Problem.

Aynkan is a single-camera distance estimation system for an assistive navigation wearable, designed for blind and low-vision users. A wall that’s three feet away should never be reported as five feet away, even if the model is uncertain.

Approach.

It pairs an off-the-shelf object detector with a learned monocular depth model and fuses the two distance signals with explicit handling of their shared bounding-box noise.

For each object, it outputs a distance estimate together with a confidence range, and it is tuned to err toward warning when unsure rather than missing an obstacle.

FIG. 1 — ASYMMETRIC CONFIDENCE RANGE

0 2 4 6 8 10 distance (m) 4.2 m
predicted 4.2 m · band [3.2, 4.6] m
Fig. 1 — Asymmetric confidence range. Drag to vary predicted distance; toggle to compare symmetric vs asymmetric ranges. The asymmetric form penalizes overestimates more strongly than underestimates — a near obstacle reported as far is the dangerous error.

Implementation.

The system runs on a smartphone or Raspberry Pi 5 class device

and uses only Apache-2.0-licensed components, so it stays redistributable.

Built with OpenCV and Ultralytics YOLO; in progress, developing toward submission to the RESNA Student Design Challenge 2027. I’m building it with a hardware collaborator.

Detection overlay
forthcoming
Pl. 01Sample frame with bounding boxes, distance estimates, and confidence values overlaid.

Result.

In progress. The detector and depth model are wired into the fusion stage; the confidence ranges are being tuned against held-out sequences.

Notes & future work.

Targeted at the RESNA Student Design Challenge. Open questions: how cautious to be in indoor versus outdoor scenes, and whether a thin temporal smoothing layer earns its compute on a wearable.

References.

rev?
  1. Conformal prediction reference — to be filled in.
  2. Monocular depth model citation — to be filled in.
  3. YOLO citation — to be filled in.

Filed under: assistive technology, perception systems, calibration.

Kingston · February MMXXVI