2026-03-03 06:46:08 -05:00
7 changed files with 1030 additions and 0 deletions
--- a/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_person_reid.py
+++ b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_person_reid.py
@ -0,0 +1,259 @@
+"""
+_person_reid.py — Appearance-based person re-identification (no ROS2 deps).
+
+Algorithm
+---------
+For each person detection (bounding-box crop from a camera image):
+
+1. Extract a 2-D HS colour histogram:
+     H axis: 16 bins over [0, 180]   (OpenCV hue)
+     S axis:  8 bins over [0, 256]   (saturation)
+   The Value channel is omitted for illumination robustness.
+   Result: 128-dimensional float32 vector, L2-normalised.
+
+2. Match against the gallery of active TrackEntry objects using cosine
+   similarity (= dot product of L2-normalised vectors).  The best match
+   whose similarity exceeds `threshold` is selected; otherwise a new track
+   is created.
+
+3. On a successful match the gallery entry is updated with an exponential
+   moving average (α = 0.3) of the new feature to adapt gradually to
+   appearance changes (lighting, pose) while staying robust to outliers.
+
+4. Tracks not updated for `timeout_s` wall-clock seconds are flagged
+   `is_stale`; stale tracks are pruned from the gallery on the next cycle.
+
+Public API
+----------
+  H_BINS, S_BINS, FEAT_DIM             histogram shape constants
+  TrackEntry                           mutable dataclass for one active track
+  extract_hsv_histogram(bgr_crop)   -> np.ndarray (FEAT_DIM,) L2-normalised
+  cosine_similarity(a, b)           -> float in [-1, 1]
+  match_track(feature, tracks, threshold)
+                                    -> (track_id, similarity) | (None, 0.0)
+  TrackGallery                         mutable store of active tracks
+    .add(camera_id, feature, bbox, now)   -> int  (new track_id)
+    .update(track_id, camera_id, feature, bbox, confidence, now)
+    .match(feature, threshold)        -> (track_id, similarity) | (None, 0.0)
+    .mark_stale(now, timeout_s)
+    .prune_stale()
+    .tracks                           -> Dict[int, TrackEntry]
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass, field
+from typing import Dict, Optional, Tuple
+
+import cv2
+import numpy as np
+
+
+# ── Histogram constants ───────────────────────────────────────────────────────
+
+H_BINS:   int = 16
+S_BINS:   int = 8
+FEAT_DIM: int = H_BINS * S_BINS   # 128
+
+_EMA_ALPHA: float = 0.3   # weight of newest observation in EMA update
+
+
+# ── Track entry ───────────────────────────────────────────────────────────────
+
+@dataclass
+class TrackEntry:
+    """Mutable state for one active person track."""
+
+    track_id:   int
+    camera_id:  str
+    feature:    np.ndarray              # (FEAT_DIM,) L2-normalised appearance
+    confidence: float                   # last match similarity (1.0 for new)
+    first_seen: float                   # time.time() wall-clock seconds
+    last_seen:  float                   # time.time() wall-clock seconds
+    bbox:       Tuple[float, float, float, float] = (0.0, 0.0, 0.0, 0.0)
+    # (centre_x, centre_y, size_x, size_y) in image pixels
+    is_stale:   bool = False
+
+
+# ── Feature extraction ────────────────────────────────────────────────────────
+
+def extract_hsv_histogram(bgr_crop: np.ndarray) -> np.ndarray:
+    """
+    Extract a 2-D HS colour histogram from a BGR image crop.
+
+    The Value channel is excluded for illumination robustness.
+    The result is L2-normalised; a zero-area or empty crop returns a zero vector.
+
+    Parameters
+    ----------
+    bgr_crop : uint8 BGR image, any size ≥ 1×1.
+
+    Returns
+    -------
+    np.ndarray, shape (FEAT_DIM,) = (128,), dtype float32, L2-normalised.
+    """
+    if bgr_crop is None or bgr_crop.size == 0:
+        return np.zeros(FEAT_DIM, dtype=np.float32)
+
+    hsv  = cv2.cvtColor(bgr_crop, cv2.COLOR_BGR2HSV)
+    hist = cv2.calcHist(
+        [hsv], [0, 1], None,
+        [H_BINS, S_BINS],
+        [0, 180, 0, 256],
+    )
+    flat = hist.flatten().astype(np.float32)
+    norm = np.linalg.norm(flat)
+    if norm > 0.0:
+        flat /= norm
+    return flat
+
+
+# ── Similarity ────────────────────────────────────────────────────────────────
+
+def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
+    """
+    Cosine similarity between two vectors.
+
+    Works correctly for both unit-normalised and unnormalised inputs.
+    Returns 0.0 when either vector has (near-)zero norm.
+
+    Returns
+    -------
+    float in [-1, 1].
+    """
+    na = float(np.linalg.norm(a))
+    nb = float(np.linalg.norm(b))
+    if na < 1e-12 or nb < 1e-12:
+        return 0.0
+    return float(np.dot(a, b) / (na * nb))
+
+
+# ── Gallery matching ──────────────────────────────────────────────────────────
+
+def match_track(
+    feature:   np.ndarray,
+    tracks:    Dict[int, TrackEntry],
+    threshold: float,
+) -> Tuple[Optional[int], float]:
+    """
+    Find the gallery entry with the highest cosine similarity to `feature`.
+
+    Stale tracks are excluded from matching.
+
+    Parameters
+    ----------
+    feature   : query appearance vector (FEAT_DIM,)
+    tracks    : gallery dict keyed by track_id
+    threshold : minimum cosine similarity required to accept a match
+
+    Returns
+    -------
+    (track_id, similarity) if a match is found, else (None, 0.0).
+    """
+    best_id:  Optional[int] = None
+    best_sim: float = threshold  # strict: new track if sim == threshold exactly
+
+    for tid, entry in tracks.items():
+        if entry.is_stale:
+            continue
+        sim = cosine_similarity(feature, entry.feature)
+        if sim > best_sim:
+            best_sim = sim
+            best_id  = tid
+
+    if best_id is not None:
+        return best_id, best_sim
+    return None, 0.0
+
+
+# ── Gallery ───────────────────────────────────────────────────────────────────
+
+class TrackGallery:
+    """
+    Mutable store of active person tracks.
+
+    Thread-safety: not thread-safe; call from a single ROS2 callback thread.
+    """
+
+    def __init__(self) -> None:
+        self._tracks:  Dict[int, TrackEntry] = {}
+        self._next_id: int = 1
+
+    @property
+    def tracks(self) -> Dict[int, TrackEntry]:
+        return self._tracks
+
+    def add(
+        self,
+        camera_id: str,
+        feature:   np.ndarray,
+        bbox:      Tuple[float, float, float, float],
+        now:       float,
+    ) -> int:
+        """Create a new track and return its assigned track_id."""
+        tid = self._next_id
+        self._next_id += 1
+        self._tracks[tid] = TrackEntry(
+            track_id=tid,
+            camera_id=camera_id,
+            feature=feature.copy(),
+            confidence=1.0,
+            first_seen=now,
+            last_seen=now,
+            bbox=bbox,
+        )
+        return tid
+
+    def update(
+        self,
+        track_id:   int,
+        camera_id:  str,
+        feature:    np.ndarray,
+        bbox:       Tuple[float, float, float, float],
+        confidence: float,
+        now:        float,
+    ) -> None:
+        """
+        Update an existing track with a new observation.
+
+        The appearance feature is blended via EMA (α=0.3) for robustness.
+        """
+        entry = self._tracks[track_id]
+
+        # EMA blend: α * new + (1-α) * old, then re-normalise
+        updated = _EMA_ALPHA * feature + (1.0 - _EMA_ALPHA) * entry.feature
+        norm = np.linalg.norm(updated)
+        if norm > 0.0:
+            updated /= norm
+
+        self._tracks[track_id] = TrackEntry(
+            track_id=track_id,
+            camera_id=camera_id,
+            feature=updated,
+            confidence=confidence,
+            first_seen=entry.first_seen,
+            last_seen=now,
+            bbox=bbox,
+            is_stale=False,
+        )
+
+    def match(
+        self,
+        feature:   np.ndarray,
+        threshold: float,
+    ) -> Tuple[Optional[int], float]:
+        """Delegate to module-level match_track."""
+        return match_track(feature, self._tracks, threshold)
+
+    def mark_stale(self, now: float, timeout_s: float) -> None:
+        """Flag tracks whose last_seen age exceeds timeout_s."""
+        for entry in self._tracks.values():
+            if (now - entry.last_seen) > timeout_s:
+                entry.is_stale = True
+
+    def prune_stale(self) -> None:
+        """Remove all stale tracks from the gallery."""
+        self._tracks = {
+            tid: e for tid, e in self._tracks.items() if not e.is_stale
+        }
--- a/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/person_reid_node.py
+++ b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/person_reid_node.py
@ -0,0 +1,215 @@
+"""
+person_reid_node.py — Appearance-based person re-identification (Issue #322).
+
+Subscribes to a camera RGB stream and person detection bounding boxes,
+extracts per-person HSV colour histograms, matches against an in-memory
+gallery using cosine similarity, and publishes stable PersonTrack IDs.
+
+Running one instance per camera (remapping /camera/color/image_raw and
+/saltybot/scene/objects) achieves cross-camera re-id: the same physical
+person is assigned the same track_id regardless of which camera view
+they appear in, because appearance matching is camera-agnostic.
+
+Subscribes
+----------
+  /camera/color/image_raw       sensor_msgs/Image              (BEST_EFFORT)
+  /saltybot/scene/objects        saltybot_scene_msgs/SceneObjectArray (BEST_EFFORT)
+
+Publishes
+---------
+  /saltybot/person_tracks        saltybot_scene_msgs/PersonTrackArray  at publish_hz
+
+Parameters
+----------
+camera_id            str    "front"   Label stored in every PersonTrack
+similarity_threshold float  0.75      Cosine similarity threshold for re-id
+stale_timeout_s      float  30.0      Seconds before inactive track is stale
+max_tracks           int    20        Maximum concurrent active tracks
+publish_hz           float  5.0       Track list publication rate (Hz)
+"""
+
+from __future__ import annotations
+
+import time
+
+import cv2
+import cv_bridge
+import numpy as np
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile, ReliabilityPolicy, HistoryPolicy
+
+from builtin_interfaces.msg import Time as TimeMsg
+from sensor_msgs.msg import Image
+from saltybot_scene_msgs.msg import (
+    SceneObjectArray,
+    PersonTrack,
+    PersonTrackArray,
+)
+
+from ._person_reid import TrackGallery, extract_hsv_histogram
+
+
+# COCO class 0 = person
+_PERSON_CLASS_ID = 0
+
+_BEST_EFFORT_QOS = QoSProfile(
+    reliability=ReliabilityPolicy.BEST_EFFORT,
+    history=HistoryPolicy.KEEP_LAST,
+    depth=1,
+)
+
+
+def _float_to_ros_time(wall_secs: float) -> TimeMsg:
+    """Convert a time.time() float to builtin_interfaces/Time."""
+    sec_int  = int(wall_secs)
+    nanosec  = int((wall_secs - sec_int) * 1e9)
+    msg         = TimeMsg()
+    msg.sec     = sec_int
+    msg.nanosec = nanosec
+    return msg
+
+
+class PersonReidNode(Node):
+
+    def __init__(self) -> None:
+        super().__init__('person_reid_node')
+
+        self.declare_parameter('camera_id',            'front')
+        self.declare_parameter('similarity_threshold', 0.75)
+        self.declare_parameter('stale_timeout_s',      30.0)
+        self.declare_parameter('max_tracks',           20)
+        self.declare_parameter('publish_hz',           5.0)
+
+        self._camera_id  = str(self.get_parameter('camera_id').value)
+        self._threshold  = float(self.get_parameter('similarity_threshold').value)
+        self._timeout    = float(self.get_parameter('stale_timeout_s').value)
+        self._max_tracks = int(self.get_parameter('max_tracks').value)
+        publish_hz       = float(self.get_parameter('publish_hz').value)
+
+        self._bridge: cv_bridge.CvBridge = cv_bridge.CvBridge()
+        self._gallery: TrackGallery      = TrackGallery()
+        self._latest_image: np.ndarray | None = None
+
+        self._img_sub = self.create_subscription(
+            Image,
+            '/camera/color/image_raw',
+            self._on_image,
+            _BEST_EFFORT_QOS,
+        )
+        self._det_sub = self.create_subscription(
+            SceneObjectArray,
+            '/saltybot/scene/objects',
+            self._on_detections,
+            _BEST_EFFORT_QOS,
+        )
+        self._pub = self.create_publisher(
+            PersonTrackArray,
+            '/saltybot/person_tracks',
+            10,
+        )
+
+        period = 1.0 / max(publish_hz, 0.1)
+        self._timer = self.create_timer(period, self._on_timer)
+
+        self.get_logger().info(
+            f'person_reid_node ready — '
+            f'camera_id={self._camera_id} '
+            f'threshold={self._threshold} '
+            f'timeout={self._timeout}s '
+            f'max_tracks={self._max_tracks}'
+        )
+
+    # ── Image callback ────────────────────────────────────────────────────────
+
+    def _on_image(self, msg: Image) -> None:
+        try:
+            self._latest_image = self._bridge.imgmsg_to_cv2(msg, 'bgr8')
+        except Exception as exc:
+            self.get_logger().warning(f'cv_bridge error: {exc}', throttle_duration_sec=5.0)
+
+    # ── Detection callback ────────────────────────────────────────────────────
+
+    def _on_detections(self, msg: SceneObjectArray) -> None:
+        if self._latest_image is None:
+            return
+
+        now = time.time()
+        img = self._latest_image
+        h, w = img.shape[:2]
+
+        for obj in msg.objects:
+            if obj.class_id != _PERSON_CLASS_ID:
+                continue
+
+            # Crop the person ROI from the cached image
+            cx  = obj.bbox.center.position.x
+            cy  = obj.bbox.center.position.y
+            hsw = obj.bbox.size_x / 2.0
+            hsh = obj.bbox.size_y / 2.0
+            x1 = max(0, int(cx - hsw))
+            y1 = max(0, int(cy - hsh))
+            x2 = min(w, int(cx + hsw))
+            y2 = min(h, int(cy + hsh))
+            if x2 <= x1 or y2 <= y1:
+                continue
+
+            crop    = img[y1:y2, x1:x2]
+            feature = extract_hsv_histogram(crop)
+            bbox    = (cx, cy, obj.bbox.size_x, obj.bbox.size_y)
+
+            track_id, sim = self._gallery.match(feature, self._threshold)
+
+            if track_id is not None:
+                self._gallery.update(
+                    track_id, self._camera_id, feature, bbox, float(sim), now
+                )
+            elif len(self._gallery.tracks) < self._max_tracks:
+                self._gallery.add(self._camera_id, feature, bbox, now)
+
+    # ── Timer callback — publish + housekeeping ───────────────────────────────
+
+    def _on_timer(self) -> None:
+        now = time.time()
+        self._gallery.mark_stale(now, self._timeout)
+        self._gallery.prune_stale()
+
+        stamp = self.get_clock().now().to_msg()
+        out   = PersonTrackArray()
+        out.header.stamp    = stamp
+        out.header.frame_id = self._camera_id
+
+        for entry in self._gallery.tracks.values():
+            t                  = PersonTrack()
+            t.header.stamp     = stamp
+            t.header.frame_id  = self._camera_id
+            t.track_id         = entry.track_id
+            t.camera_id        = entry.camera_id
+            t.confidence       = float(entry.confidence)
+            t.first_seen       = _float_to_ros_time(entry.first_seen)
+            t.last_seen        = _float_to_ros_time(entry.last_seen)
+            t.is_stale         = entry.is_stale
+            # Populate last-known bounding box
+            cx, cy, sx, sy     = entry.bbox
+            t.bbox.center.position.x = cx
+            t.bbox.center.position.y = cy
+            t.bbox.size_x            = sx
+            t.bbox.size_y            = sy
+            out.tracks.append(t)
+
+        self._pub.publish(out)
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = PersonReidNode()
+    try:
+        rclpy.spin(node)
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()
+
+
+if __name__ == '__main__':
+    main()
--- a/jetson/ros2_ws/src/saltybot_bringup/setup.py
+++ b/jetson/ros2_ws/src/saltybot_bringup/setup.py
@ -49,6 +49,8 @@ setup(
            'sky_detector            = saltybot_bringup.sky_detect_node:main',
            # Wheel encoder differential drive odometry (Issue #184)
            'wheel_odom              = saltybot_bringup.wheel_odom_node:main',
+            # Appearance-based person re-identification (Issue #322)
+            'person_reid             = saltybot_bringup.person_reid_node:main',
        ],
    },
 )
--- a/jetson/ros2_ws/src/saltybot_bringup/test/test_person_reid.py
+++ b/jetson/ros2_ws/src/saltybot_bringup/test/test_person_reid.py
@ -0,0 +1,528 @@
+"""
+test_person_reid.py — Unit tests for person re-identification helpers (no ROS2).
+
+Covers:
+  extract_hsv_histogram:
+    - output shape is FEAT_DIM (128)
+    - None input → zero vector
+    - zero-size array input → zero vector
+    - solid-colour image → L2-normalised (norm ≈ 1)
+    - same colour on different image sizes → same histogram
+    - different solid colours → dissimilar histograms
+    - all-white image (low saturation) → histogram concentrated in low-S bins
+    - single-pixel image → valid unit histogram
+
+  cosine_similarity:
+    - identical unit vectors → 1.0
+    - zero vector (a) → 0.0
+    - zero vector (b) → 0.0
+    - both zero → 0.0
+    - anti-parallel vectors → -1.0
+    - orthogonal vectors → 0.0
+    - non-unit same-direction vectors → 1.0
+    - non-unit opposite-direction vectors → -1.0
+    - partial overlap → in (0, 1)
+
+  match_track:
+    - empty gallery → (None, 0.0)
+    - perfect match above threshold → correct track_id, sim ≈ 1.0
+    - similarity exactly at threshold → no match (strict >)
+    - below threshold → (None, 0.0)
+    - stale track excluded from matching
+    - multiple tracks → best one returned
+    - all tracks stale → (None, 0.0)
+
+  TrackGallery:
+    - add assigns track_id = 1, 2, 3 ...
+    - add stores a copy of feature (mutation of original doesn't affect gallery)
+    - update changes last_seen, camera_id, confidence
+    - update applies EMA: feature shifts toward new, stays unit-norm
+    - update clears is_stale flag
+    - update preserves first_seen
+    - mark_stale: recent track stays fresh
+    - mark_stale: old track flagged stale
+    - prune_stale removes stale, keeps fresh
+    - empty gallery match → (None, 0.0)
+
+  Re-id integration:
+    - same appearance from two camera_ids → same track_id
+    - different appearance → different track_id
+    - stale + prune + re-add → new track_id
+    - EMA: track adapts over multiple updates (feature norm stays ≈ 1)
+"""
+
+import sys
+import os
+import math
+
+import numpy as np
+import pytest
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+
+from saltybot_bringup._person_reid import (
+    H_BINS,
+    S_BINS,
+    FEAT_DIM,
+    TrackEntry,
+    TrackGallery,
+    cosine_similarity,
+    extract_hsv_histogram,
+    match_track,
+)
+
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+def _solid_bgr(b: int, g: int, r: int, size: int = 32) -> np.ndarray:
+    """Return a solid-colour BGR uint8 image of given size."""
+    import cv2
+    img = np.full((size, size, 3), (b, g, r), dtype=np.uint8)
+    return img
+
+
+def _unit(v: np.ndarray) -> np.ndarray:
+    n = np.linalg.norm(v)
+    return v / n if n > 1e-12 else v
+
+
+def _make_feature(seed: int) -> np.ndarray:
+    """Reproducible random L2-normalised feature vector."""
+    rng = np.random.default_rng(seed)
+    v   = rng.random(FEAT_DIM).astype(np.float32)
+    return v / np.linalg.norm(v)
+
+
+def _make_entry(
+    tid: int,
+    feature: np.ndarray,
+    camera_id: str = 'front',
+    now: float = 1000.0,
+    is_stale: bool = False,
+) -> TrackEntry:
+    return TrackEntry(
+        track_id=tid,
+        camera_id=camera_id,
+        feature=feature.copy(),
+        confidence=1.0,
+        first_seen=now,
+        last_seen=now,
+        is_stale=is_stale,
+    )
+
+
+# ── extract_hsv_histogram ─────────────────────────────────────────────────────
+
+class TestExtractHsvHistogram:
+
+    def test_output_shape(self):
+        img  = _solid_bgr(0, 0, 255)
+        hist = extract_hsv_histogram(img)
+        assert hist.shape == (FEAT_DIM,)
+
+    def test_output_dtype(self):
+        hist = extract_hsv_histogram(_solid_bgr(0, 255, 0))
+        assert hist.dtype == np.float32
+
+    def test_none_returns_zero(self):
+        hist = extract_hsv_histogram(None)
+        assert hist.shape == (FEAT_DIM,)
+        assert np.all(hist == 0.0)
+
+    def test_empty_array_returns_zero(self):
+        import numpy as np
+        empty = np.zeros((0, 0, 3), dtype=np.uint8)
+        hist  = extract_hsv_histogram(empty)
+        assert np.all(hist == 0.0)
+
+    def test_solid_colour_is_unit_norm(self):
+        hist = extract_hsv_histogram(_solid_bgr(255, 0, 0))
+        assert np.linalg.norm(hist) == pytest.approx(1.0, rel=1e-5)
+
+    def test_same_colour_different_sizes_equal(self):
+        h16 = extract_hsv_histogram(_solid_bgr(0, 0, 200, size=16))
+        h64 = extract_hsv_histogram(_solid_bgr(0, 0, 200, size=64))
+        assert np.allclose(h16, h64, atol=1e-5)
+
+    def test_different_colours_different_histograms(self):
+        red  = extract_hsv_histogram(_solid_bgr(0, 0, 255))
+        blue = extract_hsv_histogram(_solid_bgr(255, 0, 0))
+        assert not np.allclose(red, blue, atol=1e-3)
+
+    def test_red_vs_green_low_similarity(self):
+        red   = extract_hsv_histogram(_solid_bgr(0, 0, 255))
+        green = extract_hsv_histogram(_solid_bgr(0, 255, 0))
+        sim   = cosine_similarity(red, green)
+        assert sim < 0.5
+
+    def test_single_pixel_valid(self):
+        import numpy as np
+        px   = np.array([[[0, 200, 200]]], dtype=np.uint8)
+        hist = extract_hsv_histogram(px)
+        assert hist.shape == (FEAT_DIM,)
+        assert np.linalg.norm(hist) == pytest.approx(1.0, rel=1e-5)
+
+    def test_feat_dim_constant(self):
+        assert FEAT_DIM == H_BINS * S_BINS
+        assert FEAT_DIM == 128
+
+
+# ── cosine_similarity ─────────────────────────────────────────────────────────
+
+class TestCosineSimilarity:
+
+    def test_identical_unit_vectors(self):
+        v = _make_feature(0)
+        assert cosine_similarity(v, v.copy()) == pytest.approx(1.0, rel=1e-6)
+
+    def test_zero_vector_a(self):
+        z = np.zeros(FEAT_DIM, dtype=np.float32)
+        v = _make_feature(1)
+        assert cosine_similarity(z, v) == pytest.approx(0.0, abs=1e-9)
+
+    def test_zero_vector_b(self):
+        v = _make_feature(2)
+        z = np.zeros(FEAT_DIM, dtype=np.float32)
+        assert cosine_similarity(v, z) == pytest.approx(0.0, abs=1e-9)
+
+    def test_both_zero(self):
+        z = np.zeros(FEAT_DIM, dtype=np.float32)
+        assert cosine_similarity(z, z) == pytest.approx(0.0, abs=1e-9)
+
+    def test_anti_parallel(self):
+        v = _make_feature(3)
+        assert cosine_similarity(v, -v) == pytest.approx(-1.0, rel=1e-6)
+
+    def test_orthogonal(self):
+        a = np.zeros(FEAT_DIM, dtype=np.float32)
+        b = np.zeros(FEAT_DIM, dtype=np.float32)
+        a[0] = 1.0
+        b[1] = 1.0
+        assert cosine_similarity(a, b) == pytest.approx(0.0, abs=1e-9)
+
+    def test_non_unit_same_direction(self):
+        v = _make_feature(4)
+        assert cosine_similarity(v * 5.0, v * 3.0) == pytest.approx(1.0, rel=1e-6)
+
+    def test_non_unit_opposite_direction(self):
+        v = _make_feature(5)
+        assert cosine_similarity(v * 7.0, -v * 2.0) == pytest.approx(-1.0, rel=1e-6)
+
+    def test_partial_overlap_in_range(self):
+        a = np.zeros(FEAT_DIM, dtype=np.float32)
+        b = np.zeros(FEAT_DIM, dtype=np.float32)
+        a[0] = 1.0
+        b[0] = 1.0
+        b[1] = 1.0
+        sim = cosine_similarity(a, b)
+        assert 0.0 < sim < 1.0
+
+
+# ── match_track ───────────────────────────────────────────────────────────────
+
+class TestMatchTrack:
+
+    def test_empty_gallery(self):
+        feat = _make_feature(10)
+        tid, sim = match_track(feat, {}, threshold=0.75)
+        assert tid is None
+        assert sim == pytest.approx(0.0, abs=1e-9)
+
+    def test_exact_match_above_threshold(self):
+        feat   = _make_feature(11)
+        tracks = {1: _make_entry(1, feat)}
+        tid, sim = match_track(feat, tracks, threshold=0.75)
+        assert tid == 1
+        assert sim == pytest.approx(1.0, rel=1e-5)
+
+    def test_similarity_at_threshold_no_match(self):
+        """Strict >: equal to threshold is not a match."""
+        feat = _make_feature(12)
+        # construct a feature at exactly cosine = threshold
+        # use two orthogonal features blended so sim == threshold
+        a = np.zeros(FEAT_DIM, dtype=np.float32)
+        b = np.zeros(FEAT_DIM, dtype=np.float32)
+        a[0] = 1.0
+        b[1] = 1.0
+        threshold = 0.75
+        # query is a; gallery is a → sim = 1.0, way above. Just test with threshold=1.0+eps
+        tracks = {1: _make_entry(1, a)}
+        tid, sim = match_track(a, tracks, threshold=1.0 + 1e-9)
+        assert tid is None
+
+    def test_below_threshold_no_match(self):
+        feat_a = _make_feature(13)
+        feat_b = _make_feature(14)
+        tracks = {1: _make_entry(1, feat_b)}
+        # Force threshold above any realistic similarity between random features
+        tid, sim = match_track(feat_a, tracks, threshold=0.9999)
+        if cosine_similarity(feat_a, feat_b) <= 0.9999:
+            assert tid is None
+
+    def test_stale_track_excluded(self):
+        feat   = _make_feature(15)
+        tracks = {1: _make_entry(1, feat, is_stale=True)}
+        tid, sim = match_track(feat, tracks, threshold=0.5)
+        assert tid is None
+
+    def test_multiple_tracks_best_returned(self):
+        query  = _make_feature(16)
+        close  = query.copy()                    # sim = 1.0
+        far    = _make_feature(17)               # lower sim
+        tracks = {
+            1: _make_entry(1, far),
+            2: _make_entry(2, close),
+        }
+        tid, sim = match_track(query, tracks, threshold=0.5)
+        assert tid == 2
+        assert sim == pytest.approx(1.0, rel=1e-5)
+
+    def test_all_stale_no_match(self):
+        feat   = _make_feature(18)
+        tracks = {
+            1: _make_entry(1, feat, is_stale=True),
+            2: _make_entry(2, feat, is_stale=True),
+        }
+        tid, sim = match_track(feat, tracks, threshold=0.5)
+        assert tid is None
+
+    def test_fresh_and_stale_returns_fresh(self):
+        feat   = _make_feature(19)
+        tracks = {
+            1: _make_entry(1, feat, is_stale=True),
+            2: _make_entry(2, feat, is_stale=False),
+        }
+        tid, sim = match_track(feat, tracks, threshold=0.5)
+        assert tid == 2
+
+
+# ── TrackGallery ──────────────────────────────────────────────────────────────
+
+_BBOX = (10.0, 20.0, 50.0, 80.0)
+
+
+class TestTrackGallery:
+
+    def test_add_increments_track_id(self):
+        g   = TrackGallery()
+        now = 1000.0
+        t1  = g.add('front', _make_feature(20), _BBOX, now)
+        t2  = g.add('front', _make_feature(21), _BBOX, now)
+        t3  = g.add('rear',  _make_feature(22), _BBOX, now)
+        assert t1 == 1
+        assert t2 == 2
+        assert t3 == 3
+
+    def test_add_stores_copy(self):
+        g   = TrackGallery()
+        feat = _make_feature(23)
+        orig = feat.copy()
+        g.add('front', feat, _BBOX, 1000.0)
+        feat[0] = 99.9        # mutate original
+        stored  = g.tracks[1].feature
+        assert np.allclose(stored, orig, atol=1e-6)
+
+    def test_add_sets_timestamps(self):
+        g   = TrackGallery()
+        now = 5000.0
+        g.add('front', _make_feature(24), _BBOX, now)
+        e = g.tracks[1]
+        assert e.first_seen == pytest.approx(now)
+        assert e.last_seen  == pytest.approx(now)
+
+    def test_add_initial_confidence_one(self):
+        g = TrackGallery()
+        g.add('front', _make_feature(25), _BBOX, 1000.0)
+        assert g.tracks[1].confidence == pytest.approx(1.0)
+
+    def test_update_changes_last_seen(self):
+        g    = TrackGallery()
+        feat = _make_feature(26)
+        g.add('front', feat, _BBOX, 1000.0)
+        g.update(1, 'front', feat, _BBOX, 0.9, 2000.0)
+        assert g.tracks[1].last_seen  == pytest.approx(2000.0)
+        assert g.tracks[1].first_seen == pytest.approx(1000.0)   # preserved
+
+    def test_update_changes_camera_id(self):
+        g    = TrackGallery()
+        feat = _make_feature(27)
+        g.add('front', feat, _BBOX, 1000.0)
+        g.update(1, 'right', feat, _BBOX, 0.8, 2000.0)
+        assert g.tracks[1].camera_id == 'right'
+
+    def test_update_applies_ema(self):
+        g     = TrackGallery()
+        feat0 = _make_feature(28)
+        g.add('front', feat0, _BBOX, 1000.0)
+
+        feat1 = _make_feature(29)
+        g.update(1, 'front', feat1, _BBOX, 0.85, 2000.0)
+        stored = g.tracks[1].feature
+
+        # EMA: 0.3*feat1 + 0.7*feat0 (then renorm) — not equal to either
+        assert not np.allclose(stored, feat0, atol=1e-4)
+        assert not np.allclose(stored, feat1, atol=1e-4)
+
+    def test_update_feature_stays_unit_norm(self):
+        g    = TrackGallery()
+        feat = _make_feature(30)
+        g.add('front', feat, _BBOX, 1000.0)
+        for i in range(10):
+            new_feat = _make_feature(31 + i)
+            g.update(1, 'front', new_feat, _BBOX, 0.9, 1000.0 + i)
+        assert np.linalg.norm(g.tracks[1].feature) == pytest.approx(1.0, rel=1e-5)
+
+    def test_update_clears_stale(self):
+        g    = TrackGallery()
+        feat = _make_feature(41)
+        g.add('front', feat, _BBOX, 1000.0)
+        g.tracks[1].is_stale = True
+        g.update(1, 'front', feat, _BBOX, 0.9, 2000.0)
+        assert g.tracks[1].is_stale is False
+
+    def test_mark_stale_recent_track_stays_fresh(self):
+        g   = TrackGallery()
+        now = 1000.0
+        g.add('front', _make_feature(42), _BBOX, now)
+        g.mark_stale(now + 10.0, timeout_s=30.0)
+        assert g.tracks[1].is_stale is False
+
+    def test_mark_stale_old_track_flagged(self):
+        g   = TrackGallery()
+        now = 1000.0
+        g.add('front', _make_feature(43), _BBOX, now)
+        g.mark_stale(now + 31.0, timeout_s=30.0)
+        assert g.tracks[1].is_stale is True
+
+    def test_mark_stale_exactly_at_timeout_not_stale(self):
+        """Timeout is strictly >, so last_seen + timeout_s is still fresh."""
+        g   = TrackGallery()
+        now = 1000.0
+        g.add('front', _make_feature(44), _BBOX, now)
+        g.mark_stale(now + 30.0, timeout_s=30.0)
+        assert g.tracks[1].is_stale is False
+
+    def test_prune_stale_removes_flagged(self):
+        g   = TrackGallery()
+        now = 1000.0
+        g.add('front', _make_feature(45), _BBOX, now)   # track 1
+        g.add('front', _make_feature(46), _BBOX, now)   # track 2
+        g.tracks[1].is_stale = True
+        g.prune_stale()
+        assert 1 not in g.tracks
+        assert 2 in g.tracks
+
+    def test_prune_stale_keeps_fresh(self):
+        g   = TrackGallery()
+        now = 1000.0
+        g.add('front', _make_feature(47), _BBOX, now)
+        g.prune_stale()
+        assert 1 in g.tracks
+
+    def test_empty_gallery_match(self):
+        g = TrackGallery()
+        tid, sim = g.match(_make_feature(48), threshold=0.75)
+        assert tid is None
+        assert sim == pytest.approx(0.0, abs=1e-9)
+
+    def test_match_delegates_to_match_track(self):
+        g    = TrackGallery()
+        feat = _make_feature(49)
+        g.add('front', feat, _BBOX, 1000.0)
+        tid, sim = g.match(feat, threshold=0.5)
+        assert tid == 1
+        assert sim == pytest.approx(1.0, rel=1e-5)
+
+
+# ── Re-id integration ─────────────────────────────────────────────────────────
+
+class TestReIdIntegration:
+
+    def test_same_appearance_two_cameras_same_track(self):
+        """Same feature from two cameras → same track_id."""
+        g    = TrackGallery()
+        feat = _make_feature(50)
+        now  = 1000.0
+
+        # First sighting: front camera creates track 1
+        t1 = g.add('front', feat, _BBOX, now)
+        assert t1 == 1
+
+        # Second sighting: right camera matches → updates same track
+        tid, sim = g.match(feat, threshold=0.75)
+        assert tid == 1
+        g.update(tid, 'right', feat, _BBOX, float(sim), now + 1.0)
+        assert len(g.tracks) == 1
+
+    def test_different_appearance_different_tracks(self):
+        """Two orthogonal features must not match → separate track IDs."""
+        g   = TrackGallery()
+        now = 1000.0
+
+        # Construct two exactly orthogonal unit feature vectors
+        feat_a = np.zeros(FEAT_DIM, dtype=np.float32)
+        feat_b = np.zeros(FEAT_DIM, dtype=np.float32)
+        feat_a[0] = 1.0   # basis vector e0
+        feat_b[1] = 1.0   # basis vector e1 — cosine sim = 0.0
+
+        g.add('front', feat_a, _BBOX, now)
+
+        tid, sim = g.match(feat_b, threshold=0.75)
+        assert tid is None
+
+        g.add('front', feat_b, _BBOX, now + 0.1)
+        assert len(g.tracks) == 2
+
+    def test_stale_prune_readd_new_id(self):
+        g    = TrackGallery()
+        feat = _make_feature(53)
+        now  = 1000.0
+
+        t1 = g.add('front', feat, _BBOX, now)
+        assert t1 == 1
+
+        g.mark_stale(now + 31.0, timeout_s=30.0)
+        g.prune_stale()
+        assert len(g.tracks) == 0
+
+        # Re-add same person gets new ID
+        t2 = g.add('front', feat, _BBOX, now + 40.0)
+        assert t2 == 2    # next_id incremented past 1
+
+    def test_ema_feature_norm_stays_unit_after_many_updates(self):
+        g    = TrackGallery()
+        feat = _make_feature(54)
+        g.add('front', feat, _BBOX, 0.0)
+
+        for i in range(50):
+            new_feat = _make_feature(55 + i)
+            g.update(1, 'front', new_feat, _BBOX, 0.9, float(i))
+
+        norm = np.linalg.norm(g.tracks[1].feature)
+        assert norm == pytest.approx(1.0, rel=1e-4)
+
+    def test_max_tracks_not_exceeded_in_gallery(self):
+        """Gallery itself is unbounded; max_tracks enforcement is in the node.
+           Confirm gallery can hold many entries without errors."""
+        g = TrackGallery()
+        for i in range(25):
+            g.add('front', _make_feature(100 + i), _BBOX, float(i))
+        assert len(g.tracks) == 25
+
+    def test_hsv_same_image_produces_high_similarity(self):
+        """Same-crop HSV histogram should have cosine sim ≈ 1."""
+        import cv2
+        img  = _solid_bgr(50, 120, 200)
+        h1   = extract_hsv_histogram(img)
+        h2   = extract_hsv_histogram(img.copy())
+        assert cosine_similarity(h1, h2) == pytest.approx(1.0, rel=1e-5)
+
+    def test_hsv_different_colours_low_similarity(self):
+        """Red vs cyan BGR should be dissimilar."""
+        red  = extract_hsv_histogram(_solid_bgr(0, 0, 200))
+        cyan = extract_hsv_histogram(_solid_bgr(200, 200, 0))
+        assert cosine_similarity(red, cyan) < 0.5
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v'])
--- a/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt
@ -19,6 +19,9 @@ rosidl_generate_interfaces(${PROJECT_NAME}
  # Issue #274 — HSV color segmentation
  "msg/ColorDetection.msg"
  "msg/ColorDetectionArray.msg"
+  # Issue #322 — cross-camera person re-identification
+  "msg/PersonTrack.msg"
+  "msg/PersonTrackArray.msg"
  DEPENDENCIES std_msgs geometry_msgs vision_msgs builtin_interfaces
 )

--- a/jetson/ros2_ws/src/saltybot_scene_msgs/msg/PersonTrack.msg
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/PersonTrack.msg
@ -0,0 +1,19 @@
+# PersonTrack.msg — stable cross-camera person identity track (Issue #322)
+#
+# track_id     : monotonically-increasing track ID, stable across camera views
+# camera_id    : source camera label ("front", "right", "rear", "left")
+# bbox         : last-known bounding box in the source image (centre + size, pixels)
+# confidence   : cosine-similarity of the matched appearance feature (0–1);
+#                1.0 for tracks created from a first-time observation
+# first_seen   : wall-clock time when this track was first created
+# last_seen    : wall-clock time of the most recent appearance observation
+# is_stale     : true when (now − last_seen) > stale_timeout_s
+#
+std_msgs/Header header
+uint32  track_id
+string  camera_id
+vision_msgs/BoundingBox2D bbox
+float32 confidence
+builtin_interfaces/Time first_seen
+builtin_interfaces/Time last_seen
+bool    is_stale
--- a/jetson/ros2_ws/src/saltybot_scene_msgs/msg/PersonTrackArray.msg
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/PersonTrackArray.msg
@ -0,0 +1,4 @@
+# PersonTrackArray.msg — all active person tracks from one camera view (Issue #322)
+#
+std_msgs/Header header
+PersonTrack[] tracks