1. The Mathematics of Eye Aspect Ratio (EAR)

Traditional blink detection relied on heavy CNN classifiers running per-frame bounding boxes, which is computationally prohibitive on low-power vehicle hardware. In this architecture, 6 localized 3D ocular landmarks are extracted to compute the scalar aspect ratio:

# Soukupová and Čech EAR Equation Formulation
def calculate_ear(eye_landmarks):
    # Vertical euclidean distances
    A = np.linalg.norm(eye_landmarks[1] - eye_landmarks[5])
    B = np.linalg.norm(eye_landmarks[2] - eye_landmarks[4])
    # Horizontal euclidean distance
    C = np.linalg.norm(eye_landmarks[0] - eye_landmarks[3])
    
    # Eye Aspect Ratio calculation
    ear = (A + B) / (2.0 * C)
    return ear

2. Consecutive Frame Hysteresis (Filtering Blinks)

Normal involuntary blinks last between 100 ms and 250 ms ($3\text{--}7$ frames at 30 FPS). A microsleep event is defined when the eyelid remains closed past 500 ms ($N \ge 16$ consecutive frames where $\text{EAR} < \text{Threshold}$). An integrated state counter prevents false alarm fatigue while ensuring immediate escalation when genuine sleep is identified.

3. Key System Capabilities