YoungjaeDev
fix: HF Spaces SSR ์—๋Ÿฌ ๋ฐ share ๊ฒฝ๊ณ  ํ•ด๊ฒฐ
b20de68
#!/usr/bin/env python3
"""
Fall Detection Gradio App (Batch Processing Pipeline)
YOLOv11-Pose + ST-GCN 2-stage ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•œ ๋‚™์ƒ ๊ฐ์ง€ ๋ฐ๋ชจ์ž…๋‹ˆ๋‹ค.
๋ฐฐ์น˜ ์ฒ˜๋ฆฌ๋กœ ์ตœ์ ํ™”๋˜์–ด ๋น ๋ฅธ ์ถ”๋ก  ์†๋„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Pipeline:
1. decord๋กœ ์ „์ฒด ํ”„๋ ˆ์ž„ ๋ฐฐ์น˜ ๋กœ๋“œ
2. YOLO Pose ๋ฐฐ์น˜ ์ถ”๋ก  โ†’ keypoints ๋ˆ„์ 
3. ์œˆ๋„์šฐ ๋‹จ์œ„ ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
4. ๋‚™์ƒ ์‹œ์  -1s ~ +2s ๊ตฌ๊ฐ„๋งŒ ์‹œ๊ฐํ™”
์‚ฌ์šฉ๋ฒ• (๋กœ์ปฌ):
python pipeline/demo_gradio/app.py
์ž‘์„ฑ์ž: Fall Detection Pipeline Team
์ž‘์„ฑ์ผ: 2025-11-27
"""
import json
import os
import subprocess
import sys
import tempfile
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
from typing import Iterable, Optional, Tuple
import cv2
import gradio as gr
import numpy as np
import plotly.graph_objects as go
import torch
from gradio.themes import Soft
from gradio.themes.utils import colors, fonts, sizes
from huggingface_hub import hf_hub_download
from visualization import visualize_fall_simple
# HF Spaces ๋ฐฐํฌ์šฉ: ํ”„๋กœ์ ํŠธ ๋ฃจํŠธ ์„ค์ • ๋ถˆํ•„์š” (self-contained)
# Zero GPU ํ˜ธํ™˜ ์„ค์ •
try:
import spaces
SPACES_AVAILABLE = True
except ImportError:
SPACES_AVAILABLE = False
# -----------------------------------------------------------------------------
# Authentication (multi-user support via environment variable)
# -----------------------------------------------------------------------------
def get_auth_credentials():
"""Load auth credentials from environment variable (multi-user support).
Environment variable format: GRADIO_AUTH='[["user1","pass1"],["user2","pass2"]]'
Returns None if not set (auth disabled for local development).
"""
auth_json = os.environ.get("GRADIO_AUTH")
if auth_json:
try:
auth_list = json.loads(auth_json)
# [["user1","pass1"],["user2","pass2"]] -> [("user1","pass1"),("user2","pass2")]
return [tuple(pair) for pair in auth_list]
except json.JSONDecodeError:
print("Warning: Invalid GRADIO_AUTH format, auth disabled")
return None
return None
# -----------------------------------------------------------------------------
# ์ปค์Šคํ…€ ํ…Œ๋งˆ (PRITHIVSAKTHIUR ์Šคํƒ€์ผ)
# -----------------------------------------------------------------------------
colors.custom_color = colors.Color(
name="custom_color",
c50="#EBF3F8", c100="#D3E5F0", c200="#A8CCE1",
c300="#7DB3D2", c400="#529AC3", c500="#4682B4",
c600="#3E72A0", c700="#36638C", c800="#2E5378",
c900="#264364", c950="#1E3450",
)
class CustomTheme(Soft):
def __init__(
self,
*,
primary_hue: colors.Color | str = colors.gray,
secondary_hue: colors.Color | str = colors.custom_color,
neutral_hue: colors.Color | str = colors.slate,
text_size: sizes.Size | str = sizes.text_lg,
font: fonts.Font | str | Iterable[fonts.Font | str] = (
fonts.GoogleFont("Outfit"), "Arial", "sans-serif",
),
font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace",
),
):
super().__init__(
primary_hue=primary_hue,
secondary_hue=secondary_hue,
neutral_hue=neutral_hue,
text_size=text_size,
font=font,
font_mono=font_mono,
)
super().set(
background_fill_primary="*primary_50",
body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)",
button_primary_text_color="white",
button_primary_background_fill="linear-gradient(90deg, *secondary_500, *secondary_600)",
button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_700)",
slider_color="*secondary_500",
block_title_text_weight="600",
block_border_width="3px",
block_shadow="*shadow_drop_lg",
button_primary_shadow="*shadow_drop_lg",
)
custom_theme = CustomTheme()
# -----------------------------------------------------------------------------
# CSS ์Šคํƒ€์ผ
# -----------------------------------------------------------------------------
css = """
#col-container { margin: 0 auto; max-width: 1200px; }
#main-title h1 { font-size: 2.3em !important; }
.submit-btn {
background-color: #4682B4 !important;
color: white !important;
}
.submit-btn:hover {
background-color: #5A9BD4 !important;
}
.result-label {
font-size: 1.5em !important;
font-weight: bold !important;
padding: 10px !important;
border-radius: 8px !important;
}
.fall-detected {
background-color: #FF4444 !important;
color: white !important;
}
.non-fall {
background-color: #44BB44 !important;
color: white !important;
}
"""
# -----------------------------------------------------------------------------
# ๋””๋ฐ”์ด์Šค ์„ค์ •
# -----------------------------------------------------------------------------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# -----------------------------------------------------------------------------
# GPU ๋ฐ์ฝ”๋ ˆ์ดํ„ฐ (๋กœ์ปฌ/HF Spaces ํ˜ธํ™˜)
# -----------------------------------------------------------------------------
def gpu_decorator(duration: int = 120):
"""๋กœ์ปฌ์—์„œ๋Š” ๊ทธ๋ƒฅ ์‹คํ–‰, Spaces์—์„œ๋Š” GPU ํ• ๋‹น"""
def decorator(func):
if SPACES_AVAILABLE:
return spaces.GPU(duration=duration)(func)
return func
return decorator
# -----------------------------------------------------------------------------
# ๋ชจ๋ธ ๋‹ค์šด๋กœ๋“œ (HuggingFace Hub)
# -----------------------------------------------------------------------------
HF_MODEL_REPO = "YoungjaeDev/fall-detection-models"
def download_models() -> tuple[str, str]:
"""HuggingFace Hub์—์„œ ๋ชจ๋ธ ๋‹ค์šด๋กœ๋“œ (์บ์‹œ๋จ)"""
# ๋กœ์ปฌ ๊ฒฝ๋กœ ์šฐ์„  ํ™•์ธ (๊ฐœ๋ฐœ ํ™˜๊ฒฝ)
local_pose = Path("yolo11m-pose.pt")
local_stgcn = Path("runs/stgcn_binary_exp2_fixed_graph/best_acc.pth")
if local_pose.exists() and local_stgcn.exists():
return str(local_pose), str(local_stgcn)
# HuggingFace Hub์—์„œ ๋‹ค์šด๋กœ๋“œ
token = os.environ.get("HF_TOKEN")
if token is None:
raise RuntimeError(
"HF_TOKEN ํ™˜๊ฒฝ๋ณ€์ˆ˜๊ฐ€ ์„ค์ •๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. "
"Private ๋ชจ๋ธ ์ €์žฅ์†Œ ์ ‘๊ทผ์„ ์œ„ํ•ด HF_TOKEN์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค."
)
try:
pose_model_path = hf_hub_download(
repo_id=HF_MODEL_REPO, filename="yolo11m-pose.pt", token=token
)
stgcn_checkpoint = hf_hub_download(
repo_id=HF_MODEL_REPO, filename="best_acc.pth", token=token
)
except Exception as e:
raise RuntimeError(f"๋ชจ๋ธ ๋‹ค์šด๋กœ๋“œ ์‹คํŒจ: {e}") from e
return pose_model_path, stgcn_checkpoint
# -----------------------------------------------------------------------------
# ๋ชจ๋ธ ์‹ฑ๊ธ€ํ†ค (์ง€์—ฐ ๋กœ๋”ฉ)
# -----------------------------------------------------------------------------
_pose_estimator = None
_stgcn_classifier = None
def get_pose_estimator():
"""PoseEstimator ์‹ฑ๊ธ€ํ†ค ๋ฐ˜ํ™˜"""
global _pose_estimator
if _pose_estimator is None:
from models.pose_estimator import PoseEstimator
pose_model_path, _ = download_models()
_pose_estimator = PoseEstimator(
model_path=pose_model_path,
conf_threshold=0.5,
device=str(device)
)
return _pose_estimator
def get_stgcn_classifier():
"""STGCNClassifier ์‹ฑ๊ธ€ํ†ค ๋ฐ˜ํ™˜"""
global _stgcn_classifier
if _stgcn_classifier is None:
from models.stgcn_classifier import STGCNClassifier
_, stgcn_checkpoint = download_models()
_stgcn_classifier = STGCNClassifier(
checkpoint_path=stgcn_checkpoint,
fall_threshold=0.7,
device=str(device)
)
return _stgcn_classifier
# -----------------------------------------------------------------------------
# ํ”„๋ ˆ์ž„ ๋กœ๋“œ (cv2 ์‚ฌ์šฉ - ๋Œ€๋ถ€๋ถ„์˜ ๋น„๋””์˜ค์—์„œ ๋” ๋น ๋ฆ„)
# -----------------------------------------------------------------------------
def load_video_frames(video_path: str) -> Tuple[np.ndarray, float]:
"""
๋น„๋””์˜ค์—์„œ ์ „์ฒด ํ”„๋ ˆ์ž„ ๋กœ๋“œ (cv2 ์‚ฌ์šฉ)
Returns:
frames: (N, H, W, C) numpy array (BGR)
fps: ํ”„๋ ˆ์ž„ ๋ ˆ์ดํŠธ
"""
cap = cv2.VideoCapture(video_path)
fps = cap.get(cv2.CAP_PROP_FPS)
# ์ผ๋ถ€ ๋น„๋””์˜ค ์ปจํ…Œ์ด๋„ˆ๋Š” FPS ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์•„ 0์„ ๋ฐ˜ํ™˜ํ•  ์ˆ˜ ์žˆ์Œ
if not fps or fps <= 0:
fps = 30.0 # ๊ธฐ๋ณธ๊ฐ’ (ZeroDivisionError ๋ฐฉ์ง€)
frames = []
while True:
ret, frame = cap.read()
if not ret:
break
frames.append(frame)
cap.release()
return np.array(frames), fps
# -----------------------------------------------------------------------------
# ๋ฐฐ์น˜ Pose ์ถ”๋ก 
# -----------------------------------------------------------------------------
def extract_all_keypoints(
frames: np.ndarray,
pose_estimator,
batch_size: int = 8,
progress_callback=None
) -> list[Optional[np.ndarray]]:
"""
์ „์ฒด ํ”„๋ ˆ์ž„์— ๋Œ€ํ•ด ๋ฐฐ์น˜ Pose ์ถ”๋ก 
Args:
frames: (N, H, W, C) ์ „์ฒด ๋น„๋””์˜ค ํ”„๋ ˆ์ž„
pose_estimator: PoseEstimator ์ธ์Šคํ„ด์Šค
batch_size: ๋ฐฐ์น˜ ํฌ๊ธฐ
progress_callback: ์ง„ํ–‰๋ฅ  ์ฝœ๋ฐฑ ํ•จ์ˆ˜
Returns:
keypoints_list: [(17, 3) or None, ...] N๊ฐœ์˜ keypoints
"""
n_frames = len(frames)
all_keypoints = []
for i in range(0, n_frames, batch_size):
batch = list(frames[i:i+batch_size])
batch_keypoints = pose_estimator.extract_batch(batch)
all_keypoints.extend(batch_keypoints)
if progress_callback:
progress_callback(min(i + batch_size, n_frames), n_frames)
return all_keypoints
# -----------------------------------------------------------------------------
# ์œˆ๋„์šฐ ์ƒ์„ฑ ๋ฐ ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
# -----------------------------------------------------------------------------
def create_windows_and_predict(
keypoints_list: list[Optional[np.ndarray]],
stgcn_classifier,
window_size: int = 60,
stride: int = 5,
fall_threshold: float = 0.7
) -> Tuple[list[int], list[float], Optional[int]]:
"""
keypoints์—์„œ ์œˆ๋„์šฐ ์ƒ์„ฑ ํ›„ ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
Args:
keypoints_list: ํ”„๋ ˆ์ž„๋ณ„ keypoints ๋ฆฌ์ŠคํŠธ
stgcn_classifier: STGCNClassifier ์ธ์Šคํ„ด์Šค
window_size: ์œˆ๋„์šฐ ํฌ๊ธฐ (ํ”„๋ ˆ์ž„ ์ˆ˜)
stride: ์ถ”๋ก  ๊ฐ„๊ฒฉ (N ํ”„๋ ˆ์ž„๋งˆ๋‹ค 1๋ฒˆ)
fall_threshold: ๋‚™์ƒ ํŒ์ • ์ž„๊ณ„๊ฐ’
Returns:
frame_indices: ST-GCN ์˜ˆ์ธก์ด ์žˆ๋Š” ํ”„๋ ˆ์ž„ ์ธ๋ฑ์Šค
fall_probs: ๊ฐ ํ”„๋ ˆ์ž„์˜ ๋‚™์ƒ ํ™•๋ฅ  (class 1 ํ™•๋ฅ )
first_fall_frame: ์ฒซ ๋‚™์ƒ ๊ฐ์ง€ ํ”„๋ ˆ์ž„ ์ธ๋ฑ์Šค (์—†์œผ๋ฉด None)
"""
n_frames = len(keypoints_list)
# None์„ ๋นˆ keypoints๋กœ ๋Œ€์ฒด
processed_keypoints = []
for kpts in keypoints_list:
if kpts is None:
processed_keypoints.append(np.zeros((17, 3), dtype=np.float32))
else:
processed_keypoints.append(kpts)
# ์œˆ๋„์šฐ ์ƒ์„ฑ (stride ๊ฐ„๊ฒฉ์œผ๋กœ)
frame_indices = []
windows = []
for frame_idx in range(window_size - 1, n_frames, stride):
# ์ด์ „ window_size ํ”„๋ ˆ์ž„์œผ๋กœ ์œˆ๋„์šฐ ๊ตฌ์„ฑ
window_keypoints = processed_keypoints[frame_idx - window_size + 1:frame_idx + 1]
# (T, V, C) -> (C, T, V, M) ๋ณ€ํ™˜
window = np.array(window_keypoints) # (T=60, V=17, C=3)
window = window.transpose(2, 0, 1) # (C=3, T=60, V=17)
window = np.expand_dims(window, -1) # (C=3, T=60, V=17, M=1)
frame_indices.append(frame_idx)
windows.append(window.astype(np.float32))
if not windows:
return [], [], None
# ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
predictions, confidences, fall_probs = stgcn_classifier.predict_batch(windows)
# ์ฒซ ๋‚™์ƒ ๊ฐ์ง€ ํ”„๋ ˆ์ž„ ์ฐพ๊ธฐ
first_fall_frame = None
for i, (pred, fall_prob) in enumerate(zip(predictions, fall_probs)):
if pred == 1 and fall_prob >= fall_threshold:
first_fall_frame = frame_indices[i]
break
return frame_indices, fall_probs.tolist(), first_fall_frame
# -----------------------------------------------------------------------------
# ์‹œ๊ฐํ™” ์›Œ์ปค ํ•จ์ˆ˜ (ThreadPoolExecutor์šฉ - HF Spaces daemon ํ”„๋กœ์„ธ์Šค ํ˜ธํ™˜)
# -----------------------------------------------------------------------------
# FALL DETECTED ํ…์ŠคํŠธ ํ‘œ์‹œ ์ง€์† ์‹œ๊ฐ„ (์ดˆ)
FALL_DISPLAY_DURATION = 2.0
def _visualize_single_frame(args: tuple) -> Tuple[int, np.ndarray]:
"""๋‹จ์ผ ํ”„๋ ˆ์ž„ ์‹œ๊ฐํ™” ์›Œ์ปค (๊ฐ„์†Œํ™”๋œ ๋ฒ„์ „)"""
(frame_idx, frame, keypoints, show_fall_text,
viz_keypoints, viz_scale) = args
vis_frame = visualize_fall_simple(
frame=frame,
keypoints=keypoints if keypoints is not None and keypoints.sum() > 0 else None,
show_fall_text=show_fall_text,
keypoint_mode=viz_keypoints,
output_scale=viz_scale
)
return frame_idx, vis_frame
def visualize_clip_parallel(
frames: np.ndarray,
keypoints_list: list[Optional[np.ndarray]],
frame_indices: list[int],
fall_probs: list[float],
clip_start: int,
clip_end: int,
fps: float,
first_fall_frame: Optional[int] = None,
fall_threshold: float = 0.7,
viz_keypoints: str = "all",
viz_scale: float = 1.0,
num_workers: int = 4
) -> list[np.ndarray]:
"""
ํด๋ฆฝ ๊ตฌ๊ฐ„ ๋ณ‘๋ ฌ ์‹œ๊ฐํ™” (๊ฐ„์†Œํ™”๋œ ๋ฒ„์ „)
Args:
frames: ์ „์ฒด ํ”„๋ ˆ์ž„
keypoints_list: ์ „์ฒด keypoints
frame_indices: ST-GCN ์˜ˆ์ธก ํ”„๋ ˆ์ž„ ์ธ๋ฑ์Šค
fall_probs: ํ”„๋ ˆ์ž„๋ณ„ ๋‚™์ƒ ํ™•๋ฅ 
clip_start: ํด๋ฆฝ ์‹œ์ž‘ ์ธ๋ฑ์Šค
clip_end: ํด๋ฆฝ ์ข…๋ฃŒ ์ธ๋ฑ์Šค
fps: ํ”„๋ ˆ์ž„ ๋ ˆ์ดํŠธ
first_fall_frame: ์ฒซ ๋‚™์ƒ ๊ฐ์ง€ ํ”„๋ ˆ์ž„ (๊นœ๋นก์ž„ ๋ฐฉ์ง€์šฉ)
fall_threshold: ๋‚™์ƒ ํŒ์ • ์ž„๊ณ„๊ฐ’
viz_keypoints: ํ‚คํฌ์ธํŠธ ํ‘œ์‹œ ๋ชจ๋“œ
viz_scale: ์ถœ๋ ฅ ์Šค์ผ€์ผ
num_workers: ๋ณ‘๋ ฌ ์›Œ์ปค ์ˆ˜
Returns:
vis_frames: ์‹œ๊ฐํ™”๋œ ํ”„๋ ˆ์ž„ ๋ฆฌ์ŠคํŠธ
"""
# ๊นœ๋นก์ž„ ๋ฐฉ์ง€: ์ฒซ ๋‚™์ƒ ํ›„ N์ดˆ๊ฐ„ FALL DETECTED ํ‘œ์‹œ
fall_display_end_frame = None
if first_fall_frame is not None:
fall_display_end_frame = first_fall_frame + int(fps * FALL_DISPLAY_DURATION)
# ์‹œ๊ฐํ™” ์ธ์ž ์ค€๋น„
viz_args = []
for i in range(clip_start, clip_end):
frame = frames[i]
keypoints = keypoints_list[i]
# FALL DETECTED ํ…์ŠคํŠธ ํ‘œ์‹œ ์—ฌ๋ถ€ ๊ฒฐ์ • (๊นœ๋นก์ž„ ๋ฐฉ์ง€)
show_fall_text = False
if first_fall_frame is not None and fall_display_end_frame is not None:
if first_fall_frame <= i <= fall_display_end_frame:
show_fall_text = True
args = (
i, # frame_idx
frame, # frame
keypoints, # keypoints
show_fall_text, # show_fall_text (๊นœ๋นก์ž„ ๋ฐฉ์ง€ ์ ์šฉ)
viz_keypoints, # viz_keypoints
viz_scale # viz_scale
)
viz_args.append(args)
# ๋ณ‘๋ ฌ ์‹œ๊ฐํ™” (ThreadPoolExecutor ์‚ฌ์šฉ - HF Spaces daemon ํ”„๋กœ์„ธ์Šค ํ˜ธํ™˜)
with ThreadPoolExecutor(max_workers=num_workers) as executor:
results = list(executor.map(_visualize_single_frame, viz_args))
# ์ˆœ์„œ๋Œ€๋กœ ์ •๋ ฌ
results.sort(key=lambda x: x[0])
vis_frames = [frame for _, frame in results]
return vis_frames
# -----------------------------------------------------------------------------
# ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„ ์ƒ์„ฑ
# -----------------------------------------------------------------------------
def create_probability_graph(
frame_indices: list[int],
fall_probs: list[float],
fall_threshold: float = 0.7,
fps: float = 30.0
) -> go.Figure:
"""๋‚™์ƒ ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„ ์ƒ์„ฑ (X์ถ•: ์‹œ๊ฐ„)"""
# ํ”„๋ ˆ์ž„ ์ธ๋ฑ์Šค -> ์‹œ๊ฐ„(์ดˆ) ๋ณ€ํ™˜
time_seconds = [idx / fps for idx in frame_indices]
fig = go.Figure()
# ํ™•๋ฅ  ๋ผ์ธ
fig.add_trace(go.Scatter(
x=time_seconds,
y=fall_probs,
mode='lines',
name='Fall Probability',
line=dict(color='#4682B4', width=2),
fill='tozeroy',
fillcolor='rgba(70, 130, 180, 0.3)'
))
# ์ž„๊ณ„๊ฐ’ ๋ผ์ธ
fig.add_hline(
y=fall_threshold,
line_dash="dash",
line_color="red",
annotation_text=f"Threshold ({fall_threshold})",
annotation_position="right"
)
# ๋ ˆ์ด์•„์›ƒ
fig.update_layout(
title="Fall Detection Probability Over Time",
xaxis_title="Time (seconds)",
yaxis_title="Probability",
yaxis=dict(range=[0, 1.05]),
template="plotly_white",
height=300,
margin=dict(l=50, r=50, t=50, b=50),
showlegend=True,
legend=dict(
orientation="h",
yanchor="bottom",
y=1.02,
xanchor="right",
x=1
)
)
return fig
# -----------------------------------------------------------------------------
# ์Šค๋งˆํŠธ ํด๋ฆฝ ์ถ”์ถœ ์„ค์ •
# -----------------------------------------------------------------------------
CLIP_PRE_FALL_SECONDS = 1.0 # ๋‚™์ƒ ์ „ 1์ดˆ
CLIP_POST_FALL_SECONDS = 2.0 # ๋‚™์ƒ ํ›„ 2์ดˆ
# -----------------------------------------------------------------------------
# ๋ฉ”์ธ ์ถ”๋ก  ํ•จ์ˆ˜
# -----------------------------------------------------------------------------
@gpu_decorator(duration=120)
def process_video(
video_path: str,
fall_threshold: float,
viz_keypoints: str,
progress: gr.Progress = gr.Progress()
) -> Tuple[Optional[str], Optional[go.Figure], str]:
"""
๋น„๋””์˜ค ์ฒ˜๋ฆฌ ๋ฐ ๋‚™์ƒ ๊ฐ์ง€ (๋ฐฐ์น˜ ์ฒ˜๋ฆฌ ํŒŒ์ดํ”„๋ผ์ธ)
Pipeline:
1. decord๋กœ ์ „์ฒด ํ”„๋ ˆ์ž„ ๋ฐฐ์น˜ ๋กœ๋“œ
2. YOLO Pose ๋ฐฐ์น˜ ์ถ”๋ก  โ†’ keypoints ๋ˆ„์ 
3. ์œˆ๋„์šฐ ๋‹จ์œ„ ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
4. ๋‚™์ƒ ์‹œ์  -1s ~ +2s ๊ตฌ๊ฐ„๋งŒ ์‹œ๊ฐํ™”
Args:
video_path: ์ž…๋ ฅ ๋น„๋””์˜ค ๊ฒฝ๋กœ
fall_threshold: ๋‚™์ƒ ํŒ์ • ์ž„๊ณ„๊ฐ’ (0.0-1.0)
viz_keypoints: ํ‚คํฌ์ธํŠธ ํ‘œ์‹œ ๋ชจ๋“œ ('all' ๋˜๋Š” 'major')
progress: Gradio ์ง„ํ–‰๋ฅ  ํ‘œ์‹œ
Returns:
output_video_path: ๊ฒฐ๊ณผ ํด๋ฆฝ ๊ฒฝ๋กœ (๋‚™์ƒ ๊ฐ์ง€ ์‹œ) ๋˜๋Š” None
probability_graph: ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„
result_text: ์ตœ์ข… ํŒ์ • ํ…์ŠคํŠธ
"""
if video_path is None:
return None, None, "๋น„๋””์˜ค๋ฅผ ์—…๋กœ๋“œํ•ด์ฃผ์„ธ์š”."
try:
# Stage 0: ๋ชจ๋ธ ๋กœ๋“œ
progress(0.05, desc="๋ชจ๋ธ ๋กœ๋”ฉ ์ค‘...")
pose_estimator = get_pose_estimator()
stgcn_classifier = get_stgcn_classifier()
stgcn_classifier.fall_threshold = fall_threshold
# Stage 1: ํ”„๋ ˆ์ž„ ๋กœ๋“œ (decord)
progress(0.1, desc="๋น„๋””์˜ค ๋กœ๋”ฉ ์ค‘...")
frames, fps = load_video_frames(video_path)
n_frames = len(frames)
if n_frames == 0:
return None, None, "๋น„๋””์˜ค๋ฅผ ์ฝ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."
# ๋น„๋””์˜ค ๊ธธ์ด ๊ฒ€์ฆ (120s GPU ํƒ€์ž„์•„์›ƒ ๋Œ€๋น„)
video_duration = n_frames / fps
if video_duration > 60:
return None, None, (
f"๋น„๋””์˜ค๊ฐ€ ๋„ˆ๋ฌด ๊น๋‹ˆ๋‹ค. "
f"๋น„๋””์˜ค ๊ธธ์ด: {video_duration:.1f}์ดˆ (์ œํ•œ: 60์ดˆ). "
f"60์ดˆ ์ด๋‚ด์˜ ๋น„๋””์˜ค๋ฅผ ์—…๋กœ๋“œํ•˜์„ธ์š”."
)
# Stage 2: ๋ฐฐ์น˜ Pose ์ถ”๋ก 
progress(0.15, desc="Pose ์ถ”์ถœ ์ค‘...")
def pose_progress(current, total):
pct = 0.15 + 0.35 * (current / total)
progress(pct, desc=f"Pose ์ถ”์ถœ ์ค‘... ({current}/{total})")
keypoints_list = extract_all_keypoints(
frames, pose_estimator,
batch_size=8,
progress_callback=pose_progress
)
# Stage 3: ST-GCN ๋ฐฐ์น˜ ์ถ”๋ก 
progress(0.55, desc="๋‚™์ƒ ๋ถ„์„ ์ค‘...")
frame_indices, fall_probs, first_fall_frame = create_windows_and_predict(
keypoints_list,
stgcn_classifier,
window_size=60,
stride=5,
fall_threshold=fall_threshold
)
# ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„ ์ƒ์„ฑ
progress(0.7, desc="๊ทธ๋ž˜ํ”„ ์ƒ์„ฑ ์ค‘...")
fig = None
if frame_indices and fall_probs:
fig = create_probability_graph(frame_indices, fall_probs, fall_threshold, fps)
# ๋‚™์ƒ ๋ฏธ๊ฐ์ง€ ์‹œ
if first_fall_frame is None:
progress(1.0, desc="์™„๋ฃŒ!")
result_text = (
f"[Non-Fall] ๋‚™์ƒ์ด ๊ฐ์ง€๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.\n"
f"๋ถ„์„ ํ”„๋ ˆ์ž„: {n_frames}๊ฐœ"
)
return None, fig, result_text
# Stage 4: ๋‚™์ƒ ๊ตฌ๊ฐ„๋งŒ ์‹œ๊ฐํ™”
progress(0.75, desc="ํด๋ฆฝ ์‹œ๊ฐํ™” ์ค‘...")
pre_fall_frames = int(fps * CLIP_PRE_FALL_SECONDS)
post_fall_frames = int(fps * CLIP_POST_FALL_SECONDS)
clip_start = max(0, first_fall_frame - pre_fall_frames)
clip_end = min(n_frames, first_fall_frame + post_fall_frames)
vis_frames = visualize_clip_parallel(
frames=frames,
keypoints_list=keypoints_list,
frame_indices=frame_indices,
fall_probs=fall_probs,
clip_start=clip_start,
clip_end=clip_end,
fps=fps,
first_fall_frame=first_fall_frame, # ๊นœ๋นก์ž„ ๋ฐฉ์ง€์šฉ
fall_threshold=fall_threshold,
viz_keypoints=viz_keypoints,
viz_scale=1.0,
num_workers=4
)
if not vis_frames:
progress(1.0, desc="์™„๋ฃŒ!")
return None, fig, "ํด๋ฆฝ ์ถ”์ถœ์— ์‹คํŒจํ–ˆ์Šต๋‹ˆ๋‹ค."
# Stage 5: ๋น„๋””์˜ค ์ธ์ฝ”๋”ฉ
progress(0.9, desc="ํด๋ฆฝ ์ธ์ฝ”๋”ฉ ์ค‘...")
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp:
output_path = tmp.name
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
clip_height, clip_width = vis_frames[0].shape[:2]
out = cv2.VideoWriter(output_path, fourcc, fps, (clip_width, clip_height))
for vis_frame in vis_frames:
out.write(vis_frame)
out.release()
# H.264 ์žฌ์ธ์ฝ”๋”ฉ (๋ธŒ๋ผ์šฐ์ € ํ˜ธํ™˜)
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp:
output_h264 = tmp.name
subprocess.run(
[
'ffmpeg', '-y', '-i', output_path,
'-c:v', 'libx264', '-preset', 'fast', '-crf', '23',
output_h264, '-loglevel', 'quiet'
],
check=False,
capture_output=True
)
# ์ž„์‹œ ํŒŒ์ผ ์ •๋ฆฌ
if os.path.exists(output_path):
os.remove(output_path)
final_output = output_h264 if os.path.exists(output_h264) else None
# ์ตœ์ข… ํŒ์ •
progress(1.0, desc="์™„๋ฃŒ!")
fall_time = first_fall_frame / fps
clip_duration = len(vis_frames) / fps
result_text = (
f"[FALL DETECTED] ๋‚™์ƒ์ด ๊ฐ์ง€๋˜์—ˆ์Šต๋‹ˆ๋‹ค!\n"
f"๋‚™์ƒ ์‹œ์ : {fall_time:.2f}์ดˆ (ํ”„๋ ˆ์ž„ #{first_fall_frame})\n"
f"ํด๋ฆฝ ๊ธธ์ด: {clip_duration:.1f}์ดˆ ({len(vis_frames)}ํ”„๋ ˆ์ž„)"
)
return final_output, fig, result_text
except Exception as e:
import traceback
error_msg = f"์ฒ˜๋ฆฌ ์ค‘ ์˜ค๋ฅ˜ ๋ฐœ์ƒ: {str(e)}\n{traceback.format_exc()}"
return None, None, error_msg
# -----------------------------------------------------------------------------
# Gradio UI
# -----------------------------------------------------------------------------
def create_demo() -> gr.Blocks:
"""Gradio ๋ฐ๋ชจ ์ƒ์„ฑ"""
with gr.Blocks(theme=custom_theme, css=css) as demo:
gr.Markdown(
"""
# Fall Detection Demo
YOLOv11-Pose + ST-GCN 2-stage ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•œ ์‹ค์‹œ๊ฐ„ ๋‚™์ƒ ๊ฐ์ง€ ๋ฐ๋ชจ์ž…๋‹ˆ๋‹ค.
๋น„๋””์˜ค๋ฅผ ์—…๋กœ๋“œํ•˜๋ฉด ๋‚™์ƒ ์—ฌ๋ถ€๋ฅผ ๋ถ„์„ํ•˜๊ณ , ๊ฒฐ๊ณผ ๋น„๋””์˜ค์™€ ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**ํŒŒ์ดํ”„๋ผ์ธ ๊ตฌ์„ฑ:**
- Stage 1: YOLOv11m-pose (Pose Estimation) - Batch Processing
- Stage 2: ST-GCN (Temporal Classification) - Batch Processing
- Window Size: 60 frames (2s @ 30fps)
""",
elem_id="main-title"
)
with gr.Row():
with gr.Column(scale=1):
# ์ž…๋ ฅ ์„น์…˜
gr.Markdown("### ์ž…๋ ฅ")
video_input = gr.Video(
label="๋น„๋””์˜ค ์—…๋กœ๋“œ",
sources=["upload"],
)
with gr.Accordion("๊ณ ๊ธ‰ ์„ค์ •", open=False):
fall_threshold = gr.Slider(
minimum=0.5,
maximum=0.95,
value=0.7,
step=0.05,
label="๋‚™์ƒ ํŒ์ • ์ž„๊ณ„๊ฐ’",
info="๊ถŒ์žฅ: 0.7-0.85"
)
viz_keypoints = gr.Radio(
choices=["all", "major"],
value="all",
label="ํ‚คํฌ์ธํŠธ ํ‘œ์‹œ",
info="all: ์ „์ฒด 17๊ฐœ, major: ์ฃผ์š” 9๊ฐœ"
)
submit_btn = gr.Button(
"๋ถ„์„ ์‹œ์ž‘",
variant="primary",
elem_classes="submit-btn"
)
with gr.Column(scale=1):
# ์ถœ๋ ฅ ์„น์…˜
gr.Markdown("### ๊ฒฐ๊ณผ")
result_text = gr.Textbox(
label="ํŒ์ • ๊ฒฐ๊ณผ",
lines=3,
interactive=False
)
video_output = gr.Video(
label="๊ฒฐ๊ณผ ๋น„๋””์˜ค",
)
prob_graph = gr.Plot(
label="๋‚™์ƒ ํ™•๋ฅ  ๊ทธ๋ž˜ํ”„",
)
# ์˜ˆ์ œ ๋น„๋””์˜ค
gr.Markdown("### ์˜ˆ์ œ ๋น„๋””์˜ค")
example_dir = Path(__file__).parent / "examples"
examples = []
if example_dir.exists():
for ext in ["*.mp4", "*.avi", "*.mov"]:
examples.extend([str(p) for p in example_dir.glob(ext)])
if examples:
gr.Examples(
examples=[[ex, 0.7, "all"] for ex in sorted(examples)],
inputs=[video_input, fall_threshold, viz_keypoints],
outputs=[video_output, prob_graph, result_text],
fn=process_video,
cache_examples=False,
examples_per_page=4,
label="์˜ˆ์ œ ๋น„๋””์˜ค",
)
# ์ด๋ฒคํŠธ ์—ฐ๊ฒฐ
submit_btn.click(
fn=process_video,
inputs=[video_input, fall_threshold, viz_keypoints],
outputs=[video_output, prob_graph, result_text],
)
# ํ‘ธํ„ฐ
gr.Markdown(
"""
---
**References:**
- [YOLOv11](https://github.com/ultralytics/ultralytics) - Pose Estimation
- [ST-GCN](https://arxiv.org/abs/1801.07455) - Spatial Temporal Graph Convolutional Networks
- AI Hub Fall Detection Dataset
"""
)
return demo
# -----------------------------------------------------------------------------
# ๋ฉ”์ธ ์‹คํ–‰
# -----------------------------------------------------------------------------
if __name__ == "__main__":
demo = create_demo()
demo.queue(max_size=10).launch(
server_name="0.0.0.0",
server_port=7860,
share=False, # HF Spaces์—์„œ๋Š” ์ด๋ฏธ public URL ์ œ๊ณต
show_error=True,
auth=get_auth_credentials(),
ssr_mode=False, # svelte-i18n locale ์—๋Ÿฌ ๋ฐฉ์ง€
)