feat(brain): add single-stream visual prototype (T-017)

Brain/yovision_brain/source.py

@@ -0,0 +1,201 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Iterable
+from urllib.parse import urlsplit
+
+from .domain import Box, Detection
+
+try:
+    import cv2  # type: ignore
+    import numpy as np  # type: ignore
+except ImportError:  # pragma: no cover - exercised by the startup failure path
+    cv2 = None
+    np = None
+
+
+@dataclass(frozen=True)
+class FramePacket:
+    frame: Any
+    captured_at: datetime
+    scripted_detections: tuple[Detection, ...] | None
+
+
+def require_opencv() -> None:
+    if cv2 is None or np is None:
+        raise RuntimeError("Brain demo requires the pinned NumPy and OpenCV packages")
+
+
+def read_stream_url(path_value: str) -> str:
+    path = Path(path_value)
+    if not path.is_absolute():
+        raise ValueError("stream URL file must be an absolute external path")
+    try:
+        resolved = path.resolve(strict=True)
+    except OSError as exc:
+        raise ValueError("stream URL file does not exist") from exc
+    if not resolved.is_file():
+        raise ValueError("stream URL file does not exist")
+    repository_root = Path(__file__).resolve().parents[2]
+    try:
+        resolved.relative_to(repository_root)
+    except ValueError:
+        pass
+    else:
+        raise ValueError("stream URL file must be outside the repository")
+    try:
+        with resolved.open("rb") as stream:
+            raw = stream.read(4097)
+    except OSError as exc:
+        raise ValueError("stream URL file cannot be read") from exc
+    if len(raw) > 4096:
+        raise ValueError("stream URL file exceeds 4096 bytes")
+    try:
+        lines = raw.decode("utf-8").splitlines()
+    except UnicodeError as exc:
+        raise ValueError("stream URL file must be UTF-8") from exc
+    values = [line.strip() for line in lines if line.strip()]
+    if len(values) != 1:
+        raise ValueError("stream URL file must contain exactly one non-empty line")
+    parsed = urlsplit(values[0])
+    if parsed.scheme not in {"rtsp", "rtsps"} or not parsed.hostname:
+        raise ValueError("stream URL must be an RTSP URL")
+    return values[0]
+
+
+class SyntheticSource:
+    mode = "synthetic"
+    label = "合成回放"
+    fixture = True
+    source_ref = "demo-camera-01"
+
+    def __init__(self, width: int = 960, height: int = 540) -> None:
+        require_opencv()
+        self.width = width
+        self.height = height
+        self._sequence = 0
+
+    def read(self) -> FramePacket:
+        self._sequence += 1
+        frame = np.zeros((self.height, self.width, 3), dtype=np.uint8)
+        frame[:] = (20, 28, 42)
+        cv2.rectangle(frame, (0, int(self.height * 0.72)), (self.width, self.height), (31, 42, 58), -1)
+        for x in range(0, self.width, 80):
+            cv2.line(frame, (x, int(self.height * 0.72)), (x + 80, self.height), (43, 57, 75), 1)
+        cv2.putText(frame, "SYNTHETIC FIXTURE - NOT MODEL OUTPUT", (24, 38), cv2.FONT_HERSHEY_SIMPLEX, 0.72, (82, 190, 245), 2)
+
+        phase = ((self._sequence - 1) % 160) / 159.0
+        center_x = -0.04 + phase * 1.08
+        x1 = max(0.0, center_x - 0.04)
+        x2 = min(1.0, center_x + 0.04)
+        detections: tuple[Detection, ...] = ()
+        if x2 - x1 > 0.01:
+            box = Box(x1, 0.34, x2, 0.82)
+            detections = (Detection("P-DEMO-001", "person", box, None),)
+            px = int(center_x * self.width)
+            head_y = int(self.height * 0.40)
+            cv2.circle(frame, (px, head_y), 17, (195, 210, 225), -1)
+            cv2.line(frame, (px, head_y + 18), (px, int(self.height * 0.65)), (195, 210, 225), 12)
+            cv2.line(frame, (px, int(self.height * 0.52)), (px - 30, int(self.height * 0.60)), (195, 210, 225), 8)
+            cv2.line(frame, (px, int(self.height * 0.52)), (px + 30, int(self.height * 0.60)), (195, 210, 225), 8)
+            cv2.line(frame, (px, int(self.height * 0.65)), (px - 24, int(self.height * 0.79)), (195, 210, 225), 9)
+            cv2.line(frame, (px, int(self.height * 0.65)), (px + 24, int(self.height * 0.79)), (195, 210, 225), 9)
+        return FramePacket(frame, datetime.now(timezone.utc), detections)
+
+    def close(self) -> None:
+        return
+
+
+class StreamSource:
+    mode = "stream"
+    label = "外部 MediaMTX / RTSP"
+    fixture = False
+    source_ref = "configured-video-source"
+
+    def __init__(self, stream_url: str) -> None:
+        require_opencv()
+        self._stream_url = stream_url
+        self._capture: Any = None
+
+    def _open(self) -> None:
+        if self._capture is not None:
+            self._capture.release()
+        self._capture = cv2.VideoCapture(self._stream_url)
+        self._capture.set(cv2.CAP_PROP_BUFFERSIZE, 1)
+
+    def read(self) -> FramePacket:
+        if self._capture is None or not self._capture.isOpened():
+            self._open()
+        ok, frame = self._capture.read()
+        if not ok or frame is None:
+            self._open()
+            raise RuntimeError("stream frame unavailable")
+        return FramePacket(frame, datetime.now(timezone.utc), None)
+
+    def close(self) -> None:
+        if self._capture is not None:
+            self._capture.release()
+            self._capture = None
+
+
+class HOGPersonDetector:
+    name = "opencv_hog_person_demo"
+    production_ready = False
+
+    def __init__(self) -> None:
+        require_opencv()
+        self._hog = cv2.HOGDescriptor()
+        self._hog.setSVMDetector(cv2.HOGDescriptor_getDefaultPeopleDetector())
+
+    def detect(self, frame: Any) -> list[tuple[Box, float]]:
+        height, width = frame.shape[:2]
+        scale = min(1.0, 960.0 / max(width, 1))
+        working = frame if scale == 1.0 else cv2.resize(frame, (int(width * scale), int(height * scale)))
+        boxes, weights = self._hog.detectMultiScale(working, winStride=(8, 8), padding=(8, 8), scale=1.05)
+        result: list[tuple[Box, float]] = []
+        work_height, work_width = working.shape[:2]
+        for raw_box, weight in zip(boxes, weights):
+            x, y, box_width, box_height = (int(value) for value in raw_box)
+            x1 = max(0.0, min(1.0, x / work_width))
+            y1 = max(0.0, min(1.0, y / work_height))
+            x2 = max(0.0, min(1.0, (x + box_width) / work_width))
+            y2 = max(0.0, min(1.0, (y + box_height) / work_height))
+            if x2 > x1 and y2 > y1:
+                result.append((Box(x1, y1, x2, y2), float(weight)))
+        return result
+
+
+class CentroidTracker:
+    def __init__(self, max_distance: float = 0.18, ttl_frames: int = 8) -> None:
+        self._max_distance = max_distance
+        self._ttl_frames = ttl_frames
+        self._next_id = 1
+        self._tracks: dict[str, tuple[Box, int]] = {}
+
+    def update(self, boxes: Iterable[tuple[Box, float]], sequence: int) -> list[Detection]:
+        incoming = list(boxes)
+        available = set(self._tracks)
+        detections: list[Detection] = []
+        for box, score in incoming:
+            center = box.center
+            selected: str | None = None
+            selected_distance = self._max_distance
+            for track_id in available:
+                old_center = self._tracks[track_id][0].center
+                distance = ((center.x - old_center.x) ** 2 + (center.y - old_center.y) ** 2) ** 0.5
+                if distance < selected_distance:
+                    selected = track_id
+                    selected_distance = distance
+            if selected is None:
+                selected = f"P-{self._next_id:04d}"
+                self._next_id += 1
+            else:
+                available.remove(selected)
+            self._tracks[selected] = (box, sequence)
+            detections.append(Detection(selected, "person", box, score))
+        for track_id, (_, last_seen) in list(self._tracks.items()):
+            if sequence - last_seen > self._ttl_frames:
+                del self._tracks[track_id]
+        return detections