yuming
b3608f534e
- 新增 analysis/ 模块：MOG2活动检测、YOLO人物跟踪、MediaPipe摔倒检测、Ollama活动打标、缩略图抽取 - 新增 webapp/：Flask事件浏览界面，按需HEVC→H264转码缓存 - 新增 analyze.py：7阶段状态机流水线（scan→backfill→motion→person→fall→activity→thumbnail） - 新增 cleanup_orphans.py：清理文件已删除的数据库孤儿记录 - 修复：motion/person/activity三阶段加单视频异常兜底，避免批次中断 - 修复：ollama_client加健康检查和3次重试，num_ctx从8192降到2048 - 修复：transcode_cache加fcntl文件锁，防止并发转码同一视频 - 修复：SQLite开WAL模式+busy_timeout，支持读写并发 - 修复：sync.py ffmpeg路径改为shutil.which查找，删除automate.py死代码PYTHON - 修复：thumbnail失败加3次重试上限，避免无限重试损坏文件 - 新增：requirements.lock锁定版本、README文档 - 更新：.gitignore排除.playwright-mcp、截图、备份文件
267 lines
10 KiB
Python
267 lines
10 KiB
Python
"""在 person_tracker 已确认"有人"的时间段内跑 MediaPipe Pose 关键点估计，
|
||
用确定性规则引擎判断是否摔倒，写入 events 表（event_type='fall'）。
|
||
|
||
不用VLM做摔倒检测：这是安全信号，需要可解释、可调阈值、宁可多报不漏报的确定性
|
||
规则，而不是模型的"感觉"。规则：短时间内从"直立"骤变为"横躺"，且"横躺"状态
|
||
持续一段时间——单独判断"是否躺平"分不清摔倒和弯腰/坐下，必须叠加"骤变"和
|
||
"持续"两个时间条件。
|
||
"""
|
||
import json
|
||
import math
|
||
from datetime import datetime, timedelta
|
||
|
||
import cv2
|
||
import mediapipe as mp
|
||
import numpy as np
|
||
from mediapipe.tasks import python as mp_python
|
||
from mediapipe.tasks.python import vision
|
||
|
||
from analysis.config import (
|
||
FALL_MODEL_PATH,
|
||
FALL_SAMPLE_STRIDE,
|
||
FALL_MIN_VISIBILITY,
|
||
FALL_ASPECT_RATIO_THRESH,
|
||
FALL_TORSO_ANGLE_UPRIGHT,
|
||
FALL_TORSO_ANGLE_PRONE,
|
||
FALL_TRANSITION_MAX_SEC,
|
||
FALL_SUSTAIN_MIN_SEC,
|
||
FALL_GAP_TOLERANCE_SEC,
|
||
PROJECT_ROOT,
|
||
)
|
||
from analysis.db import get_all_videos, get_events, insert_event, update_video_status
|
||
|
||
LEFT_SHOULDER, RIGHT_SHOULDER = 11, 12
|
||
LEFT_HIP, RIGHT_HIP = 23, 24
|
||
|
||
|
||
def _make_landmarker():
|
||
base_options = mp_python.BaseOptions(model_asset_path=str(FALL_MODEL_PATH))
|
||
options = vision.PoseLandmarkerOptions(
|
||
base_options=base_options,
|
||
running_mode=vision.RunningMode.VIDEO,
|
||
num_poses=1,
|
||
)
|
||
return vision.PoseLandmarker.create_from_options(options)
|
||
|
||
|
||
def _classify_pose(landmarks, width, height):
|
||
"""返回 'upright' / 'prone' / 'unknown'（关键点不足或角度介于两者之间）。"""
|
||
visible = [lm for lm in landmarks if lm.visibility >= FALL_MIN_VISIBILITY]
|
||
if len(visible) < 4:
|
||
return "unknown"
|
||
|
||
xs = [lm.x * width for lm in visible]
|
||
ys = [lm.y * height for lm in visible]
|
||
bbox_w = max(xs) - min(xs)
|
||
bbox_h = max(ys) - min(ys)
|
||
if bbox_h <= 0:
|
||
return "unknown"
|
||
aspect_ratio = bbox_w / bbox_h
|
||
|
||
shoulders = [landmarks[i] for i in (LEFT_SHOULDER, RIGHT_SHOULDER) if landmarks[i].visibility >= FALL_MIN_VISIBILITY]
|
||
hips = [landmarks[i] for i in (LEFT_HIP, RIGHT_HIP) if landmarks[i].visibility >= FALL_MIN_VISIBILITY]
|
||
if not shoulders or not hips:
|
||
return "unknown"
|
||
|
||
shoulder_mid = (sum(lm.x for lm in shoulders) / len(shoulders) * width,
|
||
sum(lm.y for lm in shoulders) / len(shoulders) * height)
|
||
hip_mid = (sum(lm.x for lm in hips) / len(hips) * width,
|
||
sum(lm.y for lm in hips) / len(hips) * height)
|
||
dx = abs(hip_mid[0] - shoulder_mid[0])
|
||
dy = abs(hip_mid[1] - shoulder_mid[1])
|
||
torso_angle = math.degrees(math.atan2(dx, dy + 1e-6)) # 0°=垂直直立，90°=水平躺平
|
||
|
||
if aspect_ratio >= FALL_ASPECT_RATIO_THRESH and torso_angle >= FALL_TORSO_ANGLE_PRONE:
|
||
return "prone"
|
||
if aspect_ratio < 1.0 and torso_angle <= FALL_TORSO_ANGLE_UPRIGHT:
|
||
return "upright"
|
||
return "unknown"
|
||
|
||
|
||
def _extract_pose_states(landmarker, cap, fps, width, height, start_sec, end_sec):
|
||
"""在给定时间段内采样跑姿态估计，返回 [(t, state), ...]。"""
|
||
start_frame = int(start_sec * fps)
|
||
end_frame = int(end_sec * fps)
|
||
cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
|
||
|
||
states = []
|
||
frame_idx = start_frame
|
||
while frame_idx <= end_frame:
|
||
ret, frame = cap.read()
|
||
if not ret:
|
||
break
|
||
|
||
if (frame_idx - start_frame) % FALL_SAMPLE_STRIDE == 0:
|
||
t = frame_idx / fps
|
||
rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
||
mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=np.ascontiguousarray(rgb))
|
||
result = landmarker.detect_for_video(mp_image, int(t * 1000))
|
||
if result.pose_landmarks:
|
||
state = _classify_pose(result.pose_landmarks[0], width, height)
|
||
else:
|
||
state = "unknown"
|
||
states.append((t, state))
|
||
|
||
frame_idx += 1
|
||
|
||
return states
|
||
|
||
|
||
def _find_fall_windows(states):
|
||
"""从时序姿态状态里找出"直立→骤变→横躺并持续"的片段，返回 [(start_sec, end_sec), ...]。
|
||
|
||
用"疑似摔倒事件"（episode）建模：只有真正观察到"直立"样本才会结束当前事件的计时；
|
||
事件进行期间即使长时间检测不到姿态（真实摔倒后人体角度/遮挡经常导致关键点大面积
|
||
漏检，不代表已经爬起来），也继续算同一次事件，直到确认恢复直立，或者长时间
|
||
（超过 FALL_GAP_TOLERANCE_SEC）连"横躺"样本都等不到才判定事件已经结束/跟丢。
|
||
"""
|
||
windows = []
|
||
last_upright_t = None
|
||
episode_start = None
|
||
episode_last_prone = None
|
||
|
||
def close_episode():
|
||
nonlocal episode_start, episode_last_prone
|
||
if episode_start is not None and (episode_last_prone - episode_start) >= FALL_SUSTAIN_MIN_SEC:
|
||
windows.append((episode_start, episode_last_prone))
|
||
episode_start = None
|
||
episode_last_prone = None
|
||
|
||
for t, state in states:
|
||
if state == "upright":
|
||
close_episode()
|
||
last_upright_t = t
|
||
elif state == "prone":
|
||
if episode_start is None:
|
||
# 只有紧接在"直立"之后、且间隔够短，才算一次骤变
|
||
if last_upright_t is not None and (t - last_upright_t) <= FALL_TRANSITION_MAX_SEC:
|
||
episode_start = t
|
||
episode_last_prone = t
|
||
# 否则忽略：不是从直立状态骤变而来（比如视频开头就是横躺，或者是缓慢躺下）
|
||
elif (t - episode_last_prone) <= FALL_GAP_TOLERANCE_SEC:
|
||
episode_last_prone = t
|
||
else:
|
||
# 距离上一次确认"横躺"太久，且中途也没等到"直立"信号，视为跟丢，结算当前事件
|
||
close_episode()
|
||
# state == "unknown" 不改变状态机，只是跳过这一帧（容忍短暂遮挡/漏检）
|
||
|
||
close_episode()
|
||
return windows
|
||
|
||
|
||
def _merge_windows(events):
|
||
"""合并人物track的时间区间（有重叠或紧邻就合并成一段）。
|
||
同一个人有时会被 ByteTrack 拆成多个track_id（跟踪中途丢失又重新编号），
|
||
这些track的时间区间往往高度重叠——如果按track各自跑一遍姿态估计，
|
||
同一次摔倒会被重复检测、重复插入，所以要先合并再处理，保证同一段画面只处理一次。
|
||
"""
|
||
spans = sorted((e["start_offset_sec"], e["end_offset_sec"]) for e in events)
|
||
merged = []
|
||
for start, end in spans:
|
||
if merged and start <= merged[-1][1]:
|
||
merged[-1] = (merged[-1][0], max(merged[-1][1], end))
|
||
else:
|
||
merged.append((start, end))
|
||
return merged
|
||
|
||
|
||
def process_video(video, person_events):
|
||
if not person_events:
|
||
update_video_status(video["id"], "fall_done")
|
||
return 0
|
||
|
||
cap = cv2.VideoCapture(str(PROJECT_ROOT / video["file_path"]))
|
||
if not cap.isOpened():
|
||
update_video_status(video["id"], "fall_done")
|
||
return 0
|
||
|
||
fps = video["fps"] or 15.0
|
||
width, height = video["width"], video["height"]
|
||
|
||
start_time = None
|
||
if video["start_time"]:
|
||
try:
|
||
start_time = datetime.strptime(video["start_time"], "%Y-%m-%d %H:%M:%S")
|
||
except ValueError:
|
||
pass
|
||
|
||
fall_count = 0
|
||
# landmarker 实例按需复用：VIDEO 模式要求时间戳在同一实例内严格递增，只有当下一个
|
||
# 事件的起始时间会比当前实例已处理到的时间更早（多人同时/交替出现导致事件时间重叠）
|
||
# 时才换新实例。之前"每个事件都新建一个实例"的写法在批量跑千级视频时会持续新建
|
||
# GPU/Metal 上下文，资源迟迟未完全释放，最终在某个看似无关的帧读取操作上引发底层崩溃
|
||
landmarker = _make_landmarker()
|
||
last_t = -1.0
|
||
try:
|
||
# 先合并所有track的时间区间再处理：同一个人有时会被拆成多个重叠的track_id，
|
||
# 按原始track各自跑一遍会导致同一次摔倒被重复检测、重复插入
|
||
for merged_start, merged_end in _merge_windows(person_events):
|
||
start_frame_t = int(merged_start * fps) / fps
|
||
if start_frame_t <= last_t:
|
||
landmarker.close()
|
||
landmarker = _make_landmarker()
|
||
|
||
states = _extract_pose_states(
|
||
landmarker, cap, fps, width, height,
|
||
merged_start, merged_end,
|
||
)
|
||
if states:
|
||
last_t = states[-1][0]
|
||
|
||
for start_sec, end_sec in _find_fall_windows(states):
|
||
timestamp_abs = (start_time + timedelta(seconds=start_sec)).isoformat() if start_time else None
|
||
insert_event({
|
||
"video_id": video["id"],
|
||
"event_type": "fall",
|
||
"label": None,
|
||
"start_offset_sec": start_sec,
|
||
"end_offset_sec": end_sec,
|
||
"timestamp_abs": timestamp_abs,
|
||
"confidence": None,
|
||
"track_id": None,
|
||
"thumbnail_path": None,
|
||
"extra_json": None,
|
||
"created_at": datetime.now().isoformat(),
|
||
})
|
||
fall_count += 1
|
||
except Exception as e:
|
||
# 兜底：单条视频处理异常不应中断整个批次，标记失败并继续下一条。
|
||
# 异常本身可能就是底层解码/GPU上下文已经出问题，close()也可能再报错，不能让它盖过原始异常
|
||
try:
|
||
landmarker.close()
|
||
except Exception:
|
||
pass
|
||
cap.release()
|
||
update_video_status(video["id"], "fall_failed", error_message=str(e))
|
||
return fall_count
|
||
|
||
landmarker.close()
|
||
cap.release()
|
||
update_video_status(video["id"], "fall_done")
|
||
return fall_count
|
||
|
||
|
||
def process_all(limit=None):
|
||
videos = get_all_videos(status="person_done")
|
||
if limit:
|
||
videos = videos[:limit]
|
||
|
||
events_by_video = {}
|
||
for e in get_events("person_presence"):
|
||
events_by_video.setdefault(e["video_id"], []).append(e)
|
||
|
||
processed = 0
|
||
total_falls = 0
|
||
for video in videos:
|
||
total_falls += process_video(video, events_by_video.get(video["id"], []))
|
||
processed += 1
|
||
|
||
return {"processed": processed, "total_falls": total_falls}
|
||
|
||
|
||
if __name__ == "__main__":
|
||
print("===== 开始检测摔倒事件 =====")
|
||
result = process_all()
|
||
print(f"→ 处理 {result['processed']} 个视频，检测到 {result['total_falls']} 次疑似摔倒")
|
||
print("===== 检测完成 =====")
|