Low-Cost Driver Supervision.

The problem

Distraction and drowsiness are behind a large share of serious crashes. In the US, distraction played a part in 8% of fatal crashes in 2024, and drowsy drivers were involved in an estimated 17.6% of fatal crashes between 2017 and 2021. Camera-based driver monitoring helps, but it’s usually built into the car. The EU made it mandatory for new vehicles from July 2026, yet millions of cars already on the road have nothing at all. We wanted to see how far an ordinary smartphone could close that gap.

What we built

An Android app that uses the phone’s front camera as a low-cost driver monitor, running entirely on the device. No video is ever recorded or stored.

• Fatigue: Google ML Kit face detection tracks eye openness to tell a normal blink from prolonged eye closure, and measures mouth openness over time to catch yawning.
• Distraction: a custom YOLOv8 model, trained to spot phone use and reading and converted to TensorFlow Lite to run in real time on the phone.
• Face not detected: if the driver turns fully away, the app flags it.
• Alerts: when it detects one of these states, the app shows a banner naming the problem with a clear instruction (e.g. “Phone use detected. Please put away your phone and focus on driving.”) plus a sound, and keeps it up until the driver is attentive again.

What we found

We ran a pilot study with 30 participants across 180 trials in a driving simulator. The system caught 147 of the 180 target behaviours.

• Fatigue signals were the strongest: prolonged eye closure scored an F1 of 0.94, and face-not-detected 0.91.
• Object-based distractions were harder: phone use (0.75) and reading (0.76) produced more false positives, because phones and small books look alike to the camera, especially when partly hidden by hands.
• Alerts fired within 2–3.5 seconds, and drivers were back to attentive within about 1.5–3.3 seconds of being alerted.

What’s next

Smoothing detections across multiple frames, adding gaze and hand-object cues to cut false positives, testing less intrusive phone positions that keep the road view clear, and measuring how drivers feel about the alerts: trust, clarity and how intrusive they are.

Check out the demo here.