning creates a known failure risk. Frustum PointNets already showed that if the guide misses something, the downstream stage cannot recover it. Here, that happens when the geometry changes abruptly, for example at a niche, a branch, or a person walking through the light sheet. Our safeguard is evidence-triggered widening. Wherever integrated evidence inside the band collapses while the image shows energy just outside it, the band is widened locally and step by step, and if necessary we fall back to full search over that arc. The recovered arc is then marked as new structure and can trigger a new keyframe. Arcs that jump radially in a way inconsistent with the rigid cross-section are flagged as dynamic and withheld from registration. This also addresses the static-scene assumption that Igaue et al. leave open.

Baselines

The naive reference point is Igaue-style extraction as published in spirit. For each radial ray of the fisheye image, the system finds the subpixel intensity peak and keeps it if it exceeds a global threshold, with no prior and no integration. Its output feeds an unchanged Rigid Coherent Point Drift registration. This is the "do nothing" condition: the current practice, re-run on our hardware.

The strongest existing competitor is the full Igaue et al. pipeline with every reasonable tuning added: adaptive local thresholds, the Coherent Point Drift outlier weight tuned on a held-out run, and multi-frame temporal averaging set to the largest window that does not blur at the rig's speed. Its reported number is 18.36 mm mean error against a total station over 48 m (Igaue et al.). That figure comes from a different tunnel, so we treat it as a sanity check for our reimplementation, not as the number to beat. The fair comparison is this tuned pipeline on our own captures, with the same camera, exposure, laser power, and filter, and the same per-frame compute budget. Every method's thresholds will be set to equal false-point rates on a calibration run. For range and point-yield-versus-distance metrics on ring lasers, there is no reported result found on Lacuna, so our baseline numbers for those will be measured rather than quoted.

Three ablations isolate the key ingredient. The first, band-only, uses the predicted band without along-band integration. It is effectively a Kalman-style tracking window and tests the $\ln N$ term alone. The second, integration-only, applies along-curve integration over the whole image against a family of hypothesized ring shapes. It resembles a Hough search, tests the $\sqrt{k}$ term without the prior, and should pay for its many hypotheses with false alarms. The third uses the immediately previous frame as the prior instead of a confidence-selected keyframe, to test whether keyframe selection matters or whether chaining frame to frame lets errors build up. Beating the tuned baseline means more than an average improvement. We require at least a 1.5× increase in the wall distance at which 90% of ring angles are recovered at equal false-point rate, with no loss of extraction accuracy on near walls, and a lower end-to-end error against the total station.

Why not just…?

Why not just raise laser power or exposure time? Power is capped by eye-safety limits for a sheet that sweeps occupied workspaces. Longer exposure smears the ring along the direction of travel on a moving rig, which directly worsens triangulation accuracy. Both raise $s$ by a constant factor, and our gain multiplies whatever $s$ the hardware provides. We will show this by running all methods at two power/exposure settings. If our method's range advantage persists at both, the two approaches combine rather than compete.

Why not just average several frames? Temporal averaging is the most serious threat, because it also gives a $\sqrt{\text{frames}}$ gain. It does so only if the ring stays on the same pixels, which a moving rig violates unless it moves very slowly. Our band prediction is, in effect, motion-compensated integration. If plain averaging matches us at realistic survey speeds, the claim fails. That is why tuned averaging is built into the strongest baseline, and why we report range as a function of rig speed.

Why not just rely on Coherent Point Drift's outlier model to absorb noisy points? Its uniform outlier component down-weights false points, but it cannot create points that were never extracted. At long range the problem is missing ring, not contaminated ring. The comparison against the tuned baseline, whose outlier weight is optimized, will show whether better extraction lowers final error beyond what robust registration already achieves.

Why not just train a network to segment the laser line? A learned segmenter could raise per-pixel sensitivity. However, ToF-360 shows that models trained on conventional imagery degrade badly on true single-shot panoramic data (29.56% mIoU), and no labeled ring-laser fisheye dataset exists in the file. More importantly, a segmenter's per-pixel output can itself be integrated along our predicted band. We will therefore test the band prior on top of a lightweight segmenter trained on our near-range captures, using high-SNR keyframes as automatic labels, to check that the two are complementary.

Proposed Experiments

No public ring-laser benchmark appears in the research file, so evaluation combines simulation with our own captures. In physically based simulation, we render the rig in pipes and tunnels while sweeping wall distance, albedo, rig speed, and sensor noise. This gives exact ground truth for plotting point yield against distance, and for checking whether measured range follows the predicted $\sqrt{k}/\sqrt{\ln(N/P_{fa})}$ scaling as $k$ and band width are varied. In a controlled corridor, we survey walls set at known stepped distances with a total station to measure maximum range and extraction error under real optics. In a long field run in a tunnel or utility culvert, also surveyed by total station, we compare end-to-end mapping error, drift over distance, and behavior at niches and with people crossing the sheet.

Three checks are decisive. The method must beat the tuned averaging baseline in range at realistic survey speed. Both ablations must fall clearly short of the full method, showing the gain needs prior and integration together. The evidence-triggered widening must recover abrupt geometry changes at a rate close to full search, so pruning does not silently delete real structure. If the first check fails, the idea reduces to motion-compensated averaging and should be reported as such.