🎯 How to Calibrate and Check Accuracy

The "accuracy" of this system depends on how well it is tuned for a specific user and environment.

Real-time Calibration

The two most important parameters to tune are thresh and frame_check in the Drowsiness_Detection.py script.

• Calibrating thresh (EAR Threshold):

  1. Run the script and observe the EAR: values printed in your terminal.
  2. Note the typical EAR value when your eyes are open.
  3. Note the EAR value when you blink.
  4. The ideal thresh should be between your "open-eye" and "blinking" EAR values. The default of 0.25 is a good starting point, but you may need to adjust it.
  5. Update the thresh = 0.25 line in the script with your calibrated value.

• Calibrating frame_check:

  1. This parameter sets the number of consecutive frames your eyes must be closed for the alarm to trigger.
  2. If the alarm is too sensitive (e.g., triggers on long blinks), increase this value (e.g., to 25 or 30).
  3. If the alarm is too slow to react, decrease it (e.g., to 15).

Formal Accuracy Checking

For a more systematic evaluation of the model's performance, follow these steps:

1. Create a Labeled Dataset:

   • Record short video clips under realistic conditions (e.g., different people, lighting).
   • Label each clip with a "ground truth": either "drowsy" or "alert."

2. Run the Model on Your Dataset:

   • Modify the script to process video files instead of the live webcam. Change cv2.VideoCapture(0) to cv2.VideoCapture('path/to/your/video.mp4').
   • Run the script on each video and log whether it triggered an "ALERT!" or not.

3. Evaluate Performance:

   • Compare the script's predictions to your ground truth labels. This will give you:
     • True Positives (TP): Correctly predicted "drowsy."
     • True Negatives (TN): Correctly predicted "alert."
     • False Positives (FP): Incorrectly predicted "drowsy" (false alarm).
     • False Negatives (FN): Incorrectly predicted "alert" (missed detection).

4. Calculate Metrics:

   • Use these counts to calculate standard performance metrics:
     • Accuracy: (TP + TN) / (Total Videos)
     • Precision: TP / (TP + FP) (Measures false alarm rate)
     • Recall: TP / (TP + FN) (Measures missed detection rate) ..