models/eye_scorer.py

import math

import numpy as np

_LEFT_EYE_EAR = [33, 160, 158, 133, 153, 145]
_RIGHT_EYE_EAR = [362, 385, 387, 263, 373, 380]

_LEFT_IRIS_CENTER = 468
_RIGHT_IRIS_CENTER = 473

_LEFT_EYE_INNER = 133
_LEFT_EYE_OUTER = 33
_RIGHT_EYE_INNER = 362
_RIGHT_EYE_OUTER = 263

_LEFT_EYE_TOP = 159
_LEFT_EYE_BOTTOM = 145
_RIGHT_EYE_TOP = 386
_RIGHT_EYE_BOTTOM = 374

_MOUTH_TOP = 13
_MOUTH_BOTTOM = 14
_MOUTH_LEFT = 78
_MOUTH_RIGHT = 308
_MOUTH_UPPER_1 = 82
_MOUTH_UPPER_2 = 312
_MOUTH_LOWER_1 = 87
_MOUTH_LOWER_2 = 317

MAR_YAWN_THRESHOLD = 0.55


def _distance(p1: np.ndarray, p2: np.ndarray) -> float:
    return float(np.linalg.norm(p1 - p2))


def compute_ear(landmarks: np.ndarray, eye_indices: list[int]) -> float:
    p1 = landmarks[eye_indices[0], :2]
    p2 = landmarks[eye_indices[1], :2]
    p3 = landmarks[eye_indices[2], :2]
    p4 = landmarks[eye_indices[3], :2]
    p5 = landmarks[eye_indices[4], :2]
    p6 = landmarks[eye_indices[5], :2]

    vertical1 = _distance(p2, p6)
    vertical2 = _distance(p3, p5)
    horizontal = _distance(p1, p4)

    if horizontal < 1e-6:
        return 0.0

    return (vertical1 + vertical2) / (2.0 * horizontal)


def compute_avg_ear(landmarks: np.ndarray) -> float:
    left_ear = compute_ear(landmarks, _LEFT_EYE_EAR)
    right_ear = compute_ear(landmarks, _RIGHT_EYE_EAR)
    return (left_ear + right_ear) / 2.0


def compute_gaze_ratio(landmarks: np.ndarray) -> tuple[float, float]:
    left_iris = landmarks[_LEFT_IRIS_CENTER, :2]
    left_inner = landmarks[_LEFT_EYE_INNER, :2]
    left_outer = landmarks[_LEFT_EYE_OUTER, :2]
    left_top = landmarks[_LEFT_EYE_TOP, :2]
    left_bottom = landmarks[_LEFT_EYE_BOTTOM, :2]

    right_iris = landmarks[_RIGHT_IRIS_CENTER, :2]
    right_inner = landmarks[_RIGHT_EYE_INNER, :2]
    right_outer = landmarks[_RIGHT_EYE_OUTER, :2]
    right_top = landmarks[_RIGHT_EYE_TOP, :2]
    right_bottom = landmarks[_RIGHT_EYE_BOTTOM, :2]

    left_h_total = _distance(left_inner, left_outer)
    right_h_total = _distance(right_inner, right_outer)

    if left_h_total < 1e-6 or right_h_total < 1e-6:
        return 0.5, 0.5

    left_h_ratio = _distance(left_outer, left_iris) / left_h_total
    right_h_ratio = _distance(right_outer, right_iris) / right_h_total
    h_ratio = (left_h_ratio + right_h_ratio) / 2.0

    left_v_total = _distance(left_top, left_bottom)
    right_v_total = _distance(right_top, right_bottom)

    if left_v_total < 1e-6 or right_v_total < 1e-6:
        return h_ratio, 0.5

    left_v_ratio = _distance(left_top, left_iris) / left_v_total
    right_v_ratio = _distance(right_top, right_iris) / right_v_total
    v_ratio = (left_v_ratio + right_v_ratio) / 2.0

    return float(np.clip(h_ratio, 0, 1)), float(np.clip(v_ratio, 0, 1))


def compute_mar(landmarks: np.ndarray) -> float:
    top = landmarks[_MOUTH_TOP, :2]
    bottom = landmarks[_MOUTH_BOTTOM, :2]
    left = landmarks[_MOUTH_LEFT, :2]
    right = landmarks[_MOUTH_RIGHT, :2]
    upper1 = landmarks[_MOUTH_UPPER_1, :2]
    lower1 = landmarks[_MOUTH_LOWER_1, :2]
    upper2 = landmarks[_MOUTH_UPPER_2, :2]
    lower2 = landmarks[_MOUTH_LOWER_2, :2]

    horizontal = _distance(left, right)
    if horizontal < 1e-6:
        return 0.0
    v1 = _distance(upper1, lower1)
    v2 = _distance(top, bottom)
    v3 = _distance(upper2, lower2)
    return (v1 + v2 + v3) / (2.0 * horizontal)


class EyeBehaviourScorer:
    def __init__(
        self,
        ear_open: float = 0.30,
        ear_closed: float = 0.16,
        gaze_max_offset: float = 0.28,
    ):
        self.ear_open = ear_open
        self.ear_closed = ear_closed
        self.gaze_max_offset = gaze_max_offset

    def _ear_score(self, ear: float) -> float:
        if ear >= self.ear_open:
            return 1.0
        if ear <= self.ear_closed:
            return 0.0
        return (ear - self.ear_closed) / (self.ear_open - self.ear_closed)

    def _gaze_score(self, h_ratio: float, v_ratio: float) -> float:
        h_offset = abs(h_ratio - 0.5)
        v_offset = abs(v_ratio - 0.5)
        offset = math.sqrt(h_offset**2 + v_offset**2)
        t = min(offset / self.gaze_max_offset, 1.0)
        return 0.5 * (1.0 + math.cos(math.pi * t))

    def score(self, landmarks: np.ndarray) -> float:
        left_ear = compute_ear(landmarks, _LEFT_EYE_EAR)
        right_ear = compute_ear(landmarks, _RIGHT_EYE_EAR)
        # Use minimum EAR so closing ONE eye is enough to drop the score
        ear = min(left_ear, right_ear)
        ear_s = self._ear_score(ear)
        if ear_s < 0.3:
            return ear_s
        h_ratio, v_ratio = compute_gaze_ratio(landmarks)
        gaze_s = self._gaze_score(h_ratio, v_ratio)
        return ear_s * gaze_s

    def detailed_score(self, landmarks: np.ndarray) -> dict:
        left_ear = compute_ear(landmarks, _LEFT_EYE_EAR)
        right_ear = compute_ear(landmarks, _RIGHT_EYE_EAR)
        ear = min(left_ear, right_ear)
        ear_s = self._ear_score(ear)
        h_ratio, v_ratio = compute_gaze_ratio(landmarks)
        gaze_s = self._gaze_score(h_ratio, v_ratio)
        s_eye = ear_s if ear_s < 0.3 else ear_s * gaze_s
        return {
            "ear": round(ear, 4),
            "ear_score": round(ear_s, 4),
            "h_gaze": round(h_ratio, 4),
            "v_gaze": round(v_ratio, 4),
            "gaze_score": round(gaze_s, 4),
            "s_eye": round(s_eye, 4),
        }