VERIDIC / src /feature_engine.py
turboryzen's picture
Upload folder using huggingface_hub
e95cd0c verified
Raw History Blame Contribute Delete
7.3 kB
# src/feature_engine.py
import numpy as np
from src.head_pose import head_pose_euler
from src.rppg import forehead_green, compute_hr
from src.emotion import emotion_from_blendshapes, facs_deception_features, _NEG_EMOTIONS
# ── EAR ────────────────────────────────────────────────────────────────────────
def eye_aspect_ratio(lm: np.ndarray, idx: list) -> float:
"""EAR clasico de 6 puntos."""
p1, p2, p3, p4, p5, p6 = [lm[i] for i in idx]
A = np.linalg.norm(p2 - p3)
B = np.linalg.norm(p5 - p6)
C = np.linalg.norm(p1 - p4)
return float((A + B) / (2.0 * C + 1e-8))
def blink_rate_from_ears(ears: np.ndarray, fps: float) -> float:
"""Parpadeos por segundo con umbral dinamico: 75% del EAR basal individual."""
threshold = float(ears.mean() * 0.75)
below = ears < threshold
transitions = np.diff(below.astype(int))
n_blinks = int(np.sum(transitions == 1))
duration_s = len(ears) / max(fps, 1.0)
return n_blinks / max(duration_s, 1.0)
# ── IRIS ───────────────────────────────────────────────────────────────────────
def iris_size_ratio(lm: np.ndarray,
corner_a: int, corner_b: int,
iris_center: int, iris_perimeter: list) -> float:
"""
Radio del iris normalizado por la anchura del ojo.
Usa los landmarks 468-477 (iris) que devuelve FaceLandmarker.
"""
center = lm[iris_center]
perim_pts = [lm[i] for i in iris_perimeter]
radius = np.mean([np.linalg.norm(center - p) for p in perim_pts])
eye_width = np.linalg.norm(lm[corner_a] - lm[corner_b])
return float(radius / (eye_width + 1e-8))
# ── PIPELINE TEMPORAL ──────────────────────────────────────────────────────────
def compute_video_features(
frames: list,
lm_list: list,
bs_list: list, # blendshapes por frame (dict o None)
fps: float,
audio=None,
audio_sr: int = 16000,
) -> dict | None:
"""
Agrega todas las features temporales de un video.
Las emociones se derivan de los blendshapes de MediaPipe (sin DeepFace).
Devuelve dict con FEATURE_NAMES o None si no hay suficientes datos.
"""
from src.audio_features import extract_audio_features
h, w = frames[0].shape[:2] if frames else (480, 640)
ears, ear_ls, ear_rs = [], [], []
iris_ratios = []
rolls, pitches, yaws = [], [], []
green_signal = []
emotions = []
gaze_xs, gaze_ys = [], []
brow_asym_frames = []
for frame, lm, bs in zip(frames, lm_list, bs_list):
if lm is None:
continue
# EAR β€” mantener izquierdo/derecho separados para asimetrΓ­a
ear_l = eye_aspect_ratio(lm, [33, 160, 158, 133, 153, 144])
ear_r = eye_aspect_ratio(lm, [362, 385, 387, 263, 373, 380])
ears.append((ear_l + ear_r) / 2.0)
ear_ls.append(ear_l)
ear_rs.append(ear_r)
# Iris (landmarks 468-477)
ir_l = iris_size_ratio(lm, 33, 133, 468, [469, 470, 471, 472])
ir_r = iris_size_ratio(lm, 362, 263, 473, [474, 475, 476, 477])
iris_ratios.append((ir_l + ir_r) / 2.0)
# Gaze instability β€” centro del iris en coords normalizadas
# 468 = iris izquierdo centro, 473 = iris derecho centro
gaze_xs.append((lm[468][0] + lm[473][0]) / 2.0)
gaze_ys.append((lm[468][1] + lm[473][1]) / 2.0)
# Brow asymmetry β€” altura de ceja vs pΓ‘rpado superior
# 105 = pico ceja izquierda, 334 = pico ceja derecha
# 159 = pΓ‘rpado superior izq, 386 = pΓ‘rpado superior der
brow_lift_l = lm[159][1] - lm[105][1] # positivo = ceja por encima del ojo
brow_lift_r = lm[386][1] - lm[334][1]
brow_asym_frames.append(abs(brow_lift_l - brow_lift_r))
# Head pose
r, p, y = head_pose_euler(lm, w, h)
rolls.append(r); pitches.append(p); yaws.append(y)
# rPPG: canal verde de la frente
green_signal.append(forehead_green(frame, lm))
# Emocion desde blendshapes (instantaneo, sin modelo externo)
emotions.append(emotion_from_blendshapes(bs))
if len(ears) < 10:
return None # datos insuficientes
ears_arr = np.array(ears)
iris_arr = np.array(iris_ratios)
# -- Features faciales --
ear_mean = float(ears_arr.mean())
ear_std = float(ears_arr.std())
blink_r = float(blink_rate_from_ears(ears_arr, fps))
iris_ratio_mean = float(iris_arr.mean())
iris_ratio_std = float(iris_arr.std())
# -- AsimetrΓ­a facial --
ear_asymmetry = float(np.mean(np.abs(np.array(ear_ls) - np.array(ear_rs))))
brow_asymmetry = float(np.mean(brow_asym_frames)) if brow_asym_frames else 0.0
# -- Inestabilidad de mirada --
gaze_std = float(np.sqrt(np.std(gaze_xs)**2 + np.std(gaze_ys)**2))
# -- Head pose -- unwrap para evitar saltos de fase Β±180Β° que disparan el std
head_roll_std = float(np.degrees(np.std(np.unwrap(np.radians(rolls)))))
head_pitch_std = float(np.degrees(np.std(np.unwrap(np.radians(pitches)))))
head_yaw_std = float(np.degrees(np.std(np.unwrap(np.radians(yaws)))))
# -- Emociones --
if emotions:
fear_rate = sum(1 for e in emotions if e == "fear") / len(emotions)
neg_emo_rate = sum(1 for e in emotions if e in _NEG_EMOTIONS) / len(emotions)
else:
fear_rate = neg_emo_rate = 0.0
# -- rPPG --
hr_bpm, hr_std = compute_hr(green_signal, fps)
# -- FACS deception features (display-only, no van al XGBoost aΓΊn) --
facs = facs_deception_features(bs_list)
# -- Audio --
aud_feats = extract_audio_features(audio, audio_sr)
return {
"ear_mean": ear_mean,
"ear_std": ear_std,
"blink_rate": blink_r,
"fear_rate": fear_rate,
"neg_emotion_rate": neg_emo_rate,
"iris_ratio_mean": iris_ratio_mean,
"iris_ratio_std": iris_ratio_std,
"ear_asymmetry": ear_asymmetry,
"brow_asymmetry": brow_asymmetry,
"gaze_std": gaze_std,
"head_roll_std": head_roll_std,
"head_pitch_std": head_pitch_std,
"head_yaw_std": head_yaw_std,
"hr_bpm": hr_bpm,
"hr_std": hr_std,
"pitch_mean": aud_feats["pitch_mean"],
"pitch_std": aud_feats["pitch_std"],
"energy_std": aud_feats["energy_std"],
"pause_ratio": aud_feats["pause_ratio"],
"speech_rate": aud_feats["speech_rate"],
# FACS β€” display only
"contempt_proxy": facs["contempt_proxy"],
"suppression_index": facs["suppression_index"],
"emotion_variance": facs["emotion_variance"],
"au_fear_peak": facs["au_fear_peak"],
}