Patent · 2026 · published application
One warning, at the right moment
A chest-worn camera sees many things; a blind wearer can act on one. This application claims the path between detection and the bone-conduction speaker, the vibration motors and the tones: a single threat value that decides which object gets the warning, a detector that lowers its bar while the wearer moves, and a four-stage scheduler that decides when. Both labs below use the equations and constants as published.
- Indian Patent Application No.
- 202641110070
- Status
- Published, 2026
- Output channels
- Speech · tones · vibration
- Decisions per frame
- One object, or silence
Try it
1. Which obstacle gets the one warning
Drag an object sideways; pick one to change its size.
Motion 0.07 → detector threshold 0.59Eq. 1–2: σ = max(0.35, 0.6 − 0.15 · min(M, 2.5)). A moving camera blurs the image, so confidence drops; the threshold drops with it instead of losing the obstacle.
| Object | P | C | S | M | T |
|---|---|---|---|---|---|
| chairannounced | 0.67 | 0.96 | 0.13 | 0.5 | 0.63 |
| personoutranked | 0.33 | 0.40 | 0.10 | 0.5 | 0.35 |
| bottlefindable class: never an obstacle warning | – | ||||
| doornot detected: confidence 0.48 < 0.59 | – | ||||
T = 0.4·P + 0.25·C + 0.15·S + 0.2·M; an object needs T ≥ 0.3 to be spoken about, and only the highest is.
2. When the warning is allowed out
Point at a dot. 6 warnings in 12 s from 49 cycles; everything else was held back by one of the four stages.
- below τ: silent
- seen once: wait for a second cycle
- same news within the cooldown
- another warning just played
- announced
- announced early: the threat rose by more than ε
The asymmetry is the point: escalation can skip the per-state cooldown (repeating news the wearer already has) but never the global gap, which keeps one output channel from piling up stale warnings. Set the gap to 0 and count how many warnings stack up.
What is claimed, in plain words
One measurement, used twice. The inertial sensor beside the camera measures how much the wearer is moving. That number lowers the detector's confidence threshold, because a blurred frame yields weaker detections of the same obstacle, and it also feeds the threat value, because walking toward an object makes it more urgent than standing still.
Everything reduced to one ordinal value. Proximity (calibrated per class, so a bottle and a person are “very close” at different sizes), centrality, apparent size and the wearer's motion combine into a single number. Only the highest ranks, and only if it clears a threshold; a frame of distant, peripheral objects produces silence.
A gate on the label, not the value. Small objects like a cup can fill the frame when held close without being in the way, so findable classes never trigger an obstacle warning, whatever their score.
Two intervals with different jobs. A per-object cooldown stops repeating what the wearer already knows, and may be skipped when the threat has grown; a global gap protects the shared output channel and is never skipped.
The software side of this was built and evaluated in BlindAssist. The application can be looked up by its number on the Indian Patent Office public search. Team invention.