A Gaussian Aperture Camera for Person Detection Without Face Recording
Public spaces often require cameras that locate people without recording who they are. Optical privacy cameras blur incoming light before capture, yet the strongest existing design loses 31% of the person detection average precision (AP) of an ordinary camera without optical blur. We find that existing optical privacy cameras apply a constant pixel-width blur fixed on the sensor, which makes the blur too small on a near face and too large on a distant person. Protecting near faces therefore requires sensor blur strong enough to wash out body outlines and miss distant people. We propose a camera focused at infinity with a Gaussian aperture, an opening whose radial transmission follows a smooth Gaussian curve, so the camera fixes a constant physical blur scale in the scene that removes facial detail and keeps body outlines across distances at zero computational cost. On the COCO person detection and Labeled Faces in the Wild (LFW) face verification benchmarks, our camera raises person detection AP from 36.9 to 45.7 over the state-of-the-art Meta-Optics privacy camera and lowers face verification area under the receiver operating characteristic curve (AUROC, where 1.0 is full separation and 0.5 is random guessing) from 0.72 to 0.66. We further pass real photographs of a person captured at close and far range through our camera model, and verify that the detector locates the person while the face cannot be matched to an unblurred photo of the same person even after deblurring. Defining optical blur at physical scene scale establishes a new design principle for privacy-preserving vision, where lens geometry separates human presence from personal identity before light becomes data.
Authors
- Haoyang Zhang (ORCID: https://orcid.org/0000-0002-7220-3734)
- Jie Bai
Institutions
- University of East London (GB)
- University College London (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24424341
- Primary Topic
- Privacy-Preserving Technologies in Data
- Type
- preprint