AI Driver Drowsiness Detection System -- 3

I need a deep-learning solution that watches a driver’s face through a standard camera feed, tracks eye-closure patterns and yawning frequency, then translates those cues into a clear fatigue score that updates continuously. Over a journey the model should also plot a time-based curve so I can see how alertness rises or falls.

Please build and train the full pipeline in Python, preferably with PyTorch or TensorFlow paired with OpenCV for video handling. The system must be completely vision-based; no wearables or contact sensors. I will supply sample clips for initial testing, but the code should accept any 30 fps video stream so I can later attach it to an in-car webcam.

The final hand-off should include:
• Inference script that ingests a live or recorded feed, detects eyes and mouth, classifies drowsiness level frame-by-frame, and logs a running fatigue score.
• Function that converts those scores into a simple progression curve (CSV or JSON + plotted graph).
• Trained weights and a read-me explaining dependencies, model architecture, and how to retrain with new data.

Accuracy, low latency, and robustness under different lighting conditions are key acceptance criteria.