01
See the scene
The rear camera feeds an on-device YOLO vision model. People, animals and other obstacles become detection boxes.
A little awareness goes a long way
Life moves.
Look up.
Meet NoBonk. On-device AI that watches the path ahead and gives you a heads-up when something gets too close.
See how it thinksBuilt by a curious student. Made for the real world.
MORE AWARE. LESS BONK.
An illustrated world. A real problem.A student idea with momentum
1st place STEM4All 2026
IEEE Award STEM4All 2026
3rd in category ACSEF 2026
01 / The idea
Your phone has a camera.
Give it a purpose.
One glance at your screen. One person stepping into your path. NoBonk explores a simple question: what if your phone could give you a little nudge?
02
Notice what’s closing in
The app tracks how boxes grow and move. Frame fill and approach signals help distinguish a nearby object from one closing in.
03
Make the nudge count
LOW, MEDIUM and HIGH alerts escalate through vibration and on-screen warnings, including while another app is open.
02 / Inside the decision
Don’t just read it.
Move the signal.
Make a person’s detection box grow. Add an approach signal. Watch the alert change using the app’s actual person-alert thresholds.
DETECTION LABILLUSTRATIVE INPUT
PERSON35% frame height
0%FRAME HEIGHT100%
PERSON / REFERENCE PRESET 2 mLOW
You control the input
Smaller in frameLarger in frame
The tracker sees a closing object.
On-bearing, estimated contact ≤ 1.5 s.
The decision
low alert.
At 28% frame height and above, the person ladder reaches LOW.
Educational simulation of the person alert policy at the 2 m sensitivity preset. No camera access. This is not a distance measurement; the full app also filters detections and applies tracking and alert timing.
Read the actual logic03 / Out of the lab
Real experiments.
Honest numbers.
Early results reported by the project on a Pixel 9a. Promising signals, with plenty still to learn. These are prototype observations, not independent validation.
Detection in good light
Reported detection accuracy under good lighting.
Approach trials correct
Reported approach-detection trial results.
Distance error at 1 m
Distance estimates depend on angle and scene.
The next challenge
Darkness changes the picture.
The project reports 3/10 detection in low light and about 10% battery drain per hour in background mode. Older phones can lag. NoBonk is a student-built prototype that can miss hazards. Keep looking up.
Concept illustration / not an app screenshot04 / On your phone. By design.
A little more awareness.
A lot less oversharing.
Your camera frames stay fleeting.
Processed in memory, then discarded. No saved photos or video. The app has no Internet permission.
Your history stays local.
Detection events stay in private storage on your phone and can be cleared. Approximate location is optional and off by default.
THE NEXT STEP IS ANDROID
Less bonk.
More world.
A science-fair question, built into an Android prototype.
Now getting ready for its next chapter.
COMING SOON ONGoogle Play
Android 10+ · In development · Release date to be announced Follow the project on GitHub