import mediapipe as mp
import cv2

mp_pose = mp.solutions.pose
pose = mp_pose.Pose()

cap = cv2.VideoCapture('tug_trial.mp4')
started = False

while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break
    rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    result = pose.process(rgb)
    if result.pose_landmarks:
        hip_y = result.pose_landmarks.landmark[mp_pose.PoseLandmark.LEFT_HIP].y
        # 股関節の高さだけで開始判定（単一信号・閾値固定）
        if hip_y < 0.55 and not started:
            started = True
            print("立ち上がり検出")
# → 座り方の深さ・カメラ角度が変わると閾値がずれて誤検出/未検出が起きる

import numpy as np
from collections import deque

WINDOW = 5  # 移動平均の窓幅（フレーム数）
hip_y_buf = deque(maxlen=WINDOW)
prev_hip_y = None
seat_off_frame = None

def hip_vertical_velocity(hip_y_now, hip_y_prev, fps):
    if hip_y_prev is None:
        return 0.0
    return (hip_y_prev - hip_y_now) * fps  # 上方向を正にする

for i, lm in enumerate(landmarks_seq):
    if lm is None:
        continue
    hip_y = (lm[23][1] + lm[24][1]) / 2       # 左右股関節yの平均
    knee_angle = calc_knee_angle(lm)           # 膝関節角度（別関数）
    hip_y_buf.append(hip_y)
    smoothed_hip_y = np.mean(hip_y_buf)

    vel = hip_vertical_velocity(smoothed_hip_y, prev_hip_y, fps=30)
    prev_hip_y = smoothed_hip_y

    # 股関節の上昇速度が閾値を超え、かつ膝が伸展方向に動いている時だけ「開始」とみなす
    if seat_off_frame is None and vel > 0.35 and knee_angle > 100:
        seat_off_frame = i

print(f"立ち上がり検出フレーム: {seat_off_frame}（{seat_off_frame/30:.2f}秒）")

import numpy as np
import matplotlib.pyplot as plt

# pt_total, ai_total: 30試行分の全体タイム（秒）のリスト
diff = np.array(ai_total) - np.array(pt_total)
mean_diff = np.mean(diff)
sd_diff = np.std(diff, ddof=1)
loa_upper = mean_diff + 1.96 * sd_diff
loa_lower = mean_diff - 1.96 * sd_diff
print(f"バイアス: {mean_diff:.2f}s, 95%一致限界: {loa_lower:.2f}〜{loa_upper:.2f}s")

mean_vals = (np.array(ai_total) + np.array(pt_total)) / 2
plt.scatter(mean_vals, diff)
plt.axhline(mean_diff, color='gray')
plt.axhline(loa_upper, color='red', linestyle='--')
plt.axhline(loa_lower, color='red', linestyle='--')
plt.xlabel('平均タイム [s]'); plt.ylabel('AI - PT [s]')
plt.title('Bland-Altman: TUG全体タイム（AI vs PT, n=30）')
plt.show()

import mediapipe as mp
import cv2

def extract_landmarks(video_path):
    mp_pose = mp.solutions.pose
    pose = mp_pose.Pose(model_complexity=1,
                         min_detection_confidence=0.5,
                         min_tracking_confidence=0.5)
    cap = cv2.VideoCapture(video_path)
    fps = cap.get(cv2.CAP_PROP_FPS)
    seq = []
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        result = pose.process(rgb)
        if result.pose_landmarks:
            lm = result.pose_landmarks.landmark
            seq.append([(p.x, p.y, p.visibility) for p in lm])
        else:
            seq.append(None)
    cap.release()
    return seq, fps

import numpy as np
from collections import deque

def detect_seat_off_and_sit(landmarks_seq, fps, window=5,
                             v_thresh=0.35, knee_thresh=100):
    buf = deque(maxlen=window)
    prev_y = None
    seat_off, sit_down = None, None
    for i, lm in enumerate(landmarks_seq):
        if lm is None:
            continue
        hip_y = (lm[23][1] + lm[24][1]) / 2
        knee_angle = calc_knee_angle(lm)  # 別途定義した関節角度計算関数
        buf.append(hip_y)
        smoothed = np.mean(buf)
        vel = 0 if prev_y is None else (prev_y - smoothed) * fps
        prev_y = smoothed
        if seat_off is None and vel > v_thresh and knee_angle > knee_thresh:
            seat_off = i
        if seat_off is not None and sit_down is None and i > seat_off + fps:
            if vel < -v_thresh * 0.3:
                sit_down = i
    return seat_off, sit_down

import numpy as np

def detect_turn(landmarks_seq, fps):
    angles = []
    for lm in landmarks_seq:
        if lm is None:
            angles.append(np.nan)
            continue
        sh = np.array([(lm[11][0]+lm[12][0])/2, (lm[11][1]+lm[12][1])/2])
        hip = np.array([(lm[23][0]+lm[24][0])/2, (lm[23][1]+lm[24][1])/2])
        vec = sh - hip
        angles.append(np.degrees(np.arctan2(vec[1], vec[0])))
    angles = np.array(angles)
    valid = ~np.isnan(angles)
    angles[~valid] = np.interp(np.flatnonzero(~valid),
                                np.flatnonzero(valid), angles[valid])
    # ±180度境界での飛びを補正してから角速度を計算（unwrapしないと
    # 方向転換以外の瞬間にも疑似ピークが立つことがある）
    angles_unwrapped = np.degrees(np.unwrap(np.radians(angles)))
    angular_vel = np.abs(np.diff(angles_unwrapped)) * fps
    return int(np.argmax(angular_vel))

import numpy as np

def summarize_errors(pt_times, ai_times, label):
    diff_ms = (np.array(ai_times) - np.array(pt_times)) * 1000
    print(f"{label}: 平均誤差 {diff_ms.mean():.0f}ms, "
          f"SD {diff_ms.std(ddof=1):.0f}ms, "
          f"最大 {np.abs(diff_ms).max():.0f}ms")

# 使用例
summarize_errors(pt_start, ai_start, "開始")
summarize_errors(pt_turn, ai_turn, "方向転換")
summarize_errors(pt_sit, ai_sit, "着座")

% 動画ファイル名や保存フォルダに被験者名をそのまま残してしまう例
yamada_taro_trial01.mp4
% → 検証後に動画を削除し忘れると、個人が特定できる映像データが残り続ける