One week after the initial idea: a functional app on two Samsung flagships, 27/27 tests green, APK under 30 MB. What happened on the way there.

Model selection and size

For AI traffic light recognition, there are several YOLO variants. I started with the 640px model – detection quality good, latency on the S23 Ultra around 215 ms. Too slow for live use. The 320px model achieves ~88 ms on the S23 Ultra, ~50 ms on the S25 Ultra. Usable.

APK size was initially 52 MB. arm64-only ABI filter and removal of the 640px asset reduced this to 28 MB. Better.

Freemium mechanics

Light and Pro are not two separate apps, but one. The CV model is always included (~28 MB) but locked. Upon the first Pro launch, a real benchmark runs on the device: the app measures inference latency under real conditions, not on a test image. If the average is below 150 ms, Pro is unlocked. If it is above, CV remains deactivated – ensuring no poor experience on weak hardware.

Purchase is currently a stub – unlockable locally. Play Billing will be implemented when the store rollout occurs.

What the tests revealed

Dual-load soak for 20 minutes on the S23 Ultra: camera, GPS, and display running in parallel. Thermal equilibrium at ~43 °C, CV latency stable at ~88 ms, no crash. GPS on real drives: 3-meter accuracy, excellent.

Hazard warning runs on offline OSM data (1,102 saved points in Hessen), speed limit from a local index with 921,000 points – loads for a few seconds upon first launch, then instant.

Traffic light CV during real driving: ~30 detections, but no confirmed color. Detection works, but smoothing at distance and motion is not yet good enough. This is the next open topic. Requires a proper camera mount for testing.

Status quo and next steps: tomorrow.