Drowsiness Detection

Real-time driver drowsiness detection using the Eye Aspect Ratio (EAR). Originally an IEEE conference paper; this repo packages it as a production-shaped Python library with a CLI, two interchangeable landmark backends, a stateful detector with event hooks, and a headless batch-analysis mode.

S. Santhanam et al., "Real-Time Drowsiness Detection using Computer Vision", 2021 5th ICCMC (IEEE). Read the paper →

---

What's new vs. the original script

	Old script	This package
Architecture	Single file, globals, blocking alarm	Stateful DrowsinessDetector with FSM (AWAKE/DROWSY), event callbacks
Alarm	Blocking — stalled the detection loop on every trigger	Background thread + cooldown (Alarm); never blocks frames
Alarm spam	Fired every frame while eyes closed	Fires once per AWAKE→DROWSY transition
Landmarks	dlib + face_recognition (CMake hell)	MediaPipe FaceMesh by default (pure-pip, faster, GPU-friendly) — face_recognition still available
API	None	DrowsinessDetector, DrowsinessConfig, DrowsyEvent, DetectionResult
CLI	Single mode	drowsy run (live) and drowsy analyze (headless batch → CSV/JSONL/JSON)
Tests	None	14 pytest cases, FSM and alarm fully covered, runs in CI on Python 3.9/3.11/3.12
Robustness	Crashed on missing alarm file; reset on lost detection	Soft-fails on missing audio; preserves state across brief detection dropouts
Visualization	EAR text only	EAR vs threshold, FPS counter, color-coded eye polylines, drowsy banner

---

Install

# Default install: package + OpenCV + NumPy
pip install -e .

# Recommended: add MediaPipe (no native build, very fast)
pip install -e ".[mediapipe,audio]"

# Or, original dlib-based backend (requires CMake)
pip install -e ".[legacy,audio]"

---

Quick start

CLI

# Live webcam detection with overlay
drowsy run 0

# Headless batch analysis of a video — emits CSV/JSONL/JSON + a self-contained HTML dashboard
drowsy analyze drive.mp4 --output-dir report/
# → report/report.html  (open it; no JS, no CDN, prints cleanly)

# Tune sensitivity
drowsy run 0 --ear-threshold 0.22 --frame-count 24

# Use the legacy dlib backend
drowsy run 0 --backend face_recognition

Press q or Esc to quit live mode.

Python API

import cv2
from drowsiness import DrowsinessDetector, DrowsinessConfig

cfg = DrowsinessConfig(ear_threshold=0.23, closed_frames_to_alarm=24)

with DrowsinessDetector(cfg, on_event=lambda e: print("DROWSY!", e)) as detector:
 cap = cv2.VideoCapture(0)
 while True:
 ok, frame = cap.read()
 if not ok:
 break
 result = detector.process(frame)
 # result.state: EyeState.AWAKE | EyeState.DROWSY
 # result.ear: current Eye Aspect Ratio
 # result.event: DrowsyEvent or None (fires once per drowsy episode)

---

How it works

Eye Aspect Ratio

EAR = (||p2 - p6|| + ||p3 - p5||) / (2 × ||p1 - p4||)

 p2 p3
 •--------•
 / \
p1 • • p4
 \ /
 •--------•
 p6 p5

• Eyes open → EAR ≈ 0.30
• Eyes closed → EAR ≈ 0.05
• EAR below threshold for N consecutive frames → drowsy event

Detection FSM

 EAR < threshold (N frames)
 ┌──────────┐ ─────────────────────────▶ ┌──────────┐
 │ AWAKE │ │ DROWSY │
 └──────────┘ ◀───────────────────────── └──────────┘
 EAR ≥ threshold (M frames)

The state machine de-bounces both directions: a brief blink doesn't trigger an alarm, and a single open frame doesn't immediately clear a drowsy state.

Pipeline

Webcam / Video
 │
 ▼
 Landmark backend ──▶ 6 EAR points per eye
 (MediaPipe FaceMesh
 or face_recognition)
 │
 ▼
 EAR computation ──▶ average EAR
 │
 ▼
 Drowsy FSM ──▶ state, closed_frames, event?
 │
 ├─ Alarm (background, debounced)
 ├─ on_event callback (your code)
 └─ Visualization overlay

---

Configuration

DrowsinessConfig(
 ear_threshold=0.25, # below = closed eye
 closed_frames_to_alarm=20, # frames below threshold → DROWSY
 open_frames_to_clear=5, # frames above threshold → AWAKE
 alarm_sound="assets/alert1.mp3",
 alarm_cooldown_s=3.0, # min seconds between alarm playbacks
 backend="mediapipe", # or "face_recognition"
 enable_alarm=True,
)

CLI flag	Default	Notes
--ear-threshold	0.25	Tune lower for darker/glasses-heavy footage.
--frame-count	20	At 30 FPS this is ~0.7s of closed eyes.
--open-frames	5	Clears the drowsy state after a sustained reopen.
--alarm-cooldown	3.0	Seconds. Prevents back-to-back alarms.
--backend	mediapipe	mediapipe or face_recognition.
--no-alarm	off	Disables audio (good for benchmarking).
--no-window	off	Headless run (no cv2.imshow).

---

Batch analysis

drowsy analyze runs a video through the same detector with audio disabled and produces a small report:

report/
├── ear.csv # frame-by-frame EAR & state
├── events.jsonl # one record per AWAKE→DROWSY transition
└── summary.json # counts, parameters, timestamps

Useful for fleet review, evaluating threshold changes, or feeding downstream analytics without a GUI.

---

Project structure

drowsiness/
├── __init__.py # public API
├── ear.py # eye_aspect_ratio() + helpers
├── landmarks.py # MediaPipe + face_recognition backends
├── alarm.py # background, cooldown-debounced audio
├── detector.py # DrowsinessDetector + FSM + events
├── visualization.py # cv2 overlay (EAR, FPS, banner)
└── cli.py # `drowsy` entrypoint
tests/
└── test_detector.py # 14 cases, mocked landmark backend
assets/
└── alert1.mp3

---

Limitations

This is still an EAR-based system. Known weaknesses:

• Sunglasses / heavy bangs / extreme head pose break landmark detection.
• A single threshold doesn't fit every driver — production deployments should calibrate per-driver during the first ~30s of awake driving.
• No yawn / head-nod / gaze-direction signal yet (PRs welcome — DrowsyEvent is the right place to fan out into multi-cue fusion).
• No tracking across multiple faces; the first detected face is used. Cabin-facing cameras typically only see one driver.

---

Citation

@inproceedings{santhanam2021drowsiness,
 title = {Real-Time Drowsiness Detection using Computer Vision},
 author = {Santhanam, S. and others},
 booktitle = {2021 5th International Conference on Computing Methodologies and Communication (ICCMC)},
 year = {2021},
 doi = {10.1109/ICCMC51019.2021.9418325}
}

License

MIT © Sanjay Santhanam