ToF Depth Calibration: Beyond the Datasheet

The hidden complexities of calibrating Time-of-Flight depth sensors for accuracy across temperature, distance, and reflectivity.

Evyatar Bluzer
3 min read

On paper, Time-of-Flight is the simplest way to measure depth: emit light, time the return, compute distance. Almost none of that simplicity survives contact with a real sensor.

Sources of ToF Error

The errors come in three classes, and they need different treatment. Systematic errors are predictable and correctable: phase non-linearity in demodulation, integration time effects, temperature-dependent delays, lens distortion. Scene-dependent errors are harder to correct - multi-path interference from light bouncing off multiple surfaces, mixed pixels at depth edges (the two ToF weaknesses I flagged a year ago), reflectivity-dependent bias, motion artifacts. And underneath both sits the noise floor: photon shot noise, sensor read noise, quantization.

The Calibration Challenge

The job is to characterize and correct the systematic errors while understanding the limits the scene-dependent and random ones impose.

Distance Calibration

We put a flat target at known distances across the 0.3-5m range we froze - 0.5m, 1m, 2m, 3m, 4m, 5m - and at each one measure the reported depth across the FOV, compare against ground truth, and fit a correction polynomial or lookup table. Simple enough, except the correction also depends on integration time, modulation frequency, and temperature, so a full characterization is a 4D calibration space.

Temperature Calibration

ToF sensors drift with temperature, the same lesson the camera calibration taught: electronic delays change and optical components shift. Our approach is to calibrate at multiple temperatures (15°C, 25°C, 35°C, 45°C), fit a thermal model for each calibration parameter, monitor device temperature in operation, and apply a temperature-compensated correction.

Reflectivity Compensation

Low-reflectivity surfaces like dark materials return less light, which means higher noise and a potential bias, since fewer photons shift the systematic errors. We're building a calibration target set with controlled reflectivity patches - 5%, 20%, 50%, 90% - to characterize that relationship.

Why is multi-path the hardest problem?

When light bounces off multiple surfaces before returning to the sensor, the measured phase is a weighted average of the paths, and neither distance is correct. We're exploring multi-frequency ToF, where different wavelengths separate direct from indirect returns; spatial coding with structured illumination patterns; and learning-based correction trained on simulated data with known multi-path. None of them is a complete fix. Multi-path may just be a limitation we end up living with.

ToF multi-path interferenceA sensor box on the left and a wall on the right; a solid direct arrow runs between them and a highlighted dashed path bounces off a second surface above before reaching the same wall spot; a dashed marker beyond the wall marks the reported depth. ToF sensor wall second surface direct path: the true distance indirect path, bounced off a second surface: longer, so it arrives later reported depth: a weighted average of both paths
Multi-path at a single pixel: the direct return and a return bounced off a second surface arrive at the same pixel, and the sensor reports a distance that is the weighted average of both paths, matching neither surface.

Calibration Data Management

Each device carries hundreds of calibration parameters, which turns into a data problem in its own right: a unique device ID mapped to its calibration record, versioned calibration algorithms, and traceability from a field issue back to the factory data. Building that infrastructure now will save enormous pain later.

Comments