MediaPipeを用いてPython環境で表情分析アプリを作ってみた

import cv2
import mediapipe as mp
from deepface import DeepFace
from collections import deque, Counter

# Mediapipe 初期化
mp_face_mesh = mp.solutions.face_mesh
face_mesh = mp_face_mesh.FaceMesh(
    static_image_mode=False,
    max_num_faces=1,
    refine_landmarks=True,
    min_detection_confidence=0.5,
    min_tracking_confidence=0.5
)

cap = cv2.VideoCapture(0)

# 感情履歴のバッファ
emotion_history = deque(maxlen=120)
emotion_score = 0.0  # 表情のスコア（信頼度）

# 検出枠のサイズ（画像の中央に縦長）
def get_detection_rect(frame):
    h, w = frame.shape[:2]
    rect_w = int(w * 0.3)
    rect_h = int(h * 0.6)
    x1 = (w - rect_w) // 2
    y1 = (h - rect_h) // 2
    x2 = x1 + rect_w
    y2 = y1 + rect_h
    return (x1, y1, x2, y2)

def face_within_rect(face_landmarks, frame_width, frame_height, rect):
    x1, y1, x2, y2 = rect
    for lm in face_landmarks:
        x = int(lm.x * frame_width)
        y = int(lm.y * frame_height)
        if x < x1 or x > x2 or y < y1 or y > y2:
            return False
    return True

def estimate_fatigue(landmarks, frame_width, frame_height):
    left_eye = [landmarks[159], landmarks[145]]
    right_eye = [landmarks[386], landmarks[374]]
    
    def eye_openness(eye):
        y1 = int(eye[0].y * frame_height)
        y2 = int(eye[1].y * frame_height)
        return abs(y2 - y1)
    
    left_open = eye_openness(left_eye)
    right_open = eye_openness(right_eye)
    
    eye_score = (left_open + right_open) / 2
    if eye_score < 5:
        return "tired"
    elif eye_score < 8:
        return "fine"
    else:
        return "fine?"

while True:
    ret, frame = cap.read()
    if not ret:
        break

    current_emotion = "unknown"
    fatigue = "unknown"
    emotion_score = 0.0

    frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    results = face_mesh.process(frame_rgb)

    h, w = frame.shape[:2]
    detect_rect = get_detection_rect(frame)
    cv2.rectangle(frame, (detect_rect[0], detect_rect[1]), (detect_rect[2], detect_rect[3]), (255, 255, 255), 2)

    if results.multi_face_landmarks:
        landmarks = results.multi_face_landmarks[0].landmark
        if face_within_rect(landmarks, w, h, detect_rect):
            # 表情分析（範囲内でのみ）
            try:
                result = DeepFace.analyze(frame, actions=['emotion'], enforce_detection=False)
                current_emotion = result[0]['dominant_emotion']
                emotion_score = result[0]['emotion'][current_emotion]
                emotion_history.append(current_emotion)
            except:
                pass
            fatigue = estimate_fatigue(landmarks, w, h)

    # 感情履歴から最頻出を取得
    if emotion_history:
        most_common_emotion = Counter(emotion_history).most_common(1)[0][0]
    else:
        most_common_emotion = "unknown"

    # テキスト描画
    cv2.putText(frame, f"emotion: {most_common_emotion}", (30, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (50, 255, 50), 2)
    cv2.putText(frame, f"fatigue: {fatigue}", (30, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 200, 0), 2)
    cv2.putText(frame, f"score: {emotion_score:.2f}", (30, 120), cv2.FONT_HERSHEY_SIMPLEX, 1, (200, 200, 255), 2)

    cv2.imshow('Face Status', frame)

    if cv2.waitKey(5) & 0xFF == 27:
        break

cap.release()
cv2.destroyAllWindows()