Sparse contour depth scan of an anonymous three-quarter profile presenter facing an opposite front-mounted camera in a conference room

3D ToF Depth Sensing for AI Auto-Framing and Face Tracking

Add metric depth to RGB and AI pipelines for stable subject selection, digital crop or PTZ control.

Send Your Tracking Requirements Compare Camera Modules

When RGB AI needs a sense of distance

RGB can detect a face. Depth helps the system decide which person belongs in the frame when distance, occlusion and background context change.

ToF adds geometry. It does not replace RGB detail or the customer tracking model.

Spatial decision

RGB + Z
01 / Semantic signal RGB

Candidate regions

02 / Metric constraint ToF

Distance and rejection gate

03 / Framing action Host

Target retained in the loop

Digital crop PTZ control

RGB input

Semantic detection can treat a displayed person as a candidate target when the scene has no spatial constraint.

Candidate on display

ToF depth gate

A calibrated distance check can reject the flat display plane and keep the foreground subject in the tracking loop.

Depth gate / target retained

Illustrative comparison. The final filter depends on calibrated depth, scene geometry and customer target-management logic.

Separate people by distance

Metric Z distance adds a spatial constraint when people stand at different depths.

Reduce target switching

Depth can help reject a face on a display or a background subject outside the expected range.

Recover a 3D subject

A body centre and scene geometry give the host more than a flat bounding box.

Keep framing logic grounded

The controller can use depth as an input for crop, deadband and PTZ decisions.

Depth is one layer in the tracking loop

A module produces sensing data. Your system still owns detection, target management, control policy and the final camera behavior.

System pipeline architecture

The boundary is explicit: DOMI supplies sensing data, the host owns target logic, and the final camera product owns execution.

DOMI hardware layer

RGB sensor

Image stream for face or body features

DOMI 3D ToF module

VCSEL, depth, IR and point-cloud data

RGB + depth data

Customer host processing

2D face or body AI

Detection and semantic candidate features

3D association and filtering

Z-distance, occlusion and target rejection

Framing controller

Track state, smoothing, deadband and control policy

Track + control

Final execution layer

Digital crop

ePTZ framing window and output composition

Mechanical PTZ

Pan, tilt, zoom and servo response

Responsibility boundary

DOMI owns

RGB-D or depth data, IR, point cloud, drivers, SDK and module-level integration support.

Customer owns

Face or body detection, subject choice, active-speaker logic, tracking state, framing rules and PTZ control.

Product team owns

RGB image quality, gimbal, thermal design, optical window, EMC, eye safety and final product claims.

Before selecting a module, confirm whether your project needs a reference algorithm, coordinate output, timestamp support or PTZ demo.

Choose the route that fits the room

ToF is not automatically the best answer. Compare the full camera architecture against distance, FOV, host path and control needs.

Auto-framing architecture route comparison
Route Best fit Value Main limit
RGB-only AI Single presenter, controlled light and cost-sensitive designs. Low hardware cost, strong facial detail and mature models. No absolute distance. Background and low-light handling stay mostly in software.
RGB plus ToF Multi-person rooms, complex backgrounds and PTZ systems. Semantic RGB features plus metric depth in one perception loop. Power, thermal design, registration and synchronization need validation.
Separate ToF plus main RGB OEMs that already have a high-quality 4K RGB camera. Video and depth sensors can be optimized independently. Requires extrinsic calibration, timestamp alignment and coordinate conversion.
Panoramic camera plus PTZ Classrooms, stages and spaces that need global reacquisition. A wide view remains available while the PTZ camera zooms in. Two camera views and the control mapping must be calibrated together.

Room coverage is geometry, not a guarantee

Use this first-pass calculation to check whether a horizontal FOV can cover the room. Edge depth quality still needs measurement in the intended setup.

Coverage width = 2 × distance × tan(horizontal FOV ÷ 2)

Browse RGB-D cameras Browse ToF modules
Ideal horizontal coverage width
Product H FOV 1 m 2 m 3 m
DM-RGBD5002A 64° 1.25 m 2.50 m 3.75 m
DM5005A 70° 1.40 m 2.80 m 4.20 m
DMOM2808D 71.8° 1.45 m 2.90 m 4.34 m

Ideal geometry only. It does not mean the edge region has the same depth quality as the centre.

DM-RGBD5002A

64° horizontal FOV. Ideal geometry reaches 3.75 m width at 3 m distance.

DMOM2808D

71.8° horizontal FOV. Wider coverage supports compact room layouts.

Valid range limit

Use the published 4.5 m or 5 m range as a starting boundary, then validate edge quality.

Top-down geometry preview. The cones show ideal horizontal coverage, not a guarantee of uniform depth quality or tracking performance.

Turn performance language into an evaluation protocol

Do not infer system tracking performance from a sensor headline. Record the setup, measure the loop and keep the failure boundary visible.

Sensor-level checks

  • Depth resolution and actual output frame rate
  • Range under target reflectivity and ambient-light conditions
  • FOV, distortion, invalid pixels, confidence and depth noise
  • Power, peak current, temperature rise and thermal drift

Auto-framing system checks

  • End-to-end latency at P50 and P95
  • Framing centre error and track loss rate
  • ID switch rate and re-acquisition time
  • PTZ settling time, overshoot and motion smoothness
  • Host CPU, NPU, memory and interface bandwidth

Recommended scene set

Every result should name the module, firmware, algorithm version, distance, light, sample size and host platform.

  • Single presenter, two people crossing and group entry
  • Short occlusion, turning away and subject re-entry
  • Displayed face, poster, window backlight and low light
  • White clothing, black clothing and low-reflectivity targets
  • Digital crop and mechanical PTZ control loops

Evaluation Protocol

Until complete data exists, publish the test method instead of a tracking accuracy claim. ToF can support filtering and reacquisition, but it cannot see through an occlusion or guarantee a target choice.

Three hardware paths to prototype

Choose the module around the system boundary you need to validate first: combined RGB-D, compact MIPI depth or USB depth-only.

DOMI DM-RGBD5002A RGB-D ToF camera module

RGB-D prototype and algorithm validation

DM-RGBD5002A

Fast evaluation when RGB and depth should arrive as one calibrated stream.

Representative specs
640 × 480 depth at 25 fps1280 × 960 RGB at 30 fps0.2-5.0 m specified range64° H × 51° V depth FOV
Required host and interface
USB Type-C, USB 2.0, SDK host

Validate first

  • RGB-depth synchronization and timestamp behavior
  • Registration API and output combinations
  • Ambient light and target reflectivity in the final room
View Module Details Send Your Tracking Requirements
DOMI DMOM2808D compact MIPI ToF camera module

Embedded MIPI production path

DMOM2808D

Space-constrained designs that need a compact depth head and direct embedded integration.

Representative specs
320 × 240 at 10-30 fps0.25-4.5 m specified range71.8° H × 56.5° V FOV21 × 9.5 × 6.33 mm module head
Required host and interface
2-lane MIPI CSI-2, SPI, I2C and sync signals

Validate first

  • Host driver, power sequencing and lane timing
  • External RGB calibration and timestamp alignment
  • Temperature, optical window and eye-safety documents
View Module Details Send Your Tracking Requirements
DOMI DM5005A USB UVC 3D ToF camera platform

USB and UVC depth companion sensor

DM5005A

Rapid depth-only prototyping when the product already has its own RGB camera.

Representative specs
320 × 240 depth at up to 30 fps940 nm: 0.2-2 m850 nm: 0.3-5 mPoint cloud, depth and IR output
Required host and interface
USB 2.0, Micro USB, UVC 1.2

Validate first

  • Latency and SDK output-mode combinations
  • Enclosure, heat path and optical-window transmission
  • Ambient conditions for the selected wavelength and range
View Module Details Send Your Tracking Requirements

These three paths cover the main evaluation decisions. Other DOMI modules may fit a different range, output or form factor and should be confirmed against the complete project brief.

Move from sample output to a production-ready camera

The fastest prototype is not always the lowest-risk product path. Make data, timing, optics, host constraints and compliance explicit before design freeze.

Data and timing

  • Depth, IR, RGB, point cloud, confidence and invalid pixels
  • Hardware or software synchronization and timestamp source
  • Frame skew, detector interval and end-to-end latency

Optics and mechanics

  • Intrinsics, extrinsics, distortion and RGB-depth registration
  • Optical window, stray light, mounting tolerance and FOV clearance
  • VCSEL duty cycle, peak current, temperature and heat path

Host and compliance

  • Windows, Linux, Android, Jetson or target SoC support
  • USB bandwidth, UVC mode combinations, MIPI lanes and drivers
  • Multi-camera interference, laser safety, EMC and ESD

OEM project path

  1. Requirements review

    Confirm room geometry, people count, tracking mode, host and interface.

  2. Module shortlist

    Map range, FOV, optical form, bandwidth and RGB requirements to candidate hardware.

  3. Evaluation kit and SDK

    Confirm samples, drivers, output modes, timestamps and any reference examples.

  4. Scene validation

    Test people, occlusion, background displays, light and the complete control loop.

  5. Optical and mechanical integration

    Lock the window, mounting, thermal path, calibration and host wiring.

  6. EVT and DVT

    Verify firmware, reliability, EMC, eye safety and system-level acceptance criteria.

  7. Production readiness

    Align calibration, MOQ, lead time, PCN, EOL and change-control expectations.

Optics and mechanics

Keep the optical paths physically separated

The cover glass and isolation structure are part of the depth design. Separate the transmitter and receiver openings, then re-check ranging after the final window, adhesive and enclosure are fixed.

Cover glass
Transmission, thickness, coating and internal reflections
Isolation foam
Two separate openings for the transmit and receive paths
Sensor board
Tx VCSEL and Rx lens placement, heat path and calibration
Exploded-view concept for mechanical review. Final spacing, window design, eye-safety documentation and EMC validation depend on the selected module and enclosure.
Review SDK support Explore other solutions

Questions to resolve before a camera test

Direct answers to the depth, algorithm, environment and interface questions that shape an evaluation brief.

Send Your Tracking Requirements

Technical ownership: DOMI FAE team. Content prepared September 2026. Confirm controlled specifications and deliverables for the selected project.

Does ToF replace RGB face detection?

No. RGB normally provides facial and body features for AI models. ToF adds metric depth so the system can separate people by distance, reject unsuitable background regions and estimate a subject position in 3D.

Does the module output face coordinates directly?

Standard hardware primarily outputs RGB, depth, IR or point-cloud data, depending on the model. Face detection, subject tracking and PTZ commands normally run on the host. Confirm any reference algorithm or coordinate API for the selected module and firmware.

How can depth reduce false tracking of a face on a display?

Depth can help identify that a displayed face sits on a flat background plane instead of at the expected subject distance. Use it as an additional constraint, not as the only anti-false-target mechanism.

Can the system follow a selected presenter or active speaker?

Depth can provide the 3D position of candidate subjects, but selected-presenter and active-speaker modes require target-management logic. Active-speaker tracking may also need microphone-array or audio-direction data.

Is 25 to 30 fps sufficient for smooth PTZ tracking?

It can be sufficient for normal indoor movement, but frame rate alone does not determine smoothness. Test end-to-end latency, detector interval, filtering, PTZ response, deadband, acceleration and zoom settling as one control loop.

How is a separate ToF module aligned with a 4K RGB camera?

The system needs RGB and depth intrinsics, distortion parameters, camera-to-camera extrinsics and timestamp alignment. Verify calibration after the final mount, optical window and lens configuration are fixed.

What environmental conditions can affect ToF performance?

Strong ambient light, low-reflectivity clothing, reflective or transparent surfaces, multipath, optical-window losses, temperature and interference from other active depth cameras can affect valid range and depth quality.

Should we choose USB RGB-D, USB depth-only or MIPI?

Use USB RGB-D for the fastest combined evaluation. Use USB depth-only when the product already has a separate RGB camera. Use MIPI when size, embedded integration and production architecture matter more than plug-and-play development.

Get a Sensor Shortlist and Integration Review

Share your room geometry, tracking mode and host platform. DOMI can help evaluate working distance, field of view, hardware path and integration risks.

  • Recommended RGB-D or depth-only hardware path
  • First-pass working distance and FOV review
  • USB, UVC or MIPI integration risk notes
  • Evaluation sample, SDK and controlled datasheet availability
  • Priority tracking and environment scenes to validate

Send your engineering brief

Work email, company, application and a short project brief are enough to start. Add dimensions, host details and failure cases if available.