feat: 实现 Brain 匿名检测与单路跟踪 (#14)

合入 dev，#13 保持待验收。

Brain/src/yovision_brain/vision/__init__.py

@@ -0,0 +1,16 @@
+"""Anonymous detection and single-stream tracking."""
+
+from .detector import LumaBlobDetector, TorchLumaBlobDetector
+from .models import BoundingBox, Detection, Detector, DetectorMetadata, TrackedObject
+from .tracker import SingleStreamTracker
+
+__all__ = [
+    "BoundingBox",
+    "Detection",
+    "Detector",
+    "DetectorMetadata",
+    "LumaBlobDetector",
+    "SingleStreamTracker",
+    "TorchLumaBlobDetector",
+    "TrackedObject",
+]

Brain/src/yovision_brain/vision/detector.py

@@ -0,0 +1,120 @@
+"""Deterministic anonymous blob detectors with no biometric semantics."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from yovision_brain.decode import DecodedFrame, DecoderError
+
+from .models import BoundingBox, Detection, DetectorMetadata
+
+_METADATA = DetectorMetadata(
+    name="yovision-luma-blob",
+    version="1.0.0",
+    source="YoVision Brain first-party deterministic algorithm",
+    license="No external model license; no learned weights are distributed",
+    weights="none",
+)
+
+
+def _components(mask: Sequence[Sequence[bool]], minimum_area: int) -> tuple[BoundingBox, ...]:
+    height = len(mask)
+    width = len(mask[0]) if height else 0
+    visited: set[tuple[int, int]] = set()
+    boxes: list[BoundingBox] = []
+    for y in range(height):
+        for x in range(width):
+            if not mask[y][x] or (x, y) in visited:
+                continue
+            pending = [(x, y)]
+            visited.add((x, y))
+            points: list[tuple[int, int]] = []
+            while pending:
+                current_x, current_y = pending.pop()
+                points.append((current_x, current_y))
+                for neighbor in (
+                    (current_x - 1, current_y),
+                    (current_x + 1, current_y),
+                    (current_x, current_y - 1),
+                    (current_x, current_y + 1),
+                ):
+                    nx, ny = neighbor
+                    if 0 <= nx < width and 0 <= ny < height and mask[ny][nx] and neighbor not in visited:
+                        visited.add(neighbor)
+                        pending.append(neighbor)
+            if len(points) >= minimum_area:
+                xs, ys = zip(*points)
+                boxes.append(BoundingBox(min(xs), min(ys), max(xs) + 1, max(ys) + 1))
+    return tuple(sorted(boxes, key=lambda box: (box.top, box.left, box.bottom, box.right)))
+
+
+def _validate_frame(frame: DecodedFrame) -> None:
+    if frame.pixel_format not in {"rgb24", "yuv444p"}:
+        raise DecoderError(f"anonymous detector does not support pixel format {frame.pixel_format!r}")
+    expected = frame.width * frame.height * 3
+    if len(frame.payload) != expected:
+        raise DecoderError(f"vision frame has {len(frame.payload)} bytes; expected {expected}")
+
+
+class LumaBlobDetector:
+    """Small CPU reference detector used for deterministic integration tests."""
+
+    metadata = _METADATA
+
+    def __init__(self, *, threshold: int = 200, minimum_area: int = 1) -> None:
+        if not 0 <= threshold <= 255 or minimum_area < 1:
+            raise ValueError("invalid luma detector threshold or minimum area")
+        self._threshold = threshold
+        self._minimum_area = minimum_area
+
+    def detect(self, frame: DecodedFrame) -> tuple[Detection, ...]:
+        _validate_frame(frame)
+        if frame.pixel_format == "rgb24":
+            pixels = [
+                max(frame.payload[index : index + 3])
+                for index in range(0, len(frame.payload), 3)
+            ]
+        else:
+            pixels = list(frame.payload[: frame.width * frame.height])
+        mask = [
+            [pixels[y * frame.width + x] >= self._threshold for x in range(frame.width)]
+            for y in range(frame.height)
+        ]
+        return tuple(
+            Detection(box=box, category="anonymous_target", confidence=1.0)
+            for box in _components(mask, self._minimum_area)
+        )
+
+
+class TorchLumaBlobDetector:
+    """PyTorch CPU/GPU smoke backend; it contains no external model weights."""
+
+    metadata = DetectorMetadata(
+        name="yovision-torch-luma-blob",
+        version="1.0.0",
+        source="YoVision Brain first-party PyTorch tensor implementation",
+        license="PyTorch BSD-3-Clause; no external model weights",
+        weights="none",
+    )
+
+    def __init__(self, *, threshold: int = 200, minimum_area: int = 1, device: str = "cpu") -> None:
+        self._threshold = threshold
+        self._minimum_area = minimum_area
+        self._device = device
+
+    def detect(self, frame: DecodedFrame) -> tuple[Detection, ...]:
+        _validate_frame(frame)
+        try:
+            import torch
+        except ImportError as exc:
+            raise RuntimeError("PyTorch runtime is required for TorchLumaBlobDetector") from exc
+        values = torch.tensor(list(frame.payload), dtype=torch.uint8, device=self._device)
+        if frame.pixel_format == "rgb24":
+            luma = values.reshape(frame.height, frame.width, 3).amax(dim=2)
+        else:
+            luma = values[: frame.width * frame.height].reshape(frame.height, frame.width)
+        mask = (luma >= self._threshold).cpu().tolist()
+        return tuple(
+            Detection(box=box, category="anonymous_target", confidence=1.0)
+            for box in _components(mask, self._minimum_area)
+        )

Brain/src/yovision_brain/vision/models.py

@@ -0,0 +1,52 @@
+"""Privacy-preserving vision ports and observations."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Protocol
+
+from yovision_brain.decode import DecodedFrame
+
+
+@dataclass(frozen=True, slots=True)
+class DetectorMetadata:
+    name: str
+    version: str
+    source: str
+    license: str
+    weights: str
+
+
+@dataclass(frozen=True, slots=True)
+class BoundingBox:
+    left: int
+    top: int
+    right: int
+    bottom: int
+
+    @property
+    def area(self) -> int:
+        return max(0, self.right - self.left) * max(0, self.bottom - self.top)
+
+
+@dataclass(frozen=True, slots=True)
+class Detection:
+    box: BoundingBox
+    category: str
+    confidence: float
+
+
+@dataclass(frozen=True, slots=True)
+class TrackedObject:
+    track_id: str
+    box: BoundingBox
+    category: str
+    confidence: float
+    frame_sequence: int
+    timestamp_ns: int
+
+
+class Detector(Protocol):
+    metadata: DetectorMetadata
+
+    def detect(self, frame: DecodedFrame) -> tuple[Detection, ...]: ...

Brain/src/yovision_brain/vision/tracker.py

@@ -0,0 +1,83 @@
+"""Session-local single-stream IoU tracker."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from .models import BoundingBox, Detection, TrackedObject
+
+
+def _iou(first: BoundingBox, second: BoundingBox) -> float:
+    intersection = BoundingBox(
+        max(first.left, second.left),
+        max(first.top, second.top),
+        min(first.right, second.right),
+        min(first.bottom, second.bottom),
+    ).area
+    union = first.area + second.area - intersection
+    return intersection / union if union else 0.0
+
+
+@dataclass(slots=True)
+class _Track:
+    track_id: str
+    detection: Detection
+    missed: int = 0
+
+
+class SingleStreamTracker:
+    """Tracks anonymous boxes only within one process session and one stream."""
+
+    def __init__(self, *, iou_threshold: float = 0.2, max_missed: int = 2) -> None:
+        if not 0.0 <= iou_threshold <= 1.0 or max_missed < 0:
+            raise ValueError("invalid tracker threshold or missed-frame limit")
+        self._iou_threshold = iou_threshold
+        self._max_missed = max_missed
+        self._tracks: dict[str, _Track] = {}
+        self._next_id = 1
+
+    def update(
+        self,
+        detections: tuple[Detection, ...],
+        *,
+        frame_sequence: int,
+        timestamp_ns: int,
+    ) -> tuple[TrackedObject, ...]:
+        unmatched_tracks = set(self._tracks)
+        results: list[TrackedObject] = []
+        for detection in detections:
+            candidates = [
+                (track_id, _iou(self._tracks[track_id].detection.box, detection.box))
+                for track_id in unmatched_tracks
+                if self._tracks[track_id].detection.category == detection.category
+            ]
+            track_id, score = max(candidates, key=lambda item: item[1], default=("", -1.0))
+            if score < self._iou_threshold:
+                track_id = f"track-{self._next_id:06d}"
+                self._next_id += 1
+                self._tracks[track_id] = _Track(track_id, detection)
+            else:
+                unmatched_tracks.remove(track_id)
+                self._tracks[track_id].detection = detection
+                self._tracks[track_id].missed = 0
+            results.append(
+                TrackedObject(
+                    track_id=track_id,
+                    box=detection.box,
+                    category=detection.category,
+                    confidence=detection.confidence,
+                    frame_sequence=frame_sequence,
+                    timestamp_ns=timestamp_ns,
+                )
+            )
+        for track_id in unmatched_tracks:
+            track = self._tracks[track_id]
+            track.missed += 1
+            if track.missed > self._max_missed:
+                del self._tracks[track_id]
+        return tuple(results)
+
+    def finish(self) -> tuple[str, ...]:
+        ended = tuple(sorted(self._tracks))
+        self._tracks.clear()
+        return ended

Brain/tests/fixtures/vision/README.md

@@ -0,0 +1,5 @@
+# Brain vision fixtures
+
+Vision tests create anonymous geometric RGB frames in memory. Never add faces,
+customer recordings, biometric templates, camera credentials, or unreviewed
+model weights to this directory.

Brain/tests/vision/test_detection_tracking.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+import pytest
+
+from yovision_brain.decode import DecodedFrame
+from yovision_brain.vision import (
+    BoundingBox,
+    Detection,
+    LumaBlobDetector,
+    SingleStreamTracker,
+    TorchLumaBlobDetector,
+)
+
+
+def frame(payload: bytes, *, sequence: int = 0, width: int = 4, height: int = 3) -> DecodedFrame:
+    return DecodedFrame(sequence, sequence * 40_000_000, "camera", "main", width, height, "rgb24", payload)
+
+
+def rgb(values: list[int]) -> bytes:
+    return b"".join(bytes((value, value, value)) for value in values)
+
+
+def detection(left: int, top: int, right: int, bottom: int) -> Detection:
+    return Detection(BoundingBox(left, top, right, bottom), "anonymous_target", 0.9)
+
+
+def test_detector_emits_only_anonymous_observations() -> None:
+    payload = rgb([0, 255, 255, 0, 0, 255, 255, 0, 0, 0, 0, 0])
+    result = LumaBlobDetector(minimum_area=2).detect(frame(payload))
+    assert result == (Detection(BoundingBox(1, 0, 3, 2), "anonymous_target", 1.0),)
+    assert LumaBlobDetector.metadata.weights == "none"
+    assert "external model license" in LumaBlobDetector.metadata.license
+
+
+def test_empty_frame_has_no_detection() -> None:
+    assert LumaBlobDetector().detect(frame(rgb([0] * 12))) == ()
+
+
+def test_tracker_keeps_session_id_across_motion_and_short_occlusion() -> None:
+    tracker = SingleStreamTracker(iou_threshold=0.1, max_missed=2)
+    first = tracker.update((detection(0, 0, 3, 3),), frame_sequence=0, timestamp_ns=0)
+    assert first[0].track_id == "track-000001"
+    assert tracker.update((), frame_sequence=1, timestamp_ns=1) == ()
+    resumed = tracker.update((detection(1, 0, 4, 3),), frame_sequence=2, timestamp_ns=2)
+    assert resumed[0].track_id == "track-000001"
+    assert tracker.finish() == ("track-000001",)
+
+
+def test_disappeared_track_ends_and_new_target_gets_new_id() -> None:
+    tracker = SingleStreamTracker(max_missed=1)
+    first = tracker.update((detection(0, 0, 2, 2),), frame_sequence=0, timestamp_ns=0)
+    tracker.update((), frame_sequence=1, timestamp_ns=1)
+    tracker.update((), frame_sequence=2, timestamp_ns=2)
+    second = tracker.update((detection(0, 0, 2, 2),), frame_sequence=3, timestamp_ns=3)
+    assert first[0].track_id == "track-000001"
+    assert second[0].track_id == "track-000002"
+
+
+def test_track_ids_are_session_local() -> None:
+    one = SingleStreamTracker().update((detection(0, 0, 1, 1),), frame_sequence=0, timestamp_ns=0)
+    two = SingleStreamTracker().update((detection(0, 0, 1, 1),), frame_sequence=0, timestamp_ns=0)
+    assert one[0].track_id == two[0].track_id == "track-000001"
+
+
+def test_torch_backend_cpu_smoke_uses_no_external_weights() -> None:
+    pytest.importorskip("torch")
+    result = TorchLumaBlobDetector().detect(frame(rgb([0, 255] + [0] * 10)))
+    assert result[0].category == "anonymous_target"
+    assert TorchLumaBlobDetector.metadata.weights == "none"