Active Structured Light (ASL)

• Active Structured Light (ASL) is a 3D sensing method that projects known illumination patterns onto a scene to recover depth through calibrated triangulation.
• ASL encompasses discrete codification like binary/Gray-code systems and continuous phase-based methods, each balancing speed, resolution, and robustness.
• Recent advances integrate neural reconstruction and adaptive calibration to overcome challenges from motion, ambient noise, and specular surfaces.

Searching arXiv for the cited structured-light papers to ground the article in current indexed records. Searching arXiv for the cited structured-light papers to ground the article in current indexed records. Active Structured Light (ASL) is a class of 3D sensing methods in which a known, controllable illumination pattern is projected onto a scene and observed by one or more cameras, so that projector–camera correspondences can be inferred from the pattern deformation and converted to 3D geometry by triangulation (Gu et al., 2014). In contrast to passive stereo, ASL injects its own texture, and in contrast to time-of-flight systems it estimates shape from spatial deformation rather than light travel time (Bajestani et al., 20 Dec 2025). The field spans temporally coded binary and Gray-code systems, continuous phase-based profilometry, one-shot color codification, event-based and single-photon variants, multi-projector and active-stereo configurations, and recent neural formulations that integrate structured illumination with differentiable reconstruction or pose estimation (Lam et al., 2022).

1. Definition, scope, and historical positioning

ASL is organized around a simple operational principle: a projector emits a designed pattern, one or more cameras capture the illuminated scene, decoding recovers either per-pixel projector coordinates or an equivalent phase/code, and calibrated geometry triangulates 3D points (Wan et al., 24 Jan 2025). In the classical projector–camera view, the projector is treated as an inverse camera, and the correspondence problem is transformed from passive image matching into decoding a known optical signal (Lam et al., 2022). In camera–camera variants such as active stereo, the projected pattern serves as artificial texture that improves stereo correspondence without explicitly decoding projector pixels (Baek et al., 2020).

A persistent distinction inside ASL is between discrete codification and continuous codification. Binary or Gray-code systems assign projector pixels unique temporal identifiers, giving unambiguous correspondences at the cost of multiple projected frames (Gu et al., 2014). Phase-based methods project sinusoids and recover a continuous phase, which yields sub-stripe interpolation and higher spatial resolution per projection, but classically requires multiple phase-shifted exposures and a subsequent unwrapping stage (Je et al., 2015). One-shot and event-based systems restructure this trade-off by embedding phase shifts in color, by using sparse high-speed binary patterns, or by relying on asynchronous sensing rather than conventional frame integration (Je et al., 2015).

The literature also broadens the meaning of ASL beyond fixed-rig depth recovery. In robotic construction sensing, ASL is used as a calibrated, ROS-based sensor producing dense point clouds with phase-shifting profilometry (Lam et al., 2022). In event-based RGB-D, ASL is reframed as asynchronous depth and color annotation of individual events under known projection schedules (Bajestani et al., 20 Dec 2025). In neural reconstruction, ASL becomes a differentiable constraint on geometry, appearance, and even pose, rather than only a decoding front-end (Li et al., 2022).

2. Geometric and radiometric foundations

The geometric core of ASL is triangulation between a camera ray and a projector ray or plane. With camera intrinsics $K_c$, projector intrinsics $K_p$, and calibrated extrinsics, a camera observation $(u_c,v_c)$ and a decoded projector coordinate $(u_p,v_p)$ define a 3D point either algebraically or by ray intersection (Lam et al., 2022). In rectified formulations this reduces to the stereo relation

$Z = \frac{fB}{d},$

where $f$ is focal length, $B$ is baseline, and $d$ is disparity (Bajestani et al., 20 Dec 2025). In camera-pair systems such as 3DUNDERWORLD-SLS, the projector may be used only to assign correspondences, after which triangulation is performed between two camera rays rather than camera–projector rays (Gu et al., 2014).

For sinusoidal ASL, the projected irradiance is parameterized directly in phase. A standard three-step RGB formulation uses

$$C_1(x,y)=a(x,y)+b(x,y)\cos\big(\phi(x,y)-2\pi/3\big),$$

$C_2(x,y)=a(x,y)+b(x,y)\cos\big(\phi(x,y)\big),$

$K_p$0

with $K_p$1, where $K_p$2 is the DC term, $K_p$3 is the modulation amplitude, and $K_p$4 is the projector phase (Je et al., 2015). Under linear response and negligible channel correlation, the wrapped phase is recovered by

$K_p$5

or, in the general $K_p$6-step case,

$K_p$7

(Je et al., 2015, Lam et al., 2022).

Radiometric nonidealities are central rather than peripheral in ASL. Real projector–camera pairs exhibit crosstalk, channel correlation, nonlinearity, ambient contamination, defocus, and reflectance-dependent distortions (Je et al., 2015). One compact model is

$K_p$8

where $K_p$9 is the projected RGB vector, $(u_c,v_c)$0 models spectral mixing, $(u_c,v_c)$1 is per-channel nonlinear response, and $(u_c,v_c)$2 is sensor noise (Je et al., 2015). In binary and Gray-coded systems, gamma correction is less central, but per-pixel thresholds, inverted patterns, and shadow masks are used to robustly separate lit and unlit states (Gu et al., 2014). In single-photon ASL, the measurement model changes fundamentally: each exposure is binary, and the probability of a detection is governed by Poisson photon arrivals, ambient flux, and dark counts rather than analog intensity readout (Sundar et al., 2022).

3. Pattern codification strategies

A central taxonomy in ASL concerns how the projected signal encodes projector coordinates. Temporal binary and Gray codes remain canonical. For a projector of resolution $(u_c,v_c)$3, the number of patterns per axis is

$(u_c,v_c)$4

so a $(u_c,v_c)$5 projector requires $(u_c,v_c)$6 column codes and $(u_c,v_c)$7 row codes; 3DUNDERWORLD-SLS projects both axes, their inverted counterparts, and two solid-color frames, for a total of $(u_c,v_c)$8 images in that case (Gu et al., 2014). Gray coding is preferred because consecutive codes differ by one bit, which widens the thinnest stripes and reduces sensitivity to blur and color bleeding (Gu et al., 2014).

Phase-shifting profilometry (PSP in the profilometry sense of phase shifting, not phase sampling) uses continuous sinusoidal fringes and recovers phase with sub-pixel precision. In the SL Sensor, the primary codification is a “3+3” pattern: three high-frequency sinusoidal fringes with phase offsets of $(u_c,v_c)$9, followed by three unit-frequency fringes for temporal unwrapping (Lam et al., 2022). The unwrapped phase is computed as

$(u_p,v_p)$0

$(u_p,v_p)$1

and then mapped to a projector coordinate through the fringe wavelength (Lam et al., 2022). Compared with binary codes, this improves spatial resolution and correspondence precision but reduces motion tolerance and ambient-light robustness (Lam et al., 2022).

Single-shot codification reduces temporal burden by moving information into spatial, chromatic, or asynchronous dimensions. One-shot RGB sinusoidal projection encodes three phase shifts in a single color frame and recovers phase from one video image, making it suitable for moving or deforming objects (Je et al., 2015). Phase Sampling Profilometry projects a sampled phase pattern and reconstructs the 1D phase signal in the spatial-frequency domain under a Nyquist condition, thereby obtaining an unambiguous phase coordinate from one image without temporal phase shifts or phase unwrapping (Wang, 2020). Single-shot feature-decoding systems instead treat the known IR pattern as an input to a learned matcher that builds feature-space cost volumes between the pattern and the captured IR image, replacing hand-crafted pixel-domain correlation (Li et al., 16 Dec 2025).

Other coding strategies target failure modes rather than only speed. Redundancy codes append error-correcting or error-detecting bits to conventional structured-light sequences. In the ambient-light regime, error correction codes such as $(u_p,v_p)$2 or $(u_p,v_p)$3 increase temporal codeword separability and can reduce disparity error rates by approximately $(u_p,v_p)$4 to $(u_p,v_p)$5 in the reported experiments (Sun et al., 2022). In global-illumination settings, CRC-based error detection supports iterative adaptive reconstruction with $(u_p,v_p)$6 frames per iteration and convergence in $(u_p,v_p)$7 rounds on the reported scenes (Sun et al., 2022). In single-photon ASL, BCH coding on the most significant bits and binary shift coding on the least significant bits are combined to remain robust both to photon-noise-induced bit flips and to short-range blur and resolution mismatch (Sundar et al., 2022).

4. Dynamic-scene and high-speed ASL

A common misconception is that ASL is inherently restricted to static scenes. The literature directly contradicts this, but the mechanisms differ across system classes. One-shot RGB sinusoidal ASL acquires all three phase steps in a single projected color frame and reconstructs moving or deforming objects such as human faces from one video image (Je et al., 2015). Structured-light flow goes further by treating motion blur as signal rather than nuisance: two projected line patterns from two projectors produce two blur widths whose ratio yields depth along each camera ray, so fast-moving objects can be reconstructed from a single blurred exposure (Furukawa et al., 2017).

Multi-shot phase systems can also be adapted to motion if the motion model is constrained. The SL Sensor assumes linear translation without rotation and perpendicular to fringe direction, so the encoded projector coordinate does not change during the sequence; per-image phase-correlation alignment then compensates motion before standard PSP decoding (Lam et al., 2022). In a controlled scan of a mask moved $(u_p,v_p)$8 mm per frame at $(u_p,v_p)$9 cm distance, this strategy removes the reported $Z = \frac{fB}{d},$0–$Z = \frac{fB}{d},$1 mm motion ripples, and in robot scanning it supports motion at approximately $Z = \frac{fB}{d},$2 m/s over a $Z = \frac{fB}{d},$3–$Z = \frac{fB}{d},$4 m stand-off (Lam et al., 2022).

Event-based ASL moves high-speed sensing into the asynchronous domain. ESL uses a raster-scanned laser point and an event camera, constructs camera and projector time maps over a single $Z = \frac{fB}{d},$5 ms scan, and estimates depth by maximizing patchwise spatio-temporal consistency rather than matching individual events (Muglikar et al., 2021). The key consistency condition is

$Z = \frac{fB}{d},$6

with the cost

$Z = \frac{fB}{d},$7

which is minimized over depth or disparity (Muglikar et al., 2021). Reported experiments show an average RMSE reduction of $Z = \frac{fB}{d},$8 relative to the cited event-based baseline for the same acquisition time (Muglikar et al., 2021).

More recent event-based ASL uses DLP projection rather than raster scanning. E-RGB-D projects binary line or dot patterns at approximately $Z = \frac{fB}{d},$9 kHz, tags each event with the active pattern and color state, and computes depth per event by direct disparity lookup under rectified camera–projector geometry (Bajestani et al., 20 Dec 2025). Reported performance includes color detection speed equivalent to approximately $f$0 fps and per-pixel depth detection up to approximately $f$1 kHz (Bajestani et al., 20 Dec 2025). At the hardware level, swept-plane ASL with an acousto-optic scanner has pushed plane generation to approximately $f$2 planes per second and full-frame depth to $f$3 fps, with ROI scanning around $f$4 kHz in the reported system (Sirikonda et al., 2024). This suggests that, in current high-speed ASL, the limiting bandwidth may have shifted from the light source to the sensor.

5. Calibration, reconstruction, and optimization paradigms

Classical ASL relies on explicit calibration and explicit decoding. Camera and projector are modeled as pinhole devices, distortion is calibrated, and decoding produces projector coordinates or phase values that are inserted into geometric constraints (Lam et al., 2022). In projector–camera systems this often means solving the two projection equations algebraically; in multi-camera Gray-code systems it can mean triangulating between camera rays that decoded to the same projector code (Gu et al., 2014). Calibration procedures range from checkerboard-based stereo calibration to projector-as-camera calibration using projected patterns on planar targets (Lam et al., 2022).

Recent work has expanded calibration and reconstruction into optimization problems. “Towards Unified Structured Light Optimization” introduces a single-image global matching method that uses a De Bruijn grid with embedded markers to establish dense projector–camera correspondences from one projected image, then applies photometric compensation through PCNet and DeArtifact modules to optimize projector input across binary, speckle, and color-coded SL families (Wan et al., 24 Jan 2025). The reported ablation on photometric fidelity shows SSIM/PSNR/RMSE improving from $f$5 without DeArtifact to $f$6 with “Projection Net + real data” (Wan et al., 24 Jan 2025).

Another line of work integrates ASL with differentiable neural reconstruction. In “Multi-View Neural Surface Reconstruction with Structured Light,” Gray-code correspondences act as geometric constraints inside a neural implicit surface optimization that jointly refines SDF, appearance, and camera poses (Li et al., 2022). The structured-light reprojection and triangulation losses

$f$7

$f$8

supplement the photometric rendering term and reduce calibration burden because camera poses are optimized rather than assumed fixed (Li et al., 2022).

Neural-SDF formulations also generalize ASL beyond explicit decoding. ActiveNeuS models each projector as a virtual camera inside volumetric rendering, with color at a sample point given by

$f$9

and uses photometric supervision rather than projector-coordinate decoding to recover shape in low light and underwater scenes (Ichimaru et al., 2024). Neural Active Structure-from-Motion similarly optimizes a Neural-SDF and camera trajectory directly from sparse projected patterns in textureless dark environments (Sirikonda et al., 2024). A plausible implication is that the notion of “decoding” in ASL is broadening from explicit symbol recovery to any calibrated inference process that constrains geometry through known illumination.

6. Robustness, failure modes, and emerging directions

Across the literature, several failure modes recur. Ambient illumination reduces structured-light contrast and can dominate binary observations, as made explicit by the single-photon detection model

$B$0

with ambient-driven false positives becoming severe under strong illumination (Sundar et al., 2022). Glossy or specular surfaces saturate cameras, disturb disparity or phase recovery, and create missing or erroneous depth in both traditional SL sensors and object-centric ASL systems (Yang et al., 2023). Global illumination, inter-reflections, subsurface scattering, and multi-device interference violate the direct-only assumption underlying many decoders and motivate redundancy codes, adaptive reconstruction, or active viewpoint selection (Sun et al., 2022).

Motion is another recurrent source of ambiguity, but its treatment varies. Multi-frame Gray or phase-shifting systems remain vulnerable unless one-shot or motion-compensated designs are used (Je et al., 2015). Very high fringe frequencies improve resolution but increase the density of $B$1 wraps and therefore the burden on unwrapping (Je et al., 2015). Extremely high scan rates in event-based or swept-plane systems shift the bottleneck to sensor bandwidth, causing dropped events and reduced completeness unless sweeps are accumulated or ROI scanning is used [(Sirikonda et al., 2024); (Sirikonda et al., 2024)?]. Since the available evidence reports this bottleneck explicitly for the million-planes-per-second system, a more conservative statement is that bandwidth limits remain central even when scanning hardware becomes much faster (Sirikonda et al., 2024).

Several representative directions now define the frontier of ASL. One direction is practical, open, high-accuracy sensing with commodity components and reproducible software, exemplified by SL Sensor and 3DUNDERWORLD-SLS [(Lam et al., 2022); (Gu et al., 2014)]. A second direction is robust single-shot and asynchronous sensing, including RGB one-shot phase analysis, event-based RGB-D, neural single-shot feature decoding, and single-photon structured light (Je et al., 2015, Bajestani et al., 20 Dec 2025, Li et al., 16 Dec 2025, Sundar et al., 2022). A third direction is active decision-making: viewpoint selection for shiny objects based on predicted Fisher information, adaptive projection to control bandwidth or coverage percentage, and adaptive ROI scanning that allocates projected structure only where needed (Yang et al., 2023, Bajestani et al., 20 Dec 2025, Sirikonda et al., 2024). A fourth direction is joint optimization of illumination and inference, including learned diffractive patterns in active stereo and unified photometric compensation across SL pattern families (Baek et al., 2020, Wan et al., 24 Jan 2025).

The cumulative picture is that ASL is no longer a single family of projector–camera triangulation systems, but a broader research domain in which illumination coding, sensing modality, calibration strategy, and reconstruction algorithm are co-designed. The published work supports three stable conclusions: ASL remains highly effective on textureless surfaces because it controls the image formation process (Gu et al., 2014); its dominant trade-offs are still among resolution, robustness, and temporal budget (Lam et al., 2022); and current research increasingly treats pattern design, motion handling, and reconstruction as a unified optimization problem rather than isolated stages (Wan et al., 24 Jan 2025).