diff --git a/backend/app/routers.py b/backend/app/routers.py
index 8d1b1e9..edd4ccb 100644
--- a/backend/app/routers.py
+++ b/backend/app/routers.py
@@ -387,3 +387,42 @@ async def sse_events(request: Request):
             bus.unsubscribe(q)
 
     return StreamingResponse(gen(), media_type="text/event-stream")
+
+
+# ==================== 视觉识别事件上报（前端摄像头实时识别 → 后端日志/MQTT） ====================
+class VisionEventBody(BaseModel):
+    event: str                      # camera_on|camera_off|detecting|facing|triggered|silent|error|model_error
+    faces: int = 0
+    dwell_ms: int = 0
+    detail: str = ""
+
+@router.post("/api/vision/event")
+async def vision_event(body: VisionEventBody):
+    """前端摄像头识别状态/触发上报；后端记录详细日志，triggered 时广播 alert 到全屏。"""
+    log.info(
+        "vision event=%s faces=%d dwell_ms=%d detail=%s",
+        body.event, body.faces, body.dwell_ms, body.detail,
+    )
+    if body.event == "triggered":
+        hub.publish_command("alert", {"text": "有访客正对屏幕，语音助手已主动问候", "faces": body.faces})
+        log.info("vision triggered -> alert 已广播（faces=%d）", body.faces)
+    return {"ok": True}
+
+
+# ==================== YOLO 人脸检测（前端抽帧 → 后端推理） ====================
+class VisionFrameBody(BaseModel):
+    image: str = ""      # JPEG base64（不含 data: 前缀）
+    conf: float = 0.0    # 可选：覆盖置信度阈值
+
+@router.post("/api/vision/frame")
+def vision_frame(body: VisionFrameBody):
+    """接收前端抽帧 JPEG base64，后端 YOLO 推理返回人脸框（faces/boxes/latency_ms）。
+
+    普通 def 由 FastAPI 线程池执行（推理约 100-200ms），不阻塞事件循环。
+    """
+    img_b64 = (body.image or "").strip()
+    if not img_b64:
+        log.warning("vision frame: 缺少 image")
+        return {"ok": False, "error": "missing image"}
+    from .vision_yolo import predict_base64
+    return predict_base64(img_b64)
diff --git a/backend/app/vision_yolo.py b/backend/app/vision_yolo.py
new file mode 100644
index 0000000..e7de6ca
--- /dev/null
+++ b/backend/app/vision_yolo.py
@@ -0,0 +1,155 @@
+# -*- coding: utf-8 -*-
+"""YOLO 人脸检测 + 姿态估计（后端推理）
+
+链路：前端摄像头抽帧 → JPEG base64 → POST /api/vision/frame → 本模块推理 → 返回人脸/姿态/举手特征。
+模型：
+  - backend/models/yolov8n-face.pt （lindevs/yolov8-face，WIDERFACE 训练，torch 6.3MB）
+  - backend/models/yolov8n-pose.pt （ultralytics 官方，COCO 17 关键点，torch 6.8MB）
+推理：ultralytics 加载（torch 原生，自动 letterbox + NMS），CPU 设备。
+
+手势判定分工：
+  - 后端只做"单帧几何特征"（举手 = 手腕高于肩），无状态
+  - 前端做时序判定（举手持续 1.5s → 开始对话；举手后手腕摆动 → 挥手结束对话）
+"""
+import base64
+import io
+import logging
+import time
+from pathlib import Path
+
+from PIL import Image
+
+log = logging.getLogger("dpm.vision")
+
+MODEL_DIR = Path(__file__).resolve().parent.parent / "models"
+FACE_MODEL_PATH = MODEL_DIR / "yolov8n-face.pt"
+POSE_MODEL_PATH = MODEL_DIR / "yolov8n-pose.pt"
+CONF_THRESHOLD = 0.25   # 远距离小脸可调低（如 0.15）
+IMGSZ = 640
+
+# COCO 17 关键点索引
+KP_LEFT_SHOULDER = 5
+KP_RIGHT_SHOULDER = 6
+KP_LEFT_ELBOW = 7
+KP_RIGHT_ELBOW = 8
+KP_LEFT_WRIST = 9
+KP_RIGHT_WRIST = 10
+RAISE_LIFT = 0.12   # 举手：手腕高于对应肩 12% 画面高
+FIST_ELBOW_LIFT = 0.05   # 举拳：手腕高于肘 5% 画面高（前臂上举）
+FIST_SHOULDER_GAP = 0.15 # 举拳：拳不高于肩 15% 画面高（收在胸前）
+FIST_CHEST_DIST = 0.25   # 举拳：拳与肩水平距离 < 25% 画面宽（贴近躯干）
+KP_CONF_MIN = 0.3   # 关键点置信度过滤
+
+_face_model = None
+_pose_model = None
+
+
+def get_face_model():
+    global _face_model
+    if _face_model is None:
+        from ultralytics import YOLO
+        log.info("vision: 加载 YOLO 人脸模型 %s（首次约 2-5s）", FACE_MODEL_PATH)
+        _face_model = YOLO(str(FACE_MODEL_PATH), task="detect")
+        log.info("vision: YOLO 人脸模型就绪")
+    return _face_model
+
+
+def get_pose_model():
+    global _pose_model
+    if _pose_model is None:
+        from ultralytics import YOLO
+        log.info("vision: 加载 YOLO 姿态模型 %s（首次约 2-5s）", POSE_MODEL_PATH)
+        _pose_model = YOLO(str(POSE_MODEL_PATH), task="pose")
+        log.info("vision: YOLO 姿态模型就绪")
+    return _pose_model
+
+
+def _parse_pose(results):
+    """解析姿态结果 → persons（17 关键点归一化 [x,y,conf]）+ raised（举手）+ fists（举拳）"""
+    persons = []
+    raised = []
+    fists = []
+    for r in results:
+        if r.keypoints is None:
+            continue
+        kps = r.keypoints.data  # [N,17,3] 归一化坐标
+        for i in range(kps.shape[0]):
+            kp = kps[i]
+            person = [
+                [round(float(kp[j][0]), 3), round(float(kp[j][1]), 3), round(float(kp[j][2]), 3)]
+                for j in range(17)
+            ]
+            persons.append(person)
+            # 每只手臂：举手（腕明显高于肩）或 举拳（前臂上举、拳收胸前），两者互斥
+            for side, wrist_i, elbow_i, shoulder_i in (
+                ("left", KP_LEFT_WRIST, KP_LEFT_ELBOW, KP_LEFT_SHOULDER),
+                ("right", KP_RIGHT_WRIST, KP_RIGHT_ELBOW, KP_RIGHT_SHOULDER),
+            ):
+                w = kp[wrist_i]
+                e = kp[elbow_i]
+                s = kp[shoulder_i]
+                if float(w[2]) < KP_CONF_MIN or float(s[2]) < KP_CONF_MIN or float(e[2]) < KP_CONF_MIN:
+                    continue
+                wx, wy = float(w[0]), float(w[1])
+                sx, sy = float(s[0]), float(s[1])
+                ey = float(e[1])
+                if wy < sy - RAISE_LIFT:
+                    raised.append({"side": side, "x": round(wx, 3), "y": round(wy, 3)})
+                elif (
+                    wy < ey - FIST_ELBOW_LIFT      # 前臂上举（腕高于肘）
+                    and wy >= sy - FIST_SHOULDER_GAP  # 拳不高过肩太多（收在胸前）
+                    and abs(wx - sx) < FIST_CHEST_DIST  # 拳贴近躯干中线
+                ):
+                    fists.append({"side": side, "x": round(wx, 3), "y": round(wy, 3)})
+    return persons, raised, fists
+
+
+def predict_jpeg(jpeg_bytes: bytes):
+    """人脸检测 + 姿态估计 → {faces, boxes, pose, raised, latency_ms}"""
+    t0 = time.time()
+    img = Image.open(io.BytesIO(jpeg_bytes)).convert("RGB")
+
+    # 人脸
+    face_res = get_face_model().predict(img, conf=CONF_THRESHOLD, imgsz=IMGSZ, verbose=False, device="cpu")
+    boxes = []
+    for r in face_res:
+        if r.boxes is None:
+            continue
+        for b in r.boxes:
+            xyxy = [round(float(v), 1) for v in b.xyxy[0].tolist()]
+            conf = round(float(b.conf[0]), 3)
+            boxes.append({"box": xyxy, "conf": conf})
+
+    # 姿态
+    pose_res = get_pose_model().predict(img, conf=CONF_THRESHOLD, imgsz=IMGSZ, verbose=False, device="cpu")
+    persons, raised, fists = _parse_pose(pose_res)
+
+    latency_ms = round((time.time() - t0) * 1000, 1)
+    log.info(
+        "vision: 人脸 %d 姿态 %d 举手 %d 举拳 %d · %.0fms",
+        len(boxes), len(persons), len(raised), len(fists), latency_ms,
+    )
+    return {
+        "faces": len(boxes),
+        "boxes": boxes,
+        "pose": persons,
+        "raised": raised,
+        "fists": fists,
+        "latency_ms": latency_ms,
+    }
+
+
+def predict_base64(b64: str):
+    """入口：base64 JPEG → 检测结果 dict（含 ok 标记，失败时带 error）"""
+    try:
+        jpeg = base64.b64decode(b64)
+    except Exception as e:  # noqa: BLE001
+        log.warning("vision: base64 解码失败 %s", e)
+        return {"ok": False, "error": f"bad base64: {e}"}
+    try:
+        data = predict_jpeg(jpeg)
+        data["ok"] = True
+        return data
+    except Exception as e:  # noqa: BLE001
+        log.error("vision: 推理失败 %s", e, exc_info=True)
+        return {"ok": False, "error": str(e)}
diff --git a/backend/models/yolov8n-face.pt b/backend/models/yolov8n-face.pt
new file mode 100644
index 0000000..4f52f70
Binary files /dev/null and b/backend/models/yolov8n-face.pt differ
diff --git a/backend/models/yolov8n-pose.pt b/backend/models/yolov8n-pose.pt
new file mode 100644
index 0000000..f41b11f
Binary files /dev/null and b/backend/models/yolov8n-pose.pt differ