jc/jc-video-recognize
11
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(server): 新增事件聚合引擎。实现时间窗口去重、空间邻近合并和置信度加权融合的事件聚合能力，减少重复预警。

Browse Source
This commit is contained in:
wuzhuorong
2026-06-12 14:05:22 +08:00
parent bf12a29acd
commit 21829bcbae
1 changed files with 145 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

166
apps/server/services/event/aggregator.py
View File

@@ -1,10 +1,13 @@
"""事件聚合器 (MVP-1 / P5 简化版)
"""事件聚合器 (MVP-1 / P5 + MVP-2 / D20 增强版)
时间窗口去重: 对同一 ``(source_id, event_type, track_id_or_bbox_hash)``
在配置窗口内只保留一条预警事件，重复命中时更新 ``last_seen``
与 ``occurrence_count``，避免预警刷屏。
MVP-1 能力:
- 时间窗口去重: 对同一 (source_id, event_type, track_id_or_bbox_hash)
在配置窗口内只保留一条预警事件
后续 MVP-2 会扩展空间邻近合并、置信度加权融合等能力。
MVP-2 / D20 新增能力:
- 空间邻近合并: 同一 (source_id, event_type) 且 bbox IOU 高的事件视为同一目标
- 置信度加权融合: 多次命中时按时间衰减加权融合置信度
- 融合策略可选: weighted / max / avg
"""
from __future__ import annotations
@@ -12,31 +15,78 @@ from __future__ import annotations
import logging
import time
from collections import OrderedDict
from typing import Dict, List, Optional, Tuple
from typing import Dict, List, Optional, Tuple, TypeAlias
from models.event_schemas import AlertEvent, UnifiedDetection
from models.event_schemas import AlertEvent, BBox, UnifiedDetection
logger = logging.getLogger(__name__)
_AggKey = Tuple[Optional[str], str, str]
_AggKey: TypeAlias = Tuple[Optional[str], str, str]
# ---------------------------------------------------------------------------
# IOU 工具
# ---------------------------------------------------------------------------
def _bbox_iou(b1: BBox, b2: BBox) -> float:
inter_x1 = max(b1.x1, b2.x1)
inter_y1 = max(b1.y1, b2.y1)
inter_x2 = min(b1.x2, b2.x2)
inter_y2 = min(b1.y2, b2.y2)
inter_w = max(0, inter_x2 - inter_x1)
inter_h = max(0, inter_y2 - inter_y1)
inter = inter_w * inter_h
if inter == 0:
return 0.0
union = b1.area + b2.area - inter
return inter / union if union > 0 else 0.0
# ---------------------------------------------------------------------------
# EventAggregator
# ---------------------------------------------------------------------------
class EventAggregator:
"""基于时间窗口的预警去重 / 聚合器。
"""基于时间窗口 + 空间邻近的预警去重 / 融合器。
Args:
dedup_window_seconds: 去重窗口 (秒)，同 key 在窗口内不会重复产出
max_active_events: 内存中最大活跃事件数，超过时按 LRU 淘汰
enable_spatial_merge: 是否启用空间邻近合并 (MVP-2)
spatial_iou_threshold: 空间合并 IOU 阈值
fusion_strategy: 置信度融合策略: ``weighted`` / ``max`` / ``avg``
fusion_decay_factor: 历史置信度衰减因子 (越小越偏向新数据)
"""
SUPPORTED_FUSION_STRATEGIES = ("weighted", "max", "avg")
def __init__(
self,
dedup_window_seconds: float = 30.0,
max_active_events: int = 1000,
enable_spatial_merge: bool = False,
spatial_iou_threshold: float = 0.3,
fusion_strategy: str = "max",
fusion_decay_factor: float = 0.9,
) -> None:
self.dedup_window_seconds = max(0.0, dedup_window_seconds)
self.max_active_events = max(1, max_active_events)
self.enable_spatial_merge = enable_spatial_merge
self.spatial_iou_threshold = max(0.0, min(1.0, spatial_iou_threshold))
if fusion_strategy not in self.SUPPORTED_FUSION_STRATEGIES:
raise ValueError(
f"不支持的融合策略: {fusion_strategy}, "
f"支持: {self.SUPPORTED_FUSION_STRATEGIES}"
)
self.fusion_strategy = fusion_strategy
self.fusion_decay_factor = max(0.0, min(1.0, fusion_decay_factor))
# 按插入顺序保存以便 LRU 淘汰
self._active: "OrderedDict[_AggKey, AlertEvent]" = OrderedDict()
@@ -45,15 +95,24 @@ class EventAggregator:
# ------------------------------------------------------------------
def aggregate(self, alerts: List[AlertEvent]) -> List[AlertEvent]:
"""聚合一批预警事件，返回去重后真正应当对外发出的事件。"""
"""聚合一批预警事件，返回去重 / 融合后真正应当对外发出的事件。"""
now = time.time()
self._evict_expired(now)
emitted: List[AlertEvent] = []
for alert in alerts:
# 1. 优先按 key (track_id / bbox 网格) 精确匹配
key = self._make_key(alert)
existing = self._active.get(key)
# 2. 空间邻近合并: 若 key 未命中，尝试 IOU 匹配
if existing is None and self.enable_spatial_merge:
spatial_key = self._find_spatial_match(alert)
if spatial_key is not None:
existing = self._active.get(spatial_key)
key = spatial_key # 复用旧 key
if existing is None:
self._active[key] = alert
self._active.move_to_end(key)
@@ -62,16 +121,82 @@ class EventAggregator:
dropped_key, _ = self._active.popitem(last=False)
logger.debug("聚合器 LRU 淘汰事件: %s", dropped_key)
else:
# 窗口内重复：仅更新统计
existing.last_seen = now
existing.occurrence_count += 1
if alert.confidence > existing.confidence:
existing.confidence = alert.confidence
# 窗口内重复：融合统计
self._fuse(existing, alert, now)
self._active.move_to_end(key)
return emitted
# ------------------------------------------------------------------
# 内部
# 融合
# ------------------------------------------------------------------
def _fuse(self, existing: AlertEvent, new: AlertEvent, now: float) -> None:
"""将新事件融合到已有事件。"""
existing.last_seen = now
existing.occurrence_count += 1
# 置信度融合
if self.fusion_strategy == "max":
existing.confidence = max(existing.confidence, new.confidence)
elif self.fusion_strategy == "avg":
n = existing.occurrence_count
existing.confidence = (
existing.confidence * (n - 1) + new.confidence
) / n
else: # weighted
decay = self.fusion_decay_factor
existing.confidence = (
existing.confidence * decay + new.confidence * (1 - decay)
)
# 取整到 4 位避免浮点漂移
existing.confidence = round(
max(0.0, min(1.0, existing.confidence)), 4
)
# 严重性向上提升 (新事件更严重时)
severity_order = ["info", "low", "medium", "high", "critical"]
try:
if severity_order.index(new.severity.value) > severity_order.index(
existing.severity.value
):
existing.severity = new.severity
except ValueError:
pass
# 更新最新的检测目标 (用于 LLM 触发器拿到最新 bbox)
if new.detections:
existing.detections = new.detections
# ------------------------------------------------------------------
# 空间匹配
# ------------------------------------------------------------------
def _find_spatial_match(self, alert: AlertEvent) -> Optional[_AggKey]:
"""在活跃事件中寻找空间邻近的同类型事件。"""
if not alert.detections:
return None
target_bbox = alert.detections[0].bbox
best_key: Optional[_AggKey] = None
best_iou = 0.0
for key, existing in self._active.items():
# 必须同 source + 同事件类型
if key[0] != alert.source_id:
continue
if key[1] != alert.event_type.value:
continue
if not existing.detections:
continue
iou = _bbox_iou(target_bbox, existing.detections[0].bbox)
if iou >= self.spatial_iou_threshold and iou > best_iou:
best_iou = iou
best_key = key
return best_key
# ------------------------------------------------------------------
# Key 构造
# ------------------------------------------------------------------
@staticmethod
@@ -88,16 +213,17 @@ class EventAggregator:
if det.track_id is not None:
return f"t{det.track_id}"
# 按 50px 栅格归一，避免微小抖动导致漏聚合
cx, cy = det.bbox.center
return f"g{int(cx) // 50}_{int(cy) // 50}_{det.class_name}"
# ------------------------------------------------------------------
# 过期淘汰
# ------------------------------------------------------------------
def _evict_expired(self, now: float) -> None:
if self.dedup_window_seconds <= 0:
self._active.clear()
return
# 基于 first_seen 判断窗口过期：保证 "同一目标 30 秒内只产生 1 条事件"，
# 同时窗口到期后允许再次触发 (而不会因为不断收到同类检测而无限延期)。
expired_keys = [
key
for key, alert in self._active.items()
@@ -115,8 +241,6 @@ class EventAggregator:
return len(self._active)
def snapshot(self) -> Dict[str, AlertEvent]:
"""返回当前活跃事件的快照 (用于调试 / 监控)。"""
return {f"{k[0]}|{k[1]}|{k[2]}": v for k, v in self._active.items()}