1 of 1

HeadTrace: hardware-free head tracking

Tanmay Kulkarni, CREATE Community Day 2026

Overview

  • Existing eye tracking solutions are often extremely expensive for people with disabilities (such as ALS or Duchenne muscular dystrophy)
  • HeadTrace is an alternative webcam-based, external-hardware-free head tracking solution for people with mobility limitations
  • Allows a user to operate a computer hands-free, using nose direction to control a cursor

Existing technologies

  • MSR Eye Tracking framework: Framework and drivers that makes it straightforward to integrate eye trackers into Windows - both through the Windows Eye Control and a set of standard APIs.
  • Google MediaPipe: Framework for creating ML pipelines for tasks like face landmarking and segmentation (through a graph structure carrying out different subtasks).

Versions, testing, and prototyping

  • The first version utilized a very overcomplicated projection algorithm on a "nose angle vector" derived from MediaPipe and camera matrices (without calibration)
    • Through CWU’s SETC, one of their AAC (eye tracker) users and contributors tested it – however, his range of head motion was asymmetrical leading to poor performance
  • Made me realize I needed calibration, so I tried an interpolation algorithm on the nose angle, which ran into poor quality data and projection math that didn’t help:
  • I returned to this problem many months later, purely interpolating the plain nose coordinate (not any calculated angle) – yielding significantly better results

HeadTrace’s pipeline (see demo!)

Calibration:

Nose point per target (MediaPipe)

Target sequence

OLS + interpolation to solve nose-point to screen target formula

Metadata (.json)!

Live tracking:

Nose point

Metadata parameters calculate screen target

OneEuro smoothed point

Screen point injected (Eye Tracking Stack)!

Possible next steps

  • Accuracy: Exploring alternate polynomials, different point selection, and changing the smoothing parameters to get maximum accuracy and require less repeated calibration
  • Speed: running the FaceMesh every frame is expensive, not a huge latency at the moment but could be improved
  • User experience: More user testing is absolutely necessary, and exploring new UI capabilities as well as forms for HeadTrace’s format (executable, app, website, etc.)

Acknowledgements

I would like to thank Central Washington University's SETC and their users and support with distributing the first version of HeadTrace; Ann Paradiso, Dwayne Lamb, and Chris O'Dowd for their endless support through my accessibility tech journey; Dr. Mar Gonzalez-Franco for her input; and of course, Harish Kulkarni for his extensive guidance and getting me started on all of this!