backend/vision_engine.py

"""
vision_engine.py — Single source of all camera frames and AI processing.

Key principles:
  - ONE cv2.VideoCapture per camera index.  No other module opens cameras.
  - AI processing (YOLO crowd count + face recognition) is gated by `ai_enabled`.
  - When AI runs, detections are pushed into gossip_bridge for interaction tracking.
  - gossip_bridge never opens a camera itself.
"""
import cv2
import base64
import numpy as np
import logging
import threading
import os
import sys
import uuid
from datetime import datetime, timezone
from ultralytics import YOLO

# ── Add Face_Recognition/ to path so gossip_bridge can be imported
_FR_DIR = os.path.join(os.path.dirname(__file__), "Face_Recognition")
if _FR_DIR not in sys.path:
    sys.path.insert(0, _FR_DIR)
import gossip_bridge  # noqa: E402
from vision_runtime import detect_acceleration, insightface_ctx_id

logger = logging.getLogger(__name__)


def _build_face_engine():
    """Prefer FaceMatcher (buffalo_sc, enrolled-first, unknown persistence) over legacy engine."""
    try:
        from face_matcher import FaceMatcher

        logger.info(
            "VisionEngine using FaceMatcher (model=%s)",
            os.getenv("FACE_MODEL_PACK", "buffalo_sc"),
        )
        return FaceMatcher()
    except Exception as exc:
        logger.warning("FaceMatcher unavailable (%s) — using legacy FaceRecognitionEngine", exc)
        return FaceRecognitionEngine()


# ─────────────────────────────────────────────────────────────────────────────
# Face Recognition Engine (lazy-loads InsightFace)
# ─────────────────────────────────────────────────────────────────────────────

class FaceRecognitionEngine:
    def __init__(self):
        self.app  = None
        self.db: dict[str, np.ndarray] = {}
        self.lock = threading.Lock()
        self._load_model()
        self._load_db()

    def _load_model(self):
        try:
            from insightface.app import FaceAnalysis  # type: ignore

            accel = detect_acceleration()
            ctx_id = insightface_ctx_id()
            provider = "cuda" if ctx_id >= 0 else "cpu"
            fa = FaceAnalysis(name="buffalo_l", allowed_modules=["detection", "recognition"])
            fa.prepare(ctx_id=ctx_id, det_size=(640, 640))
            self.app = fa
            if provider == "cuda":
                logger.info(
                    "InsightFace model loaded (provider=cuda, device=%s).",
                    accel.get("device_name") or "cuda:0",
                )
            else:
                reason = accel.get("fallback_reason") or "no GPU detected"
                logger.info("InsightFace model loaded (provider=cpu, fallback=%s).", reason)
        except Exception as e:
            logger.warning(f"InsightFace not available ({e}). Face recognition disabled.")

    def _load_db(self):
        base = _FR_DIR
        self.db = {}
        for folder in ["faces_db", "temp_faces_db"]:
            path = os.path.join(base, folder)
            if not os.path.exists(path):
                logger.warning(f"Face DB folder not found: {path}")
                continue
            files = [f for f in os.listdir(path) if f.endswith(".npy")]
            for f in files:
                name = f.replace(".npy", "")
                try:
                    emb  = np.load(os.path.join(path, f))
                    self.db[name] = emb
                    logger.info(f"Loaded face: {name} from {folder}")
                except Exception as e:
                    logger.error(f"Failed to load face {f}: {e}")
        
        if not self.db:
            logger.warning("Face database is EMPTY. No faces will be recognized.")
        else:
            logger.info(f"Face DB loaded successfully: {list(self.db.keys())}")

    @staticmethod
    def _cosine(a: np.ndarray, b: np.ndarray) -> float:
        na, nb = np.linalg.norm(a), np.linalg.norm(b)
        if na == 0 or nb == 0:
            return 0.0
        return float(np.dot(a, b) / (na * nb))

    def _allocate_unknown_id(self) -> int:
        existing_ids = []
        for k in self.db.keys():
            if k.startswith("unknown_"):
                try:
                    existing_ids.append(int(k.split("_")[1]))
                except (IndexError, ValueError):
                    pass
        base = _FR_DIR
        for folder in ["faces_db", "temp_faces_db", "face_database", "temp_face_database"]:
            path = os.path.join(base, folder)
            if os.path.exists(path):
                for item in os.listdir(path):
                    name = item.replace(".npy", "")
                    if name.startswith("unknown_"):
                        try:
                            existing_ids.append(int(name.split("_")[1]))
                        except (IndexError, ValueError):
                            pass
        return max(existing_ids) + 1 if existing_ids else 1

    def recognize(self, frame: np.ndarray) -> list[dict]:
        """Return recognized faces with bbox, name, confidence."""
        if self.app is None:
            return []
        with self.lock:
            try:
                faces = self.app.get(frame)
            except Exception as e:
                logger.error(f"InsightFace inference error: {e}")
                return []

        results = []
        for face in faces:
            x1, y1, x2, y2 = face.bbox.astype(int)
            emb = face.embedding
            best_name  = "Unknown"
            best_score = 0.0
            for name, db_emb in self.db.items():
                # db_emb can be (512,) or (N, 512)
                if db_emb.ndim == 1:
                    score = self._cosine(emb, db_emb)
                else:
                    score = max((self._cosine(emb, v) for v in db_emb), default=0.0)
                
                if score > best_score:
                    best_score = score
                    best_name  = name
            
            # Conservative threshold to reduce false-positive identity matches.
            threshold = 0.35 if best_name.startswith("unknown") else 0.30
            if best_score < threshold:
                # Assign a new persistent unknown identity
                next_id = self._allocate_unknown_id()
                new_name = f"unknown_{next_id}"
                
                # Update in-memory db
                self.db[new_name] = np.array([emb])
                
                # Save embedding on disk
                emb_dir = os.path.join(_FR_DIR, "temp_faces_db")
                os.makedirs(emb_dir, exist_ok=True)
                np.save(os.path.join(emb_dir, f"{new_name}.npy"), np.array([emb]))
                
                # Save crop image to face_database
                img_dir = os.path.join(_FR_DIR, "face_database", new_name)
                os.makedirs(img_dir, exist_ok=True)
                
                h, w, _ = frame.shape
                x1_c, y1_c = max(0, x1), max(0, y1)
                x2_c, y2_c = min(w, x2), min(h, y2)
                face_img = frame[y1_c:y2_c, x1_c:x2_c]
                if face_img.size > 0:
                    cv2.imwrite(os.path.join(img_dir, "0.jpg"), face_img)
                
                best_name = new_name
                best_score = 1.0  # Initial creation match confidence
            
            results.append({
                "name":       best_name,
                "confidence": round(best_score, 3),
                "bbox":       [int(x1), int(y1), int(x2), int(y2)],
            })
        return results

    def match_all_faces(self, frame: np.ndarray, threshold: float | None = None) -> list[dict]:
        """Detect and identify every face in the frame. Parity with FaceMatcher."""
        return self.recognize(frame)


    def register_face_from_frame(self, name: str, frame: np.ndarray) -> bool:
        """
        Register a new face:
        1. Saves image to face_database/<name>/
        2. Generates embeddings (including occluded version)
        3. Saves .npy to faces_db/
        4. Updates in-memory DB
        """
        if self.app is None:
            return False
        try:
            import register_face
            
            # Detect face to get initial embedding (to ensure it works and for robustness)
            faces = self.app.get(frame)
            if not faces:
                logger.warning(f"Registration failed: No face detected in frame for '{name}'")
                return False
            primary_emb = faces[0].embedding
            
            # Define paths
            db_root  = os.path.join(_FR_DIR, "face_database")
            emb_root = os.path.join(_FR_DIR, "faces_db")
            os.makedirs(db_root, exist_ok=True)
            os.makedirs(emb_root, exist_ok=True)
            
            # Save temporary image for the register_face module to process
            temp_path = os.path.join(_FR_DIR, f"temp_reg_{uuid.uuid4().hex}.jpg")
            cv2.imwrite(temp_path, frame)
            
            try:
                # Call the specialized register_face logic (handles folder creation, image copying, occlusion)
                # We pass self.app to avoid redundant model loading
                new_embs = register_face.register_face(
                    name, temp_path, db_root=db_root, emb_root=emb_root, 
                    known_embedding=primary_emb, app=self.app
                )
                
                # Update in-memory DB so it's available immediately without restart
                with self.lock:
                    self.db[name] = new_embs
                
                logger.info(f"Registered face '{name}': Image saved to database and embeddings generated.")
                return True
            finally:
                if os.path.exists(temp_path):
                    os.remove(temp_path)
                    
        except Exception as e:
            logger.error(f"Error registering face: {e}")
            return False

    def search_missing_person(self, query_frame: np.ndarray, cam_frames: dict) -> dict:
        if self.app is None:
            return {"found": False, "reason": "Face recognition unavailable"}
        try:
            query_faces = self.app.get(query_frame)
        except Exception as e:
            return {"found": False, "reason": str(e)}
        if not query_faces:
            return {"found": False, "reason": "No face detected in query image"}

        query_emb = max(query_faces, key=lambda f: (f.bbox[2] - f.bbox[0]) * (f.bbox[3] - f.bbox[1])).embedding

        best_name = None
        best_score = 0.0
        for name, db_emb in self.db.items():
            if db_emb.ndim == 1:
                score = self._cosine(query_emb, db_emb)
            else:
                score = max((self._cosine(query_emb, v) for v in db_emb), default=0.0)
            if score > best_score:
                best_score = score
                best_name = name

        threshold = 0.35 if (best_name and best_name.startswith("unknown")) else 0.30
        if best_name and best_score >= threshold:
            return {
                "found": True,
                "name": best_name.replace("_", " "),
                "confidence": round(best_score, 3),
                "cam_id": "database",
                "location": "Enrolled database",
                "search_mode": "database",
                "timestamp": datetime.now(timezone.utc).isoformat(),
                "reason": "Matched enrolled face database",
            }

        best_match = {"found": False, "cam_id": None, "confidence": round(best_score, 3), "timestamp": None}
        for cam_id, frame in cam_frames.items():
            if frame is None:
                continue
            try:
                live_faces = self.app.get(frame)
            except Exception:
                continue
            for face in live_faces:
                score = self._cosine(query_emb, face.embedding)
                if score > best_match["confidence"]:
                    matched = score >= threshold
                    best_match = {
                        "found": matched,
                        "name": best_name.replace("_", " ") if best_name and matched else None,
                        "cam_id": cam_id,
                        "confidence": round(score, 3),
                        "timestamp": datetime.now(timezone.utc).isoformat(),
                        "location": cam_id,
                        "search_mode": "live" if matched else None,
                    }

        if not best_match.get("found"):
            best_match["reason"] = "Face detected but no match in enrolled database"
            best_match["best_score"] = round(best_score, 3)
            best_match["threshold"] = threshold
            best_match["enrolled_count"] = len([k for k in self.db if not k.startswith("unknown_")])
            best_match.setdefault("search_mode", "none")
        elif best_match.get("found"):
            best_match.setdefault("search_mode", "live")
        return best_match


# ─────────────────────────────────────────────────────────────────────────────
# Vision Engine — single owner of all VideoCapture objects
# ─────────────────────────────────────────────────────────────────────────────

class VisionEngine:
    def __init__(self):
        model_path = os.getenv("YOLO_MODEL_PATH", "yolov5nu.pt")
        logger.info("Loading YOLO model (%s)...", model_path)
        try:
            self.model = YOLO(model_path)
            logger.info("YOLO model loaded.")
        except Exception as e:
            logger.error(f"Failed to load YOLO: {e}")
            self.model = None

        self.active_cameras:     dict[str, cv2.VideoCapture] = {}
        self.camera_indices:     dict[str, int | str]         = {}
        self._client_cameras:    dict[str, set[str]]          = {}
        self._cam_refcounts:     dict[str, int]               = {}
        self.latest_raw_frames:  dict[str, np.ndarray | None] = {}
        self.browser_annotated_frames: dict[str, tuple[int, np.ndarray | None]] = {}
        self.lock = threading.Lock()
        self.open_lock = threading.Lock()       # strictly sequential cv2.VideoCapture
        self._pending_opens: set[str] = set()   # cam_ids currently being opened

        self.face_engine   = _build_face_engine()
        self.face_results: dict[str, list[dict]] = {}

        # ── AI toggle — controlled by the Spatial page models button ──────────
        # Dict of model_id → enabled bool.
        # Facial recognition is ON by default: it auto-starts on every camera
        # feed at launch and stays on until explicitly disabled.
        self.ai_models: dict[str, bool] = {
            "crowd":   True,   # YOLO crowd detection
            "facial":  True,    # InsightFace recognition — always-on by default
            "fight":   False,
            "anomaly": False,   # Used for Stampede Detection
            "medical": False,   # Used for Fall Detection
        }
        
        # ── Model States ──────────────────────────────────────────────────────
        self.fall_state: dict[str, int] = {}
        self.prev_gray: dict[str, np.ndarray] = {}
        self.last_event_ts: dict[str, datetime] = {}  # {cam_id_event_type: ts}
        self.last_total_count: int = 0
        self.last_crowd_count: int = 0
        self.last_face_count: int = 0
        self.room_stats: dict[str, dict] = {}

        accel = detect_acceleration()
        if self.model is not None and accel["cuda_available"]:
            logger.info(
                "YOLO will use CUDA (device=%s).",
                accel.get("device_name") or "cuda:0",
            )
        elif self.model is not None:
            logger.info(
                "YOLO using CPU (fallback=%s).",
                accel.get("fallback_reason") or "no GPU",
            )

    def warmload_models(self) -> dict:
        """Run lightweight dummy inference to populate GPU memory and warm ONNX/torch."""
        accel = detect_acceleration()
        result: dict = {
            "yolo": False,
            "insightface": False,
            "provider": accel.get("provider", "cpu"),
            "cuda_available": accel.get("cuda_available", False),
            "device_name": accel.get("device_name"),
            "fallback_reason": accel.get("fallback_reason"),
        }
        dummy_bgr = np.zeros((480, 640, 3), dtype=np.uint8)

        if self.model is not None:
            try:
                self.model.predict(dummy_bgr, verbose=False)
                result["yolo"] = True
                logger.info("Warmload: YOLO dummy inference OK (provider=%s).", result["provider"])
            except Exception as exc:
                logger.warning("Warmload: YOLO dummy inference failed: %s", exc)
                result["yolo_error"] = str(exc)

        if self.face_engine.app is not None:
            try:
                # Step 1: Call prepare() exactly once
                if hasattr(self.face_engine, "_force_prepare"):
                    self.face_engine._force_prepare()

                # Step 2: Force detection model with real-sized image
                dummy_insight = np.zeros((320, 320, 3), dtype=np.uint8)
                with self.face_engine.lock:
                    faces = self.face_engine.app.get(dummy_insight)
                
                det_ok = True
                rec_ok = True

                # Force InsightFace's internal per-model ONNX sessions — prevent lazy init on first WS frame
                import numpy as _np
                for _mname, _mobj in self.face_engine.app.models.items():
                    _session = getattr(_mobj, 'session', None)
                    if _session is not None:
                        try:
                            _inp_meta = _session.get_inputs()[0]
                            _shape = [1 if (d is None or isinstance(d, str)) else d
                                      for d in _inp_meta.shape]
                            _dummy = _np.zeros(_shape, dtype=_np.float32)
                            _session.run(None, {_inp_meta.name: _dummy})
                            logger.info("Warmload: InsightFace model '%s' ONNX session force-run OK", _mname)
                        except Exception as _exc:
                            logger.warning("Warmload: InsightFace model '%s' ONNX force-run skipped: %s",
                                           _mname, _exc)
                    else:
                        logger.warning("Warmload: InsightFace model '%s' has no session attribute — "
                                       "may still lazy-init", _mname)

                result["insightface"] = True
                ctx = insightface_ctx_id()
                logger.info(
                    "Warmload: InsightFace force-warmload OK (ctx_id=%s, faces on blank: %d, det_ok: %s, rec_ok: %s).",
                    ctx, len(faces), det_ok, rec_ok
                )
            except Exception as exc:
                logger.warning("Warmload: InsightFace dummy inference failed: %s", exc)
                result["insightface_error"] = str(exc)

        ok = result["yolo"] or result["insightface"] or (self.model is None and self.face_engine.app is None)
        result["success"] = bool(ok)
        if result["success"]:
            logger.info("Startup warmload success: %s", result)
        else:
            logger.warning("Startup warmload incomplete: %s", result)
        return result

    # ── AI toggle API ─────────────────────────────────────────────────────────

    def toggle_ai_model(self, model_id: str) -> bool:
        """Toggle a specific AI model on/off. Returns new state."""
        with self.lock:
            current = self.ai_models.get(model_id, False)
            self.ai_models[model_id] = not current
            logger.info(f"AI model '{model_id}' → {'ON' if not current else 'OFF'}")
            return not current

    def set_ai_model(self, model_id: str, enabled: bool):
        with self.lock:
            self.ai_models[model_id] = enabled
        logger.info(f"AI model '{model_id}' set to {'ON' if enabled else 'OFF'}")

    def get_ai_status(self) -> dict[str, bool]:
        with self.lock:
            return dict(self.ai_models)

    def get_ai_capabilities(self) -> dict[str, dict]:
        """Per-model support metadata for UI and cloud/local deployment hints."""
        yolo_ok = self.model is not None
        return {
            "facial": {"supported": True, "mode": "browser_frames"},
            "crowd": {"supported": yolo_ok, "mode": "yolo" if yolo_ok else "unavailable"},
            "medical": {"supported": yolo_ok, "mode": "heuristic_fall" if yolo_ok else "unavailable"},
            "anomaly": {"supported": True, "mode": "optical_flow"},
            "fight": {"supported": yolo_ok, "mode": "heuristic_proximity" if yolo_ok else "unavailable"},
            "fire": {
                "supported": False,
                "mode": "not_implemented",
                "note": "Fire alerts use SOS/IOT integrations — not a CV model",
            },
        }

    # ── Internal helpers ──────────────────────────────────────────────────────

    def _release_camera_internal(self, cam_id: str):
        if cam_id in self.active_cameras:
            self.active_cameras[cam_id].release()
            del self.active_cameras[cam_id]
        self.camera_indices.pop(cam_id, None)
        self.latest_raw_frames.pop(cam_id, None)
        logger.info(f"Released camera {cam_id}")

    @staticmethod
    def _open_with_timeout(index, backend, timeout_s: float = 10.0):
        """Open a cv2.VideoCapture in a daemon thread with a timeout.
        Returns the cap if successful, None on timeout or failure."""
        result = [None]
        def _worker():
            try:
                cap = cv2.VideoCapture(index, backend)
                result[0] = cap
            except Exception:
                result[0] = None

        t = threading.Thread(target=_worker, daemon=True)
        t.start()
        t.join(timeout=timeout_s)
        if t.is_alive():
            logger.warning(f"cv2.VideoCapture({index}, ...) timed out after {timeout_s}s")
            return None
        return result[0]

    def _open_cap(self, index) -> cv2.VideoCapture | None:
        """Try backends in order: DSHOW (fastest on Windows) → MSMF → ANY.
        Called OUTSIDE the lock to avoid blocking the entire engine.
        Each cv2.VideoCapture attempt is wrapped in a 10-second timeout."""
        if isinstance(index, str) and not index.isdigit():
            backends = [("FFMPEG", cv2.CAP_FFMPEG, 15), ("ANY", cv2.CAP_ANY, 15)]
        else:
            index = int(index)
            backends = [
                ("ANY",   cv2.CAP_ANY,  10),
                ("MSMF",  cv2.CAP_MSMF, 12),
                ("DSHOW", cv2.CAP_DSHOW, 8),
            ]
        for name, backend, timeout in backends:
            try:
                logger.info(f"Trying camera index {index} via {name} (timeout {timeout}s)...")
                cap = self._open_with_timeout(index, backend, timeout_s=timeout)
                if cap is None:
                    continue
                if cap.isOpened():
                    # We avoid setting CAP_PROP properties here (width/height/fourcc)
                    # as it frequently breaks virtual cameras (like OBS or Samsung S23).
                    
                    # Wait for virtual camera pipelines (especially MSMF) to fully buffer
                    import time
                    time.sleep(1.5)
                    
                    # Verify we can actually read a frame (retry up to 5 times for virtual cams)
                    ret = False
                    for _ in range(5):
                        ret, _ = cap.read()
                        if ret:
                            break
                        time.sleep(0.5)
                        
                    if ret:
                        logger.info(f"Camera index {index} opened via {name} ✓")
                        return cap
                    else:
                        logger.warning(f"Camera index {index} opened via {name} but read() failed after retries")
                        cap.release()
                else:
                    cap.release()
            except Exception as e:
                logger.warning(f"Backend {name} failed for index {index}: {e}")
        return None

    # ── Public camera API ─────────────────────────────────────────────────────

    def set_camera(self, cam_id: str, index) -> bool:
        if isinstance(index, str):
            if index == "browser":
                pass
            elif index.isdigit():
                index = int(index)
            elif index.lower().startswith("rstp://"):
                index = "rtsp://" + index[7:]
            elif index.lower().startswith("rtsp://") and not index.startswith("rtsp://"):
                index = "rtsp://" + index[7:]

        # Check under lock if already open or already being opened
        with self.lock:
            if cam_id in self._pending_opens:
                logger.info(f"Camera {cam_id} is already being opened, skipping duplicate request")
                return False

            if cam_id in self.active_cameras and self.camera_indices.get(cam_id) == index:
                if self.active_cameras[cam_id].isOpened():
                    logger.info(f"Camera {cam_id} already open at index {index}")
                    return True

            # Release any OTHER cam_id that claims this index
            for other_id, other_idx in list(self.camera_indices.items()):
                if other_id != cam_id and other_idx == index:
                    logger.info(f"Releasing {other_id} (index {index}) → reassigning to {cam_id}")
                    self._release_camera_internal(other_id)

            # Release current capture for this cam_id
            if cam_id in self.active_cameras:
                self._release_camera_internal(cam_id)
            if index == "browser":
                self.camera_indices[cam_id] = index
                self.latest_raw_frames[cam_id] = None
                self.browser_annotated_frames[cam_id] = (0, None)
                logger.info(f"Camera {cam_id} → index {index} (Browser Feed) OK")
                return True

            self._pending_opens.add(cam_id)

        # Open camera OUTSIDE the main lock, but INSIDE a strict open_lock.
        # Windows DirectShow crashes/deadlocks if cv2.VideoCapture is called concurrently.
        logger.info(f"Opening camera {cam_id} at index {index}...")
        try:
            with self.open_lock:
                cap = self._open_cap(index)
        finally:
            with self.lock:
                self._pending_opens.discard(cam_id)

        with self.lock:

            if cap:
                self.active_cameras[cam_id] = cap
                self.camera_indices[cam_id] = index
                self.latest_raw_frames[cam_id] = None
                logger.info(f"Camera {cam_id} → index {index} OK")
                return True

            logger.error(f"All backends failed for {cam_id} at index {index}")
            return False

    def process_browser_frame(self, cam_id: str, base64_str: str):
        if not base64_str:
            return
        try:
            import base64
            encoded = base64_str.split(",")[1] if "," in base64_str else base64_str
            nparr = np.frombuffer(base64.b64decode(encoded), np.uint8)
            frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
            if frame is None:
                return

            with self.lock:
                self.latest_raw_frames[cam_id] = frame.copy()
                crowd_on     = self.ai_models.get("crowd", True)
                facial_on    = self.ai_models.get("facial", False)
                medical_on   = self.ai_models.get("medical", False)
                anomaly_on   = self.ai_models.get("anomaly", False)
                fight_on     = self.ai_models.get("fight", False)

            annotated = frame.copy()
            count = 0
            boxes = []

            if crowd_on or medical_on or anomaly_on or fight_on:
                count, annotated, boxes = self._run_yolo(annotated)

            if medical_on:
                annotated, _ = self._run_medical_fall(annotated, cam_id, boxes)
            if anomaly_on:
                annotated, _ = self._run_anomaly_heuristic(annotated, cam_id, boxes)
            if fight_on:
                annotated, _ = self._run_fight_heuristic(annotated, cam_id, boxes)

            if facial_on:
                fe = self.face_engine
                try:
                    if hasattr(fe, "match_all_faces"):
                        faces = fe.match_all_faces(annotated)
                    elif hasattr(fe, "recognize"):
                        faces = fe.recognize(annotated)
                    else:
                        faces = []
                except (AttributeError, TypeError) as e:
                    logger.warning("[WS] face match failed: %s — returning empty result", e)
                    faces = []
                with self.lock:
                    self.face_results[cam_id] = faces if isinstance(faces, list) else []
                if isinstance(faces, list):
                    annotated = self._annotate_faces(annotated, faces)
            else:
                with self.lock:
                    self.face_results[cam_id] = []

            with self.lock:
                self.browser_annotated_frames[cam_id] = (count, annotated)
        except Exception as e:
            logger.error(f"process_browser_frame error for {cam_id}: {e}")

    def release_camera(self, cam_id: str):
        with self.lock:
            self._release_camera_internal(cam_id)

    def set_camera_for_client(self, client_id: str, cam_id: str, index) -> bool:
        """Per-client camera registration — browser feeds do not steal from other clients."""
        if isinstance(index, str):
            if index == "browser":
                pass
            elif index.isdigit():
                index = int(index)
            elif index.lower().startswith("rstp://"):
                index = "rtsp://" + index[7:]
            elif index.lower().startswith("rtsp://") and not index.startswith("rtsp://"):
                index = "rtsp://" + index[7:]

        if index == "browser":
            with self.lock:
                self._client_cameras.setdefault(client_id, set()).add(cam_id)
                self._cam_refcounts[cam_id] = self._cam_refcounts.get(cam_id, 0) + 1
                self.camera_indices[cam_id] = "browser"
                self.latest_raw_frames[cam_id] = None
                self.browser_annotated_frames[cam_id] = (0, None)
            logger.info("Camera %s → index browser (client %s) OK", cam_id, client_id[:8])
            return True

        with self.lock:
            self._client_cameras.setdefault(client_id, set()).add(cam_id)
        return self.set_camera(cam_id, index)

    def release_camera_for_client(self, client_id: str, cam_id: str) -> None:
        with self.lock:
            client_cams = self._client_cameras.get(client_id)
            if not client_cams or cam_id not in client_cams:
                return
            client_cams.discard(cam_id)
            if not client_cams:
                self._client_cameras.pop(client_id, None)
            idx = self.camera_indices.get(cam_id)

        if idx == "browser":
            with self.lock:
                ref = self._cam_refcounts.get(cam_id, 1) - 1
                if ref <= 0:
                    self._cam_refcounts.pop(cam_id, None)
                    self._release_camera_internal(cam_id)
                else:
                    self._cam_refcounts[cam_id] = ref
            return

        with self.lock:
            other_holders = any(
                cam_id in cams for cid, cams in self._client_cameras.items() if cid != client_id
            )
        if not other_holders:
            self.release_camera(cam_id)

    def release_client_cameras(self, client_id: str) -> None:
        for cam_id in list(self._client_cameras.get(client_id, set())):
            self.release_camera_for_client(client_id, cam_id)

    # ── Frame processing ──────────────────────────────────────────────────────

    def _run_yolo(self, frame: np.ndarray) -> tuple[int, np.ndarray, list]:
        """Run crowd detection YOLO. Only called when crowd or medical model is enabled."""
        if frame is None or self.model is None:
            return 0, frame, []
        results = self.model(frame, verbose=False, classes=[0])
        count   = 0
        boxes_out = []
        for r in results:
            for box in r.boxes:
                count += 1
                x1, y1, x2, y2 = map(int, box.xyxy[0])
                boxes_out.append((x1, y1, x2, y2))
                cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv2.putText(frame, f"People: {count}", (10, 30),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 2)
        return count, frame, boxes_out

    def _run_medical_fall(self, frame: np.ndarray, cam_id: str, boxes: list) -> tuple[np.ndarray, bool]:
        fall_detected = False
        for (x1, y1, x2, y2) in boxes:
            w = x2 - x1
            h = y2 - y1
            if w > h * 1.5:  # Person is horizontal
                fall_detected = True
                cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 0, 255), 3)
                cv2.putText(frame, "FALL DETECTED", (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)
        
        if cam_id not in self.fall_state:
            self.fall_state[cam_id] = 0
            
        event_triggered = False
        if fall_detected:
            self.fall_state[cam_id] += 1
            if self.fall_state[cam_id] == 20:  # Stayed down for ~1.5 seconds
                event_key = f"{cam_id}_fall"
                now = datetime.now()
                last = self.last_event_ts.get(event_key, datetime.min)
                if (now - last).total_seconds() > 60:  # 1 min cooldown
                    logger.warning(f"MEDICAL EMERGENCY: Persistent fall detected on {cam_id}")
                    self.last_event_ts[event_key] = now
                    event_triggered = True
        else:
            self.fall_state[cam_id] = max(0, self.fall_state[cam_id] - 1)
            
        return frame, event_triggered

    def _run_stampede_flow(self, frame: np.ndarray, cam_id: str, orig_frame: np.ndarray) -> tuple[np.ndarray, bool]:
        gray = cv2.cvtColor(orig_frame, cv2.COLOR_BGR2GRAY)
        gray = cv2.resize(gray, (320, 240))
        
        if cam_id not in self.prev_gray:
            self.prev_gray[cam_id] = gray
            return frame, False
            
        prev = self.prev_gray[cam_id]
        flow = cv2.calcOpticalFlowFarneback(prev, gray, None, 0.5, 3, 15, 3, 5, 1.2, 0)
        self.prev_gray[cam_id] = gray
        
        mag, _ = cv2.cartToPolar(flow[..., 0], flow[..., 1])
        avg_mag = np.mean(mag)
        
        event_triggered = False
        if avg_mag > 2.8:  # High average motion magnitude
            cv2.putText(frame, "STAMPEDE / CHAOS DETECTED", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 165, 255), 2)
            event_key = f"{cam_id}_stampede"
            now = datetime.now()
            last = self.last_event_ts.get(event_key, datetime.min)
            if (now - last).total_seconds() > 60:
                logger.warning(f"STAMPEDE DETECTED on {cam_id}")
                self.last_event_ts[event_key] = now
                event_triggered = True
            
        return frame, event_triggered

    def _run_fight_heuristic(
        self, frame: np.ndarray, cam_id: str, boxes: list
    ) -> tuple[np.ndarray, bool]:
        """Experimental fight stub: flag when multiple persons are extremely close."""
        if len(boxes) < 2:
            return frame, False

        close_pairs = 0
        for i in range(len(boxes)):
            for j in range(i + 1, len(boxes)):
                x1a, y1a, x2a, y2a = boxes[i]
                x1b, y1b, x2b, y2b = boxes[j]
                cax, cay = (x1a + x2a) / 2, (y1a + y2a) / 2
                cbx, cby = (x1b + x2b) / 2, (y1b + y2b) / 2
                dist = ((cax - cbx) ** 2 + (cay - cby) ** 2) ** 0.5
                scale = max(1.0, ((x2a - x1a) + (x2b - x1b)) / 2)
                if dist < scale * 0.55:
                    close_pairs += 1
                    cv2.rectangle(frame, (x1a, y1a), (x2a, y2a), (0, 0, 255), 3)
                    cv2.rectangle(frame, (x1b, y1b), (x2b, y2b), (0, 0, 255), 3)

        if close_pairs < 1:
            return frame, False

        cv2.putText(
            frame, "FIGHT? (heuristic)", (10, 90),
            cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2,
        )
        event_key = f"{cam_id}_fight"
        now = datetime.now()
        last = self.last_event_ts.get(event_key, datetime.min)
        if (now - last).total_seconds() > 90:
            self.last_event_ts[event_key] = now
            logger.warning("FIGHT HEURISTIC triggered on %s", cam_id)
            return frame, True
        return frame, False

    def _annotate_faces(self, frame: np.ndarray, faces: list[dict]) -> np.ndarray:
        for f in faces:
            x1, y1, x2, y2 = f["bbox"]
            name_disp = f["name"]
            if name_disp.startswith("unknown_"):
                try:
                    num = name_disp.split("_")[1]
                    name_disp = f"Unknown #{num}"
                except IndexError:
                    pass
            label  = f'{name_disp} ({f["confidence"]:.2f})'
            is_unknown = f["name"].startswith("unknown")
            color  = (0, 140, 255) if is_unknown else (0, 255, 0)
            cv2.rectangle(frame, (x1, y1), (x2, y2), color, 2)
            cv2.putText(frame, label, (x1, max(y1 - 10, 15)),
                        cv2.FONT_HERSHEY_SIMPLEX, 0.55, color, 1)
        return frame

    def get_active_frames(self) -> tuple[dict[str, tuple[int, np.ndarray | None]], list[dict]]:
        """
        Read one frame from each active camera.
        Returns:
            (results_dict, list_of_new_alerts)
        """
        results: dict[str, tuple[int, np.ndarray | None]] = {}
        new_alerts = []

        with self.lock:
            browser_feeds = [cid for cid, idx in self.camera_indices.items() if idx == "browser"]
            if not self.active_cameras and not browser_feeds:
                return results, new_alerts
            cam_items    = list(self.active_cameras.items())
            crowd_on     = self.ai_models.get("crowd",  False)
            facial_on    = self.ai_models.get("facial", False)
            medical_on   = self.ai_models.get("medical", False)
            anomaly_on   = self.ai_models.get("anomaly", False)
            fight_on     = self.ai_models.get("fight", False)

        # Periodic status log
        if not hasattr(self, "_last_status_log"): self._last_status_log = 0
        self._last_status_log += 1
        if self._last_status_log % 50 == 0:
            logger.info(f"Vision Engine AI Status: Crowd={crowd_on}, Facial={facial_on}, Medical={medical_on}")

        for cam_id, cap in cam_items:
            try:
                ret, frame = cap.read()
            except Exception as e:
                logger.warning(f"cap.read() error for {cam_id}: {e}")
                ret, frame = False, None

            if not ret or frame is None:
                results[cam_id] = (0, None)
                continue

            # Store raw frame for missing-person search
            with self.lock:
                self.latest_raw_frames[cam_id] = frame.copy()

            annotated = frame.copy()
            count     = 0

            # ── Crowd, medical & fight detection (YOLO based) ─────────────────
            boxes = []
            if crowd_on or medical_on or fight_on:
                count, annotated, boxes = self._run_yolo(annotated)

            if fight_on:
                annotated, fight_trig = self._run_fight_heuristic(annotated, cam_id, boxes)
                if fight_trig:
                    new_alerts.append({
                        "type": "fight",
                        "severity": "high",
                        "location": cam_id,
                        "message": (
                            f"POSSIBLE ALTERCATION (heuristic): close proximity of multiple "
                            f"persons on {cam_id} — verify manually"
                        ),
                    })

            if medical_on:
                annotated, fall_trig = self._run_medical_fall(annotated, cam_id, boxes)
                if fall_trig:
                    new_alerts.append({
                        "type": "medical",
                        "severity": "critical",
                        "location": cam_id,
                        "message": f"WARNING: MEDICAL EMERGENCY: Persistent fall detected on {cam_id}"
                    })
                
            # ── Stampede detection (Optical Flow based) ───────────────────────
            if anomaly_on:
                annotated, stampede_trig = self._run_stampede_flow(annotated, cam_id, frame)
                if stampede_trig:
                    new_alerts.append({
                        "type": "stampede",
                        "severity": "critical",
                        "location": cam_id,
                        "message": f"WARNING: CROWD ANOMALY: Stampede/chaos detected on {cam_id}"
                    })

            # ── Face recognition ──────────────────────────────────────────────
            faces: list[dict] = []
            if facial_on:
                fe = self.face_engine
                if hasattr(fe, "match_all_faces"):
                    faces = fe.match_all_faces(frame)
                elif hasattr(fe, "recognize"):
                    faces = fe.recognize(frame)
                else:
                    faces = []
                
                # Extract full frame as thumbnail for context
                for f in faces:
                    # We send a small version of the FULL frame instead of just the face crop.
                    # This allows the user to see exactly where the person was in the frame.
                    small_frame = cv2.resize(frame, (640, int(640 * frame.shape[0] / frame.shape[1])))
                    _, buf = cv2.imencode(".jpg", small_frame, [int(cv2.IMWRITE_JPEG_QUALITY), 65])
                    f["thumbnail"] = f"data:image/jpeg;base64,{base64.b64encode(buf).decode()}"

                annotated = self._annotate_faces(annotated, faces)

                # Push detections into gossip_bridge (no camera opened there)
                names  = [f["name"] for f in faces]
                bboxes = [tuple(f["bbox"]) for f in faces]
                if names:
                    gossip_bridge.on_detections(cam_id, names, bboxes, frame.shape[1])

            self.face_results[cam_id] = faces
            results[cam_id]           = (count, annotated)

        # Merge browser feeds
        with self.lock:
            for cid in [k for k, v in self.camera_indices.items() if v == "browser"]:
                if cid in self.browser_annotated_frames:
                    results[cid] = self.browser_annotated_frames[cid]

        return results, new_alerts

    def get_face_results(self) -> dict[str, list[dict]]:
        return dict(self.face_results)

    def search_missing_person(self, query_frame: np.ndarray) -> dict:
        fe = self.face_engine
        if hasattr(fe, "search_missing_person") and not hasattr(fe, "match_all_faces"):
            with self.lock:
                live_frames = {k: v for k, v in self.latest_raw_frames.items() if v is not None}
            return fe.search_missing_person(query_frame, live_frames)
        if hasattr(fe, "ensure_db"):
            fe.ensure_db()
        db_match = dict(fe.match_frame(query_frame))
        if db_match.get("found"):
            db_match.setdefault("search_mode", "database")
            return db_match
        try:
            import face_live_search

            live = face_live_search.search_query_in_live_feeds(query_frame, fe, self)
            if live.get("found"):
                return live
        except Exception as exc:
            logger.debug("live search fallback skipped: %s", exc)
        db_match.setdefault("search_mode", "database")
        return db_match

    # ── Legacy: base64 single-frame processing ────────────────────────────────

    def process_frame(self, base64_string: str) -> tuple[int, str]:
        try:
            encoded = base64_string.split(",")[1] if "," in base64_string else base64_string
            nparr   = np.frombuffer(base64.b64decode(encoded), np.uint8)
            frame   = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
            count, annotated, _ = self._run_yolo(frame.copy()) if self.model else (0, frame, [])
            if annotated is None:
                return 0, base64_string
            _, buf = cv2.imencode(".jpg", annotated, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
            return count, f"data:image/jpeg;base64,{base64.b64encode(buf).decode()}"
        except Exception as e:
            logger.error(f"process_frame error: {e}")
            return 0, base64_string