DriveGuard Driver Monitoring System

Distraction and drowsiness build up over seconds, not frames. A camera-based monitor needs to watch eyes, gaze, head pose, and yawns together, then escalate only when the evidence holds — not on a single blink.

• Real-time facial landmark analysis with MediaPipe
• Multi-signal cues: eye closure (PERCLOS), blinks, yawns, looking away, phone
• Tiered danger levels with distinct actions (none → soft beep → critical alarm)
• Temporal timers so brief gestures do not trip false alarms

Webcam frames → OpenCV capture → MediaPipe face landmarks → feature extractors (PERCLOS, blink/yawn counts, looking-away / mouth-open timers) → smoothed danger score → level state machine (L0 / L1 / L3) → HUD + alert action.

Stack: Python, OpenCV, MediaPipe, NumPy, Rich (console UI), Requests / python-dotenv for optional remote or config hooks.

• L0 SAFE — driver looks alert; action none
• L1 CAUTION — early cues (partial eye closure, yawn); actions like soft_beep or vibration_and_beep
• L3 DANGER — sustained risk (e.g. looking away for several seconds); action critical_alarm

Live metrics on the HUD include FPS, smoothed score, PERCLOS, blink/yawn totals, and per-cue timers (eyes closed, mouth open, looking away, phone seen).

Each cue contributes to a smoothed score. Sustained evidence — not single-frame spikes — advances the level. Landmark overlays (eyes in green, mouth in blue) make it obvious what the model is measuring while you tune thresholds.

Staged behaviors in front of the webcam: alert baseline (L0), blinks and yawns (L1), and looking down / away long enough to hit L3. Thresholds and warn/danger dwell times were tuned from those runs.

Working real-time monitor with clear L0 / L1 / L3 escalation, on-screen telemetry, and alert actions tied to blink, yawn, gaze, and PERCLOS cues.

• Lighting and camera angle change landmark quality
• Natural blinks vs. drowsiness without temporal filtering
• Balancing sensitivity so L1 fires early without constant false L3s

Timers and hysteresis matter more than any single landmark feature. PERCLOS and yawn/blink counts only become useful once you require evidence over time.

• Calibration UX for different cameras and seating positions
• On-device optimization for lower-power hardware
• More labeled clips for threshold tuning