Use max known similarity for unknown face rule

configs/full_pipeline_1080p_test_alarm.json

@@ -377,7 +377,7 @@
               "name": "unknown_face",
               "type": "unknown",
               "cooldown_ms": 7000,
-              "min_sim": 0.35,
+              "max_known_sim": 0.35,
               "min_hits": 1,
               "hit_window_ms": 1500,
               "min_face_area_ratio": 0.001,

docs/config_guide.md

@@ -571,7 +571,7 @@ python tools/analyze_face_recog_log.py .\logs\media-server_latest.log
       "name": "unknown_face",
       "type": "unknown",
       "cooldown_ms": 7000,
-      "min_sim": 0.35,
+      "max_known_sim": 0.35,
       "min_hits": 1,
       "hit_window_ms": 1500,
       "min_face_area_ratio": 0.001,
@@ -602,7 +602,8 @@ python tools/analyze_face_recog_log.py .\logs\media-server_latest.log
 | `face_track_aggregation.known.reentry_cooldown_ms` | 同一已知人短时间离开再进入时的抑制时间 | 打卡场景建议开启，例如 `8000` 以上 |
 | `face_track_aggregation.unknown.min_track_age_ms` | 陌生人候选 track 至少持续多久 | 建议 `2000` 起，避免一闪而过的小脸误报 |
 | `face_track_aggregation.unknown.min_quality_hits` | 陌生人候选需要多少次有效质量帧 | 建议 `4` 起，保证陌生人告警更准 |
-| `face_rules[].min_sim` | 进入该规则的最低相似度条件 | `known_person` 当前测试为 `0.45` |
+| `known_person.min_sim` | 确认已知人的最低相似度条件 | 当前测试为 `0.45`，越高越保守 |
+| `unknown_face.max_known_sim` | 陌生人候选允许的最高“已知人相似度” | 当前测试为 `0.35`；超过该值说明仍像库中某人，不直接报陌生人 |
 | `face_rules[].min_face_area_ratio` | 过滤小脸框 | 1080p 下 `0.0002` 约等于 `415px²`，`0.001` 约等于 `2074px²` |
 | `face_rules[].cooldown_ms` | 同一规则冷却 | 测试可 `7000`，正式按后台接收频率调整 |
 | `face_rules[].min_face_aspect / max_face_aspect` | 过滤异常长宽比人脸框 | `known_person` 当前测试为 `0.55` 到 `1.6`；`unknown_face` 建议保持 `0.6` 到 `1.6` |
@@ -689,6 +690,7 @@ python tools/analyze_face_recog_log.py .\logs\media-server_latest.log
 | `face_recog` | `threshold.accept` | 控制已知人识别最低相似度 | 误认减少，但 known 变少 | known 变多，但误认风险升高 |
 | `face_recog` | `threshold.margin` | 控制 top1 和 top2 的区分度 | 相似人员误认减少，但 known 变少 | known 变多，但相似人员更容易混淆 |
 | `known_person` | `min_face_area_ratio` | 控制已知人告警可接受的人脸最小尺寸 | 小脸告警减少，更稳 | 更容易触发，但小脸质量风险升高 |
+| `unknown_face` | `max_known_sim` | 控制陌生人候选是否“仍太像已知人” | 减少把已知人波动报成陌生人 | 更容易把未知人纳入陌生人聚合 |
 | `face_track_aggregation.known` | `min_hits` | 控制已知人需要多少次稳定识别才告警 | 更稳，适合正式打卡 | 更灵敏，适合验证链路 |
 | `alarm` | `min_duration_ms` | 控制要稳定多久才报警 | 更稳，但慢一点 | 更灵敏，但更容易闪报 |
 | `alarm` | `cooldown_ms` | 控制两次告警间隔 | 减少重复告警 | 同一事件会更频繁重复报 |
@@ -779,7 +781,7 @@ python tools/analyze_face_recog_log.py .\logs\media-server_latest.log
 
 ### Q5: unknown 和 uncertain 有什么区别？
 
-`uncertain` 表示“像某个已知人，但证据不足”，不能当作陌生人。`unknown_face` 应用于“持续存在、质量足够、且没有形成已知人证据”的人体 track。这样可以避免远处小脸、合成人脸或短暂模糊帧被误报成陌生人。
+`uncertain` 是单帧识别状态，表示“没有稳定确认是已知人”。它本身不能直接当作陌生人。`unknown_face` 是告警规则，使用开放集识别思路：人脸质量足够、不是 known、`best_sim < unknown_face.max_known_sim`、同一人体 track 持续存在并累计足够质量帧后，才认为是陌生人。这样可以避免远处小脸、合成人脸或短暂模糊帧被误报成陌生人。
 
 ---
 

plugins/alarm/alarm_node.cpp

@@ -645,6 +645,7 @@ private:
         float min_face_aspect = 0.0f;
         float max_face_aspect = 0.0f;
         float min_sim = 0.0f;
+        float max_known_sim = 0.0f;
     };
 
     struct FaceRuleEval {
@@ -704,6 +705,8 @@ private:
             r.max_face_aspect = static_cast<float>(
                 item.ValueOr<double>("max_face_aspect", static_cast<double>(r.max_face_aspect)));
             r.min_sim = static_cast<float>(item.ValueOr<double>("min_sim", static_cast<double>(r.min_sim)));
+            r.max_known_sim = static_cast<float>(
+                item.ValueOr<double>("max_known_sim", static_cast<double>(r.max_known_sim)));
 
             if (const SimpleJson* persons = item.Find("persons"); persons && persons->IsArray()) {
                 for (const auto& p : persons->AsArray()) {
@@ -765,9 +768,9 @@ private:
                 out.detail = "status=known";
                 return false;
             }
-            if (it.best_sim < rule.min_sim) {
-                out.reject_reason = "min_sim";
-                out.detail = "best_sim=" + Fixed3(it.best_sim) + " min=" + Fixed3(rule.min_sim);
+            if (rule.max_known_sim > 0.0f && it.best_sim >= rule.max_known_sim) {
+                out.reject_reason = "max_known_sim";
+                out.detail = "best_sim=" + Fixed3(it.best_sim) + " max=" + Fixed3(rule.max_known_sim);
                 return false;
             }
         } else {

tests/test_face_track_alarm.cpp

@@ -250,20 +250,35 @@ TEST(FaceTrackAlarmTest, UntrackedQualifiedFaceDoesNotTriggerAlarm) {
     EXPECT_EQ(node.alarm_count_, 0u);
 }
 
-TEST(FaceTrackAlarmTest, UnknownRuleDiagnosticsExplainMinSimRejection) {
+TEST(FaceTrackAlarmTest, UnknownRuleAcceptsLowKnownSimilarityCandidate) {
     AlarmNode::FaceRule rule;
     rule.name = "unknown_face";
     rule.kind = AlarmNode::FaceRule::Kind::Unknown;
-    rule.min_sim = 0.35f;
+    rule.max_known_sim = 0.35f;
 
     FaceRecogItem item = MakeUncertainFace(101, 7, "reg_007", 0.20f);
     item.bbox = Rect{0.0f, 0.0f, 20.0f, 20.0f};
 
+    AlarmNode::FaceRuleEval eval;
+    EXPECT_TRUE(AlarmNode::FaceItemMatchesRule(rule, item, 10000.0, &eval));
+    EXPECT_TRUE(eval.matched);
+}
+
+TEST(FaceTrackAlarmTest, UnknownRuleDiagnosticsExplainMaxKnownSimRejection) {
+    AlarmNode::FaceRule rule;
+    rule.name = "unknown_face";
+    rule.kind = AlarmNode::FaceRule::Kind::Unknown;
+    rule.max_known_sim = 0.35f;
+
+    FaceRecogItem item = MakeUncertainFace(101, 7, "reg_007", 0.42f);
+    item.bbox = Rect{0.0f, 0.0f, 20.0f, 20.0f};
+
     AlarmNode::FaceRuleEval eval;
     EXPECT_FALSE(AlarmNode::FaceItemMatchesRule(rule, item, 10000.0, &eval));
     EXPECT_FALSE(eval.matched);
-    EXPECT_EQ(eval.reject_reason, "min_sim");
-    EXPECT_NE(eval.detail.find("best_sim=0.200"), std::string::npos);
+    EXPECT_EQ(eval.reject_reason, "max_known_sim");
+    EXPECT_NE(eval.detail.find("best_sim=0.420"), std::string::npos);
+    EXPECT_NE(eval.detail.find("max=0.350"), std::string::npos);
 }
 
 TEST(FaceTrackAlarmTest, UnknownRuleDiagnosticsUseHighPrecisionAreaRatio) {
@@ -535,7 +550,8 @@ TEST(FaceTrackAlarmTest, ParsesTrackAggregationWithoutUnknownReentryCooldownCont
             {
                 "name": "unknown_face",
                 "type": "unknown",
-                "cooldown_ms": 0
+                "cooldown_ms": 0,
+                "max_known_sim": 0.35
             }
         ],
         "face_track_aggregation": {
@@ -555,6 +571,8 @@ TEST(FaceTrackAlarmTest, ParsesTrackAggregationWithoutUnknownReentryCooldownCont
     EXPECT_EQ(node.track_agg_cfg_.known_reentry_cooldown_ms, 60000);
     EXPECT_EQ(node.track_agg_cfg_.unknown_min_track_age_ms, 1200);
     EXPECT_EQ(node.track_agg_cfg_.unknown_min_quality_hits, 2);
+    ASSERT_EQ(node.face_rules_.size(), 1u);
+    EXPECT_FLOAT_EQ(node.face_rules_[0].max_known_sim, 0.35f);
 }
 
 }  // namespace