Small IoT Research ProjectMITE Cohort 19 · RUPP
DrowsyGuard
A USD 27 retrofit device that watches a driver's eyes, mouth and head on the ESP32‑S3 itself, and speaks a warning — with no cloud anywhere in the alert path.
02/08
The problem
Why one frame is not enough
A closed eye is not drowsiness.
A closed eye for long enough is.
-
01
Single-frame alarms cry wolf
Normal blinking, looking down at the gear stick, glasses glare and a moment of lost face all read as “eyes closed” to a per-frame classifier.
-
02
Appearance is not eyelid state
Our own whole-face network scored 0.81 in validation but 0.57 on drivers it had never seen — it had learned who the driver was, not whether their eyes were shut. It was removed.
-
03
Cambodia's older fleet has no monitoring
Driver monitoring ships only in new vehicles. A USD 27 retrofit that runs offline is the only version of this that reaches an existing car.
03/08
Concept — cues
Four signs, each with its own clock
What the camera measures
PERCLOS3.0 s
Share of the last three seconds with the eyes closed. The best-established fatigue measure, so it carries most of the weight.
weight 0.55Long blink>0.40 s
Slower than a blink, shorter than a microsleep. Blink speed drops before a driver knows they are tired.
weight 0.20Yawn≥1.20 s
Mouth corners drop and pull inward. Anything faster is speech, not a yawn.
weight 0.15Head nod0.3–1.5 s
Down and back. Two independent pitch channels must agree, so an open mouth cannot fake it.
weight 0.1004/08
Concept — decision
From four numbers to one spoken word
Risk has to be held, not touched
-
01
Fuse, weighted
risk = 0.55 PERCLOS + 0.20 blink + 0.15 yawn + 0.10 nod. The three minor cues corroborate; none of them can raise an alarm alone.
-
02
Hold, then fire
Risk must stay above 0.55 for 0.55 s. A single noisy frame decrements the streak instead of resetting it, so one dropout inside a real closure cannot end it.
-
03
Name the cause out loud
“Drowsy”, “Microsleep”, “Yawning”, “HeadNod” — a named warning is actionable in a way a chime is not. Three announcements per episode, then quiet.
-
04
Wi-Fi is never load-bearing
The alert path never touches the network. A dropped dashboard costs diagnostics, never safety.
05/08
Process flow
One safety path, two support paths
Everything that matters happens on the board
06/08
Equipment
Bill of materials — all bought locally
Nine parts, one breadboard, no display
ESP32‑S3‑WROOM‑1 + OV5640
N16R8 · 8 MB PSRAM
Camera, on-device inference and Wi-Fi in one board. $7.50
MAX98357A amplifier
I2S · class D · filterless
Turns the spoken clip into something audible. $2.00
Speaker
4 Ω / 3 W · 40 × 22 mm
The only thing the driver actually perceives. $0.75
Breadboard MB102
830 point · solderless
Keeps the wiring repeatable between test sessions. $1.50
Jumper wires & headers
Dupont · 1 set
Seven connections: power, ground and three I2S lines. $2.00
USB‑C cable
data-capable · 5 V
Flashing and bench power. Charge-only cables will not do. $1.50
5 V USB supply
regulated adapter
Prototype power; an automotive supply comes at pilot stage. $2.00
Inline USB power meter
measurement instrument
Turns “power draw” from a claim into a number. $5.00
Enclosure & mount
prototype materials
Same camera distance and angle every session, or the data is not comparable. $5.00
Laptop and phone are excluded — the team uses existing devices. No display, by decision: a 1.8” panel cost five GPIOs, 150 KB of PSRAM and a per-frame blit to show 8-pixel text to one person. The phone browser shows strictly more, and makes the same numbers scriptable for the acceptance tests.
$27.25Bill of materials
07/08
Where we are
Measured on the bench, not estimated
The board runs. The model does not see well enough yet.
Frame rate, idle / tracking a facetarget ≥ 10 fps
19.7 / 10.7–13.6 fps
met
Face detection on a real person20 attempts, confidence 1.00
20 / 20
met
Cost per stage39 ms detect · 15.7 ms per eye
within budget
met
PERCLOS moving on live video0.00–0.22 at rest, eye model bound
live
met
Eye-state accuracy in visible lightbase model is infrared-trained; 95.8% in its own domain
AUC 0.62
the gap
Alert end to end, and cues against a real yawnlogic and signs now correct; nobody has yawned at it yet
not yet run
next
What running it on hardware taught us: the sensor is mounted upside down, so every frame arrives mirrored. That reversed the eye pair and flipped the sign of every vertical measurement — jaw drop read −1.2 instead of +1.2 and fell as the mouth opened, so the yawn cue had never once been able to fire. No amount of desk testing would have found it.
08/08
Conclusion
Where the remaining effort goes
The hardware is done arguing.
The next win is in the model.
Our focus from here is training the lightweight on-device model — 11,250 parameters, 46 KB — on visible-light eye-state data, split by subject, so it reads eyelids instead of faces.
-
→
Fine-tune in the target domain
Infrared-trained, it does not transfer to 45 px visible-light eye crops: AUC 0.62. Grayscale, CLAHE, inversion and four patch scales did not recover it. The fix is labelled data in our own domain.
-
→
Judge it per driver, never on average
Subject-independent splits only. The 0.81-to-0.57 collapse is the standard any replacement has to clear.
-
→
Then tune the thresholds on real video
Every period on slide 3 is literature-informed and tested on synthetic traces. Labelled yawn and nod footage turns them from defaults into measurements.
- Detection quality
- F1 ≥ 0.80and event recall ≥ 0.85
- False warnings
- ≤ 1 / hourper awake control hour
- Response
- ≤ 2 safter sustained risk
- Reliability
- 4 hcontinuous, no crash
Why not a large model: the most accurate public drowsiness networks are 70–343 MB at 224×224 and cannot run on an ESP32‑S3 at any frame rate. The eye-closure route is the one that fits. Research prototype only — not a certified safety device.