cd ~/
IR IlluminationDepth SensingOpticsHardware DesignMagic Leap

IR Illumination Design for Depth Sensing

The physics and engineering of infrared illumination systems for ToF and structured light depth cameras.

Evyatar Bluzer
3 min read

Depth sensors need light to see. Passive cameras get theirs from the environment; active depth sensors emit their own, typically in the near-infrared (NIR) spectrum around 850nm or 940nm.

Why NIR?

Several properties line up. 850nm and above sits outside the human visual range, so users never see the illumination. VCSELs and LEDs are mature sources at these wavelengths. Silicon sensors still respond well, though sensitivity drops above 900nm. And narrow bandpass filters can reject most of the solar spectrum.

Illumination Architectures

Flood Illumination (ToF)

Cover the entire field of view with uniform IR light. Conceptually simple, but uniformity is a fight - edges of the FOV receive less light thanks to cosine falloff plus lens vignetting. Power density runs into eye safety limits on peak intensity as you push for range. And much of the light falls on areas we don't need, like sky and distant objects.

Patterned Illumination (Structured Light)

Project a known pattern - dots, lines, speckle - and use its deformation for triangulation. Diffractive optical elements (DOE) split a single beam into the pattern and are very power efficient. The pattern itself must be unique enough for unambiguous matching, and density is a trade-off: more points means more resolution but harder matching.

Spot Illumination (Scanning)

A single beam scanned across the scene, LiDAR-style. All the energy concentrates in the measurement point, so it's power efficient, and eye safety is easier to manage with a low duty cycle per point. The catch is speed: mechanical scanning limits frame rate.

Design Constraints

For our headset: eye safety under IEC 62471 Class 1, which limits power density at the eye. An illumination power budget under 200mW. Range of 0.3m to 5m indoor and 0.3m to 3m outdoor. Ambient rejection good enough to work in 10,000 lux sunlight. The range and sunlight requirements come from March's comparison of ToF and structured light, where the whole depth subsystem got 500mW.

The math is harsh. Sunlight at 940nm contributes ~0.2 mW/cm²/nm. Our narrow filter (10nm bandwidth) still passes 2 mW/cm². We need to exceed this with our illumination at 5m range while staying within power and safety budgets.

Sunlight versus the illuminator at the depth sensorTwo columns flowing down into one box: sunlight passes a 10 nm bandpass filter into the depth sensor, and the IR illuminator lights the scene whose return reaches the same sensor; the return arrow is highlighted and labeled with the 2 mW per square centimetre of residual sunlight it must exceed. Sunlight, 10,000 lux~0.2 mW/cm²/nm at 940 nmNarrow bandpass filter10 nm wideDepth sensorstill sees 2 mW/cm² of sunIR illuminator, VCSEL or LED at 940 nmunder 200 mW, IEC 62471 Class 1 at the eyeScene0.3-5 m indoor, 0.3-3 m outdoor full solar spectrum 2 mW/cm² passes flood, pattern, or scanned spot return must exceed2 mW/cm² at 5 m
Both light sources at the depth sensor: after the 10 nm filter, sunlight still lands 2 mW/cm² on it, and the illuminator must beat that at 5 m with under 200 mW and Class 1 eye safety. Figures from the design constraints above.

Multi-Frequency Approaches

One promising direction: modulate illumination at frequencies where ambient light doesn't contribute. Time-domain or frequency-domain filtering can then reject DC (sunlight) while preserving our modulated signal.

Next month I'll get into the optics design.

Comments