diff --git a/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_obstacle_size.py b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_obstacle_size.py
new file mode 100644
index 0000000..ca5ab45
--- /dev/null
+++ b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_obstacle_size.py
@@ -0,0 +1,361 @@
+"""
+_obstacle_size.py — Depth-based obstacle size estimation by projecting LIDAR
+clusters into the D435i depth image (no ROS2 deps).
+
+Algorithm
+---------
+1. Transform each LIDAR cluster centroid from LIDAR frame to camera frame
+   using a configurable translation-only extrinsic (typical co-mounted setup).
+2. Project to depth-image pixel coordinates via the pinhole camera model.
+3. Sample the D435i depth image in a window around the projected pixel to
+   obtain a robust Z estimate (uint16 mm → metres).
+4. Derive horizontal width (metres) directly from the LIDAR cluster bounding
+   box, which is more reliable than image-based re-measurement.
+5. Derive vertical height (metres) by scanning a vertical strip in the depth
+   image: collect pixels whose depth is within `z_tol` of the sampled Z, then
+   compute the row extent and back-project using the camera fy.
+6. Return an ObstacleSizeEstimate per cluster.
+
+Coordinate frames
+-----------------
+LIDAR frame (+X forward, +Y left, +Z up — 2-D scan at z=0):
+  x_lidar = forward distance (metres)
+  y_lidar = lateral distance (metres, positive = left)
+  z_lidar = 0   (2-D LIDAR always lies in the horizontal plane)
+
+Camera frame (+X right, +Y down, +Z forward — standard pinhole):
+  x_cam =  −y_lidar + ex
+  y_cam =  −z_lidar + ey  =  ey          (z_lidar=0 for 2-D LIDAR)
+  z_cam =   x_lidar + ez
+
+Extrinsic parameters (CameraParams.ex/ey/ez) give the LIDAR origin position
+in the camera coordinate frame:
+  ex   — lateral offset  (positive = LIDAR origin is to the right of camera)
+  ey   — vertical offset (positive = LIDAR origin is below camera)
+  ez   — forward offset  (positive = LIDAR origin is in front of camera)
+
+Public API
+----------
+  CameraParams           dataclass — camera intrinsics + co-mount extrinsics
+  ObstacleSizeEstimate   NamedTuple — per-cluster result
+
+  lidar_to_camera(x_l, y_l, params)          → (x_cam, y_cam, z_cam)
+  project_to_pixel(x_cam, y_cam, z_cam, p)   → (u, v) or None if out-of-bounds
+  sample_depth_median(depth_u16, u, v, win, scale) → (depth_m, n_valid)
+  estimate_height(depth_u16, u_c, v_c, z_ref, params,
+                  search_rows, col_hw, z_tol) → height_m
+  estimate_cluster_size(cluster, depth_u16, params,
+                        depth_window, search_rows, col_hw, z_tol,
+                        cluster_id)           → ObstacleSizeEstimate
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass, field
+from typing import NamedTuple, Optional, Tuple
+
+import numpy as np
+
+
+# ── Camera parameters ─────────────────────────────────────────────────────────
+
+@dataclass
+class CameraParams:
+    """Pinhole intrinsics and LIDAR→camera extrinsics for the D435i.
+
+    Intrinsics (D435i 640×480 depth stream defaults — override from CameraInfo):
+      fx, fy  : focal lengths in pixels
+      cx, cy  : principal point in pixels
+      width, height : image dimensions in pixels
+      depth_scale   : multiply raw uint16 value by this to get metres
+                      D435i default = 0.001 (raw value is millimetres)
+
+    Extrinsics (LIDAR origin position in camera frame):
+      ex, ey, ez : translation (metres) — see module docstring
+    """
+    fx:           float = 383.0    # D435i 640×480 depth approx
+    fy:           float = 383.0
+    cx:           float = 320.0
+    cy:           float = 240.0
+    width:        int   = 640
+    height:       int   = 480
+    depth_scale:  float = 0.001    # mm → m
+
+    # LIDAR→camera extrinsics: LIDAR origin position in camera frame
+    ex: float = 0.0    # lateral (m); 0 = LIDAR directly in front of camera
+    ey: float = 0.05   # vertical (m); +0.05 = LIDAR 5 cm below camera
+    ez: float = 0.0    # forward  (m); 0 = LIDAR at same depth as camera
+
+
+# ── Result type ───────────────────────────────────────────────────────────────
+
+class ObstacleSizeEstimate(NamedTuple):
+    """Size estimate for one LIDAR cluster projected into the depth image."""
+    obstacle_id:  int    # user-supplied cluster index (0-based) or track ID
+    centroid_x:   float  # LIDAR-frame forward distance (metres)
+    centroid_y:   float  # LIDAR-frame lateral distance (metres, +left)
+    depth_z:      float  # D435i sampled depth at projected centroid (metres)
+    width_m:      float  # horizontal size, LIDAR-derived (metres)
+    height_m:     float  # vertical size, depth-image-derived (metres)
+    pixel_u:      int    # projected centroid column (-1 = out of image)
+    pixel_v:      int    # projected centroid row    (-1 = out of image)
+    lidar_range:  float  # range from LIDAR origin to centroid (metres)
+    confidence:   float  # 0.0–1.0; based on depth sample quality
+
+
+# ── Coordinate transform ──────────────────────────────────────────────────────
+
+def lidar_to_camera(
+    x_lidar: float,
+    y_lidar: float,
+    params:  CameraParams,
+) -> Tuple[float, float, float]:
+    """
+    Transform a 2-D LIDAR point to the camera coordinate frame.
+
+    Parameters
+    ----------
+    x_lidar : forward distance in LIDAR frame (metres)
+    y_lidar : lateral distance in LIDAR frame (metres, +left)
+    params  : CameraParams with extrinsic translation ex/ey/ez
+
+    Returns
+    -------
+    (x_cam, y_cam, z_cam) in camera frame (metres)
+      x_cam = right, y_cam = down, z_cam = forward
+    """
+    x_cam = -y_lidar + params.ex
+    y_cam =  params.ey          # LIDAR z=0 for 2-D scan
+    z_cam =  x_lidar + params.ez
+    return float(x_cam), float(y_cam), float(z_cam)
+
+
+# ── Pinhole projection ────────────────────────────────────────────────────────
+
+def project_to_pixel(
+    x_cam:  float,
+    y_cam:  float,
+    z_cam:  float,
+    params: CameraParams,
+) -> Optional[Tuple[int, int]]:
+    """
+    Project a 3-D camera-frame point to image pixel coordinates.
+
+    Returns None if z_cam ≤ 0 or the projection falls outside the image.
+
+    Returns
+    -------
+    (u, v) integer pixel coordinates (column, row), or None.
+    """
+    if z_cam <= 0.0:
+        return None
+    u = params.fx * x_cam / z_cam + params.cx
+    v = params.fy * y_cam / z_cam + params.cy
+    ui, vi = int(round(u)), int(round(v))
+    if 0 <= ui < params.width and 0 <= vi < params.height:
+        return ui, vi
+    return None
+
+
+# ── Depth sampling ────────────────────────────────────────────────────────────
+
+def sample_depth_median(
+    depth_u16:   np.ndarray,
+    u:           int,
+    v:           int,
+    window_px:   int   = 5,
+    depth_scale: float = 0.001,
+) -> Tuple[float, int]:
+    """
+    Median depth in a square window centred on (u, v) in a uint16 depth image.
+
+    Zeros (invalid readings) are excluded from the median.
+
+    Parameters
+    ----------
+    depth_u16   : (H, W) uint16 depth image (raw units, e.g. mm for D435i)
+    u, v        : centre pixel (column, row)
+    window_px   : half-side of the sampling square (window of 2*window_px+1)
+    depth_scale : scale factor to convert raw units to metres (0.001 for mm)
+
+    Returns
+    -------
+    (depth_m, n_valid) — median depth in metres and count of valid pixels.
+    depth_m = 0.0 when n_valid == 0.
+    """
+    h, w = depth_u16.shape
+    r0 = max(0, v - window_px)
+    r1 = min(h, v + window_px + 1)
+    c0 = max(0, u - window_px)
+    c1 = min(w, u + window_px + 1)
+
+    patch = depth_u16[r0:r1, c0:c1].astype(np.float64)
+    valid = patch[patch > 0]
+    if len(valid) == 0:
+        return 0.0, 0
+    return float(np.median(valid) * depth_scale), int(len(valid))
+
+
+# ── Height estimation ─────────────────────────────────────────────────────────
+
+def estimate_height(
+    depth_u16:   np.ndarray,
+    u_c:         int,
+    v_c:         int,
+    z_ref:       float,
+    params:      CameraParams,
+    search_rows: int   = 120,
+    col_hw:      int   = 10,
+    z_tol:       float = 0.30,
+) -> float:
+    """
+    Estimate the vertical height (metres) of an obstacle in the depth image.
+
+    Scans a vertical strip centred on (u_c, v_c) and finds the row extent of
+    pixels whose depth is within `z_tol` metres of `z_ref`.
+
+    Parameters
+    ----------
+    depth_u16   : (H, W) uint16 depth image
+    u_c, v_c    : projected centroid pixel
+    z_ref       : reference depth (metres) — from sample_depth_median
+    params      : CameraParams (needs fy, depth_scale)
+    search_rows : half-height of the search strip (rows above/below v_c)
+    col_hw      : half-width  of the search strip (columns left/right of u_c)
+    z_tol       : depth tolerance — pixels within z_ref ± z_tol are included
+
+    Returns
+    -------
+    height_m — vertical extent in metres; 0.0 if fewer than 2 valid rows found.
+    """
+    if z_ref <= 0.0:
+        return 0.0
+
+    h, w = depth_u16.shape
+    r0 = max(0, v_c - search_rows)
+    r1 = min(h, v_c + search_rows + 1)
+    c0 = max(0, u_c - col_hw)
+    c1 = min(w, u_c + col_hw + 1)
+
+    strip = depth_u16[r0:r1, c0:c1].astype(np.float64) * params.depth_scale
+
+    # For each row in the strip, check if any valid pixel is within z_tol of z_ref
+    z_min = z_ref - z_tol
+    z_max = z_ref + z_tol
+    valid_mask = (strip > 0) & (strip >= z_min) & (strip <= z_max)
+    valid_rows = np.any(valid_mask, axis=1)   # (n_rows,) bool
+
+    if valid_rows.sum() < 2:
+        return 0.0
+
+    row_indices = np.where(valid_rows)[0]     # rows within the strip
+    v_top_strip    = int(row_indices[0])
+    v_bottom_strip = int(row_indices[-1])
+
+    # Convert to absolute image rows
+    v_top    = r0 + v_top_strip
+    v_bottom = r0 + v_bottom_strip
+
+    row_extent = v_bottom - v_top
+    if row_extent <= 0:
+        return 0.0
+
+    # Back-project row extent to metres at depth z_ref
+    height_m = float(row_extent) * z_ref / params.fy
+    return float(height_m)
+
+
+# ── Main entry point ──────────────────────────────────────────────────────────
+
+def estimate_cluster_size(
+    cluster,                        # _lidar_clustering.Cluster NamedTuple
+    depth_u16:   np.ndarray,        # (H, W) uint16 depth image
+    params:      CameraParams,
+    depth_window: int   = 5,        # half-side for depth median sampling
+    search_rows:  int   = 120,      # half-height of height search strip
+    col_hw:       int   = 10,       # half-width of height search strip
+    z_tol:        float = 0.30,     # depth tolerance for height estimation
+    obstacle_id:  int   = 0,        # caller-assigned ID
+) -> ObstacleSizeEstimate:
+    """
+    Estimate the 3-D size of one LIDAR cluster using the D435i depth image.
+
+    Parameters
+    ----------
+    cluster      : Cluster from _lidar_clustering.cluster_points()
+    depth_u16    : uint16 depth image (H × W), values in millimetres
+    params       : CameraParams with intrinsics + extrinsics
+    depth_window : half-side of median depth sampling window (pixels)
+    search_rows  : half-height for vertical height search (rows)
+    col_hw       : half-width  for vertical height search (columns)
+    z_tol        : depth tolerance when collecting obstacle pixels (metres)
+    obstacle_id  : arbitrary integer ID for this estimate
+
+    Returns
+    -------
+    ObstacleSizeEstimate
+    """
+    cx, cy = float(cluster.centroid[0]), float(cluster.centroid[1])
+    lidar_range = float(math.sqrt(cx * cx + cy * cy))
+
+    # Transform centroid to camera frame
+    x_cam, y_cam, z_cam = lidar_to_camera(cx, cy, params)
+
+    # Project to depth image pixel
+    px = project_to_pixel(x_cam, y_cam, z_cam, params)
+    if px is None:
+        return ObstacleSizeEstimate(
+            obstacle_id  = obstacle_id,
+            centroid_x   = cx,
+            centroid_y   = cy,
+            depth_z      = 0.0,
+            width_m      = float(cluster.width_m),
+            height_m     = 0.0,
+            pixel_u      = -1,
+            pixel_v      = -1,
+            lidar_range  = lidar_range,
+            confidence   = 0.0,
+        )
+
+    u, v = px
+
+    # Sample depth
+    depth_m, n_valid = sample_depth_median(
+        depth_u16, u, v, window_px=depth_window,
+        depth_scale=params.depth_scale,
+    )
+
+    # Use LIDAR range as fallback when depth image has no valid reading
+    if depth_m <= 0.0:
+        depth_m = z_cam
+        depth_conf = 0.0
+    else:
+        # Confidence: ratio of valid pixels in window + how close to expected Z
+        max_valid = (2 * depth_window + 1) ** 2
+        fill_ratio = min(1.0, n_valid / max(max_valid * 0.25, 1.0))
+        z_err = abs(depth_m - z_cam)
+        z_conf = max(0.0, 1.0 - z_err / max(z_cam, 0.1))
+        depth_conf = float(0.6 * fill_ratio + 0.4 * z_conf)
+
+    # Width: LIDAR gives reliable horizontal measurement
+    width_m = float(cluster.width_m)
+
+    # Height: from depth image vertical strip
+    height_m = estimate_height(
+        depth_u16, u, v, depth_m, params,
+        search_rows=search_rows, col_hw=col_hw, z_tol=z_tol,
+    )
+
+    return ObstacleSizeEstimate(
+        obstacle_id  = obstacle_id,
+        centroid_x   = cx,
+        centroid_y   = cy,
+        depth_z      = float(depth_m),
+        width_m      = width_m,
+        height_m     = float(height_m),
+        pixel_u      = int(u),
+        pixel_v      = int(v),
+        lidar_range  = lidar_range,
+        confidence   = float(depth_conf),
+    )
diff --git a/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/obstacle_size_node.py b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/obstacle_size_node.py
new file mode 100644
index 0000000..d35ebdd
--- /dev/null
+++ b/jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/obstacle_size_node.py
@@ -0,0 +1,262 @@
+"""
+obstacle_size_node.py — Depth-based obstacle size estimator (Issue #348).
+
+Fuses 2-D LIDAR clusters with the D435i depth image to estimate the full
+3-D width and height (metres) of each obstacle.
+
+Subscribes
+----------
+  /scan                         sensor_msgs/LaserScan    (BEST_EFFORT, ~10 Hz)
+  /camera/depth/image_rect_raw  sensor_msgs/Image        (BEST_EFFORT, ~15 Hz)
+  /camera/depth/camera_info     sensor_msgs/CameraInfo   (RELIABLE)
+
+Publishes
+---------
+  /saltybot/obstacle_sizes      saltybot_scene_msgs/ObstacleSizeArray
+
+Parameters
+----------
+  Clustering
+    distance_threshold_m  float  0.20   Max gap between consecutive scan pts (m)
+    min_points            int    3      Min LIDAR points per cluster
+    range_min_m           float  0.05   Discard ranges below this (m)
+    range_max_m           float  12.0   Discard ranges above this (m)
+
+  Depth sampling
+    depth_window_px       int    5      Half-side of median sampling window
+    search_rows_px        int    120    Half-height of height search strip
+    col_hw_px             int    10     Half-width  of height search strip
+    z_tol_m               float  0.30   Depth tolerance for height collection
+
+  Extrinsics (LIDAR origin in camera frame — metres)
+    lidar_ex              float  0.0    Camera-frame X (right) offset
+    lidar_ey              float  0.05   Camera-frame Y (down) offset
+    lidar_ez              float  0.0    Camera-frame Z (forward) offset
+
+Notes
+-----
+Camera intrinsics are read from /camera/depth/camera_info and override the
+defaults in CameraParams once the first CameraInfo message arrives.
+
+If the depth image is unavailable, the node publishes with depth_z=0 and
+height_m=0, confidence=0 (LIDAR-only width estimate).
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+import numpy as np
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile, ReliabilityPolicy, HistoryPolicy
+
+from sensor_msgs.msg import CameraInfo, Image, LaserScan
+from saltybot_scene_msgs.msg import ObstacleSize, ObstacleSizeArray
+
+from ._lidar_clustering import scan_to_cartesian, cluster_points
+from ._obstacle_size import CameraParams, estimate_cluster_size
+
+
+_SENSOR_QOS = QoSProfile(
+    reliability=ReliabilityPolicy.BEST_EFFORT,
+    history=HistoryPolicy.KEEP_LAST,
+    depth=4,
+)
+_RELIABLE_QOS = QoSProfile(
+    reliability=ReliabilityPolicy.RELIABLE,
+    history=HistoryPolicy.KEEP_LAST,
+    depth=5,
+)
+
+
+class ObstacleSizeNode(Node):
+
+    def __init__(self) -> None:
+        super().__init__('obstacle_size_node')
+
+        # ── Parameters ──────────────────────────────────────────────────────
+        # Clustering
+        self.declare_parameter('distance_threshold_m', 0.20)
+        self.declare_parameter('min_points',           3)
+        self.declare_parameter('range_min_m',          0.05)
+        self.declare_parameter('range_max_m',          12.0)
+        # Depth sampling
+        self.declare_parameter('depth_window_px',  5)
+        self.declare_parameter('search_rows_px',   120)
+        self.declare_parameter('col_hw_px',        10)
+        self.declare_parameter('z_tol_m',          0.30)
+        # Extrinsics
+        self.declare_parameter('lidar_ex', 0.0)
+        self.declare_parameter('lidar_ey', 0.05)
+        self.declare_parameter('lidar_ez', 0.0)
+
+        p = self.get_parameter
+        self._dist_thresh = float(p('distance_threshold_m').value)
+        self._min_pts     = int(p('min_points').value)
+        self._range_min   = float(p('range_min_m').value)
+        self._range_max   = float(p('range_max_m').value)
+        self._win_px      = int(p('depth_window_px').value)
+        self._rows_px     = int(p('search_rows_px').value)
+        self._col_hw      = int(p('col_hw_px').value)
+        self._z_tol       = float(p('z_tol_m').value)
+
+        # Build default CameraParams (overridden when CameraInfo arrives)
+        self._cam = CameraParams(
+            ex=float(p('lidar_ex').value),
+            ey=float(p('lidar_ey').value),
+            ez=float(p('lidar_ez').value),
+        )
+        self._cam_info_received = False
+
+        # Latest depth frame
+        self._depth_image: Optional[np.ndarray] = None
+
+        # ── Subscribers ──────────────────────────────────────────────────────
+        self._scan_sub = self.create_subscription(
+            LaserScan, '/scan', self._on_scan, _SENSOR_QOS
+        )
+        self._depth_sub = self.create_subscription(
+            Image, '/camera/depth/image_rect_raw',
+            self._on_depth, _SENSOR_QOS
+        )
+        self._info_sub = self.create_subscription(
+            CameraInfo, '/camera/depth/camera_info',
+            self._on_camera_info, _RELIABLE_QOS
+        )
+
+        # ── Publisher ────────────────────────────────────────────────────────
+        self._pub = self.create_publisher(
+            ObstacleSizeArray, '/saltybot/obstacle_sizes', 10
+        )
+
+        self.get_logger().info(
+            f'obstacle_size_node ready — '
+            f'dist_thresh={self._dist_thresh}m, z_tol={self._z_tol}m'
+        )
+
+    # ── Callbacks ─────────────────────────────────────────────────────────────
+
+    def _on_camera_info(self, msg: CameraInfo) -> None:
+        """Latch camera intrinsics on first message."""
+        if self._cam_info_received:
+            return
+        ex = self._cam.ex
+        ey = self._cam.ey
+        ez = self._cam.ez
+        self._cam = CameraParams(
+            fx=float(msg.k[0]),
+            fy=float(msg.k[4]),
+            cx=float(msg.k[2]),
+            cy=float(msg.k[5]),
+            width=int(msg.width),
+            height=int(msg.height),
+            depth_scale=self._cam.depth_scale,
+            ex=ex, ey=ey, ez=ez,
+        )
+        self._cam_info_received = True
+        self.get_logger().info(
+            f'Camera intrinsics loaded: '
+            f'fx={self._cam.fx:.1f} fy={self._cam.fy:.1f} '
+            f'cx={self._cam.cx:.1f} cy={self._cam.cy:.1f} '
+            f'{self._cam.width}×{self._cam.height}'
+        )
+
+    def _on_depth(self, msg: Image) -> None:
+        """Cache the latest depth frame as a numpy uint16 array."""
+        enc = msg.encoding.lower()
+        if enc not in ('16uc1', 'mono16'):
+            self.get_logger().warn(
+                f'Unexpected depth encoding {msg.encoding!r} — expected 16UC1',
+                throttle_duration_sec=10.0,
+            )
+            return
+        data = np.frombuffer(msg.data, dtype=np.uint16)
+        try:
+            self._depth_image = data.reshape((msg.height, msg.width))
+        except ValueError as exc:
+            self.get_logger().warn(f'Depth reshape failed: {exc}')
+
+    def _on_scan(self, msg: LaserScan) -> None:
+        """Cluster LIDAR scan, project clusters to depth image, publish sizes."""
+        points = scan_to_cartesian(
+            list(msg.ranges),
+            msg.angle_min,
+            msg.angle_increment,
+            range_min=self._range_min,
+            range_max=self._range_max,
+        )
+        clusters = cluster_points(
+            points,
+            distance_threshold_m=self._dist_thresh,
+            min_points=self._min_pts,
+        )
+
+        depth = self._depth_image
+
+        # Build output message
+        out              = ObstacleSizeArray()
+        out.header.stamp = msg.header.stamp
+        out.header.frame_id = msg.header.frame_id or 'laser'
+
+        for idx, cluster in enumerate(clusters):
+            if depth is not None:
+                est = estimate_cluster_size(
+                    cluster,
+                    depth,
+                    self._cam,
+                    depth_window = self._win_px,
+                    search_rows  = self._rows_px,
+                    col_hw       = self._col_hw,
+                    z_tol        = self._z_tol,
+                    obstacle_id  = idx,
+                )
+            else:
+                # No depth image available — publish LIDAR-only width
+                from ._obstacle_size import ObstacleSizeEstimate
+                import math as _math
+                cx = float(cluster.centroid[0])
+                cy = float(cluster.centroid[1])
+                est = ObstacleSizeEstimate(
+                    obstacle_id = idx,
+                    centroid_x  = cx,
+                    centroid_y  = cy,
+                    depth_z     = 0.0,
+                    width_m     = float(cluster.width_m),
+                    height_m    = 0.0,
+                    pixel_u     = -1,
+                    pixel_v     = -1,
+                    lidar_range = float(_math.sqrt(cx * cx + cy * cy)),
+                    confidence  = 0.0,
+                )
+
+            obs = ObstacleSize()
+            obs.header      = out.header
+            obs.obstacle_id = idx
+            obs.centroid_x  = float(est.centroid_x)
+            obs.centroid_y  = float(est.centroid_y)
+            obs.depth_z     = float(est.depth_z)
+            obs.width_m     = float(est.width_m)
+            obs.height_m    = float(est.height_m)
+            obs.pixel_u     = int(est.pixel_u)
+            obs.pixel_v     = int(est.pixel_v)
+            obs.lidar_range = float(est.lidar_range)
+            obs.confidence  = float(est.confidence)
+            out.obstacles.append(obs)
+
+        self._pub.publish(out)
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = ObstacleSizeNode()
+    try:
+        rclpy.spin(node)
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/jetson/ros2_ws/src/saltybot_bringup/setup.py b/jetson/ros2_ws/src/saltybot_bringup/setup.py
index 4d33740..e308187 100644
--- a/jetson/ros2_ws/src/saltybot_bringup/setup.py
+++ b/jetson/ros2_ws/src/saltybot_bringup/setup.py
@@ -55,6 +55,8 @@ setup(
             'obstacle_velocity       = saltybot_bringup.obstacle_velocity_node:main',
             # Lane/path edge detector (Issue #339)
             'path_edges              = saltybot_bringup.path_edges_node:main',
+            # Depth-based obstacle size estimator (Issue #348)
+            'obstacle_size           = saltybot_bringup.obstacle_size_node:main',
         ],
     },
 )
diff --git a/jetson/ros2_ws/src/saltybot_bringup/test/test_obstacle_size.py b/jetson/ros2_ws/src/saltybot_bringup/test/test_obstacle_size.py
new file mode 100644
index 0000000..d9b75af
--- /dev/null
+++ b/jetson/ros2_ws/src/saltybot_bringup/test/test_obstacle_size.py
@@ -0,0 +1,364 @@
+"""
+test_obstacle_size.py — pytest tests for _obstacle_size.py (no ROS2 required).
+
+Tests cover:
+  CameraParams  — defaults and construction
+  lidar_to_camera — coordinate transform
+  project_to_pixel — pinhole projection + bounds checks
+  sample_depth_median — empty/uniform/sparse depth images
+  estimate_height — flat/obstacle/edge cases
+  estimate_cluster_size — full pipeline with synthetic cluster + depth
+"""
+
+from __future__ import annotations
+
+import math
+from typing import List
+
+import numpy as np
+import pytest
+
+from saltybot_bringup._obstacle_size import (
+    CameraParams,
+    ObstacleSizeEstimate,
+    estimate_cluster_size,
+    estimate_height,
+    lidar_to_camera,
+    project_to_pixel,
+    sample_depth_median,
+)
+# Import Cluster helper from lidar_clustering for building test fixtures
+from saltybot_bringup._lidar_clustering import Cluster
+
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+def _default_params(**kw) -> CameraParams:
+    p = CameraParams()
+    for k, v in kw.items():
+        object.__setattr__(p, k, v)
+    return p
+
+
+def _blank_depth(h: int = 480, w: int = 640, val: int = 0) -> np.ndarray:
+    return np.full((h, w), val, dtype=np.uint16)
+
+
+def _obstacle_depth(
+    h: int = 480, w: int = 640,
+    depth_mm: int = 2000,
+    u_c: int = 320, v_c: int = 240,
+    half_w: int = 30, half_h: int = 60,
+) -> np.ndarray:
+    """Create a synthetic depth image with a rectangular obstacle."""
+    img = _blank_depth(h, w)
+    r0, r1 = max(0, v_c - half_h), min(h, v_c + half_h + 1)
+    c0, c1 = max(0, u_c - half_w), min(w, u_c + half_w + 1)
+    img[r0:r1, c0:c1] = depth_mm
+    return img
+
+
+def _make_cluster(
+    cx: float = 2.0, cy: float = 0.0,
+    width_m: float = 0.4, depth_m: float = 0.3,
+    n_pts: int = 10,
+) -> Cluster:
+    """Construct a minimal Cluster NamedTuple for testing."""
+    pts = np.column_stack([
+        np.full(n_pts, cx),
+        np.full(n_pts, cy),
+    ]).astype(np.float64)
+    centroid = np.array([cx, cy], dtype=np.float64)
+    bbox_min = centroid - np.array([width_m / 2, depth_m / 2])
+    bbox_max = centroid + np.array([width_m / 2, depth_m / 2])
+    return Cluster(
+        points=pts, centroid=centroid,
+        bbox_min=bbox_min, bbox_max=bbox_max,
+        width_m=width_m, depth_m=depth_m,
+    )
+
+
+# ── CameraParams ──────────────────────────────────────────────────────────────
+
+class TestCameraParams:
+    def test_defaults(self):
+        p = CameraParams()
+        assert p.fx == pytest.approx(383.0)
+        assert p.fy == pytest.approx(383.0)
+        assert p.cx == pytest.approx(320.0)
+        assert p.cy == pytest.approx(240.0)
+        assert p.width  == 640
+        assert p.height == 480
+        assert p.depth_scale == pytest.approx(0.001)
+        assert p.ey == pytest.approx(0.05)
+
+    def test_custom(self):
+        p = CameraParams(fx=400.0, fy=400.0, cx=330.0, cy=250.0)
+        assert p.fx == pytest.approx(400.0)
+
+
+# ── lidar_to_camera ───────────────────────────────────────────────────────────
+
+class TestLidarToCamera:
+    def test_forward_no_offset(self):
+        """A point directly in front of the LIDAR (y_lidar=0) → x_cam=0."""
+        p = CameraParams(ex=0.0, ey=0.0, ez=0.0)
+        x_cam, y_cam, z_cam = lidar_to_camera(3.0, 0.0, p)
+        assert x_cam == pytest.approx(0.0)
+        assert y_cam == pytest.approx(0.0)
+        assert z_cam == pytest.approx(3.0)
+
+    def test_lateral_maps_to_x(self):
+        """y_lidar positive (left) → x_cam negative (right convention)."""
+        p = CameraParams(ex=0.0, ey=0.0, ez=0.0)
+        x_cam, y_cam, z_cam = lidar_to_camera(0.0, 1.0, p)
+        assert x_cam == pytest.approx(-1.0)
+        assert z_cam == pytest.approx(0.0)
+
+    def test_extrinsic_offset_applied(self):
+        """Extrinsic translation is added to camera-frame coords."""
+        p = CameraParams(ex=0.1, ey=0.05, ez=-0.02)
+        x_cam, y_cam, z_cam = lidar_to_camera(2.0, 0.0, p)
+        assert x_cam == pytest.approx(0.1)      # -y_lidar + ex = 0 + 0.1
+        assert y_cam == pytest.approx(0.05)     # 0 + ey
+        assert z_cam == pytest.approx(1.98)     # 2.0 + (-0.02)
+
+    def test_zero_point(self):
+        p = CameraParams(ex=0.0, ey=0.0, ez=0.0)
+        x_cam, y_cam, z_cam = lidar_to_camera(0.0, 0.0, p)
+        assert z_cam == pytest.approx(0.0)
+
+
+# ── project_to_pixel ──────────────────────────────────────────────────────────
+
+class TestProjectToPixel:
+    def test_centre_projects_to_principal(self):
+        """A point exactly on the optical axis projects to (cx, cy)."""
+        p = CameraParams(fx=400.0, fy=400.0, cx=320.0, cy=240.0)
+        px = project_to_pixel(0.0, 0.0, 2.0, p)
+        assert px is not None
+        u, v = px
+        assert u == 320
+        assert v == 240
+
+    def test_negative_z_returns_none(self):
+        p = CameraParams()
+        assert project_to_pixel(0.0, 0.0, -1.0, p) is None
+
+    def test_zero_z_returns_none(self):
+        p = CameraParams()
+        assert project_to_pixel(0.0, 0.0, 0.0, p) is None
+
+    def test_off_to_right(self):
+        """A point to the right of axis should land to the right of cx."""
+        p = CameraParams(fx=400.0, cx=320.0, cy=240.0)
+        px = project_to_pixel(0.5, 0.0, 2.0, p)
+        assert px is not None
+        u, _ = px
+        assert u > 320
+
+    def test_out_of_image_returns_none(self):
+        p = CameraParams(fx=100.0, fy=100.0, cx=50.0, cy=50.0,
+                         width=100, height=100)
+        # Very far off-axis → outside image
+        px = project_to_pixel(10.0, 0.0, 1.0, p)
+        assert px is None
+
+    def test_known_projection(self):
+        """Verify exact pixel for known 3D point."""
+        p = CameraParams(fx=400.0, fy=400.0, cx=320.0, cy=240.0,
+                         width=640, height=480)
+        # x_cam=1.0, z_cam=4.0 → u = 400*(1/4)+320 = 420
+        # y_cam=0.0            → v = 400*(0/4)+240 = 240
+        px = project_to_pixel(1.0, 0.0, 4.0, p)
+        assert px == (420, 240)
+
+
+# ── sample_depth_median ───────────────────────────────────────────────────────
+
+class TestSampleDepthMedian:
+    def test_all_zeros_returns_zero(self):
+        img = _blank_depth()
+        d, n = sample_depth_median(img, 320, 240, window_px=5)
+        assert d == pytest.approx(0.0)
+        assert n == 0
+
+    def test_uniform_image(self):
+        """All pixels set to 2000 mm → median = 2.0 m."""
+        img = _blank_depth(val=2000)
+        d, n = sample_depth_median(img, 320, 240, window_px=5, depth_scale=0.001)
+        assert d == pytest.approx(2.0)
+        assert n > 0
+
+    def test_scale_applied(self):
+        img = _blank_depth(val=3000)
+        d, _ = sample_depth_median(img, 320, 240, window_px=3, depth_scale=0.001)
+        assert d == pytest.approx(3.0)
+
+    def test_window_clips_at_image_edge(self):
+        """Sampling near edge should not crash."""
+        img = _blank_depth(val=1500)
+        d, n = sample_depth_median(img, 0, 0, window_px=10)
+        assert d > 0.0
+        assert n > 0
+
+    def test_sparse_window(self):
+        """Only some pixels are valid — median should reflect valid ones."""
+        img = _blank_depth()
+        img[240, 320] = 1000   # single valid pixel
+        d, n = sample_depth_median(img, 320, 240, window_px=5, depth_scale=0.001)
+        assert d == pytest.approx(1.0)
+        assert n == 1
+
+    def test_mixed_window_median(self):
+        """Median of [1000, 2000, 3000] mm = 2000 mm = 2.0 m."""
+        img = _blank_depth()
+        img[240, 319] = 1000
+        img[240, 320] = 2000
+        img[240, 321] = 3000
+        d, n = sample_depth_median(img, 320, 240, window_px=1, depth_scale=0.001)
+        assert d == pytest.approx(2.0)
+        assert n == 3
+
+
+# ── estimate_height ───────────────────────────────────────────────────────────
+
+class TestEstimateHeight:
+    def test_blank_image_returns_zero(self):
+        p = CameraParams(fy=383.0)
+        img = _blank_depth()
+        h = estimate_height(img, 320, 240, z_ref=2.0, params=p)
+        assert h == pytest.approx(0.0)
+
+    def test_zero_z_ref_returns_zero(self):
+        p = CameraParams()
+        img = _blank_depth(val=2000)
+        h = estimate_height(img, 320, 240, z_ref=0.0, params=p)
+        assert h == pytest.approx(0.0)
+
+    def test_known_height(self):
+        """
+        Obstacle occupies rows 180–300 (120 row-pixel span) at depth 2.0m.
+        Expected height ≈ 120 * 2.0 / 383.0 ≈ 0.627 m.
+        """
+        p = CameraParams(fy=383.0, depth_scale=0.001)
+        img = _blank_depth()
+        # Obstacle at z=2000 mm, rows 180–300, cols 300–340
+        img[180:301, 300:341] = 2000
+        h = estimate_height(img, 320, 240, z_ref=2.0, params=p,
+                             search_rows=200, col_hw=30, z_tol=0.3)
+        expected = 120 * 2.0 / 383.0
+        assert h == pytest.approx(expected, rel=0.05)
+
+    def test_single_row_returns_zero(self):
+        """Only one valid row → span=0 → height=0."""
+        p = CameraParams(fy=383.0, depth_scale=0.001)
+        img = _blank_depth()
+        img[240, 315:326] = 2000   # single row
+        h = estimate_height(img, 320, 240, z_ref=2.0, params=p,
+                             search_rows=20, col_hw=10, z_tol=0.3)
+        assert h == pytest.approx(0.0)
+
+    def test_depth_outside_tolerance_excluded(self):
+        """Pixels far from z_ref should not contribute to height."""
+        p = CameraParams(fy=383.0, depth_scale=0.001)
+        img = _blank_depth()
+        # Two rows at very different depths — only one within z_tol
+        img[220, 315:326] = 2000   # z = 2.0 m (within tolerance)
+        img[260, 315:326] = 5000   # z = 5.0 m (outside tolerance)
+        h = estimate_height(img, 320, 240, z_ref=2.0, params=p,
+                             search_rows=100, col_hw=10, z_tol=0.3)
+        assert h == pytest.approx(0.0)   # only 1 valid row → 0
+
+
+# ── estimate_cluster_size ─────────────────────────────────────────────────────
+
+class TestEstimateClusterSize:
+    def test_returns_named_tuple(self):
+        cluster = _make_cluster()
+        img = _blank_depth(val=2000)
+        est = estimate_cluster_size(cluster, img, CameraParams())
+        assert isinstance(est, ObstacleSizeEstimate)
+
+    def test_width_comes_from_lidar(self):
+        """Width should equal the LIDAR cluster width_m."""
+        cluster = _make_cluster(cx=3.0, cy=0.0, width_m=0.5)
+        img = _blank_depth(val=3000)
+        est = estimate_cluster_size(cluster, img, CameraParams())
+        assert est.width_m == pytest.approx(0.5)
+
+    def test_centroid_preserved(self):
+        cluster = _make_cluster(cx=2.0, cy=0.3)
+        img = _blank_depth(val=2000)
+        est = estimate_cluster_size(cluster, img, CameraParams())
+        assert est.centroid_x == pytest.approx(2.0)
+        assert est.centroid_y == pytest.approx(0.3)
+
+    def test_lidar_range_correct(self):
+        cluster = _make_cluster(cx=3.0, cy=4.0)  # range = 5.0
+        img = _blank_depth(val=5000)
+        est = estimate_cluster_size(cluster, img, CameraParams())
+        assert est.lidar_range == pytest.approx(5.0)
+
+    def test_blank_depth_gives_zero_confidence(self):
+        """No valid depth pixels → confidence=0, depth_z falls back to z_cam."""
+        cluster = _make_cluster(cx=2.0, cy=0.0)
+        img = _blank_depth()
+        p   = CameraParams(ex=0.0, ey=0.0, ez=0.0)
+        est = estimate_cluster_size(cluster, img, p)
+        assert est.confidence == pytest.approx(0.0)
+        assert est.depth_z > 0.0  # LIDAR-fallback z_cam
+
+    def test_obstacle_in_depth_image(self):
+        """With a real depth patch, confidence > 0 and pixel is in-image."""
+        # Cluster at x=2m forward, y=0 → z_cam=2m, x_cam=0, y_cam=ey
+        p = CameraParams(fy=383.0, fx=383.0, cx=320.0, cy=240.0,
+                         depth_scale=0.001, ex=0.0, ey=0.05, ez=0.0)
+        cluster = _make_cluster(cx=2.0, cy=0.0)
+        # z_cam=2m, x_cam=0, y_cam=0.05
+        # u = 383*0/2+320 = 320, v = 383*0.05/2+240 ≈ 249
+        img = _obstacle_depth(depth_mm=2000, u_c=320, v_c=249,
+                               half_w=30, half_h=80)
+        est = estimate_cluster_size(cluster, img, p,
+                                    depth_window=5, search_rows=120,
+                                    col_hw=10, z_tol=0.3, obstacle_id=1)
+        assert est.confidence > 0.0
+        assert 0 <= est.pixel_u < 640
+        assert 0 <= est.pixel_v < 480
+        assert est.depth_z == pytest.approx(2.0, abs=0.1)
+        assert est.obstacle_id == 1
+
+    def test_behind_camera_returns_zero_confidence(self):
+        """Cluster behind camera (x_lidar < 0 with no ez) → not projected."""
+        p = CameraParams(ex=0.0, ey=0.0, ez=0.0)
+        # x_lidar = -1.0 → z_cam = -1 + 0 = -1 → behind camera
+        cluster = _make_cluster(cx=-1.0, cy=0.0)
+        img = _blank_depth(val=1000)
+        est = estimate_cluster_size(cluster, img, p)
+        assert est.confidence == pytest.approx(0.0)
+        assert est.pixel_u == -1
+        assert est.pixel_v == -1
+
+    def test_height_estimated_from_depth(self):
+        """When obstacle spans rows, height_m > 0."""
+        p = CameraParams(fy=383.0, fx=383.0, cx=320.0, cy=240.0,
+                         depth_scale=0.001, ex=0.0, ey=0.05, ez=0.0)
+        cluster = _make_cluster(cx=2.0, cy=0.0)
+        # Expected pixel: u≈320, v≈249; give obstacle a 100-row span
+        img = _obstacle_depth(depth_mm=2000, u_c=320, v_c=249,
+                               half_w=20, half_h=50)
+        est = estimate_cluster_size(cluster, img, p,
+                                    search_rows=150, col_hw=15, z_tol=0.3)
+        assert est.height_m > 0.0
+
+    def test_obstacle_id_passed_through(self):
+        cluster = _make_cluster()
+        img = _blank_depth(val=2000)
+        est = estimate_cluster_size(cluster, img, CameraParams(), obstacle_id=42)
+        assert est.obstacle_id == 42
+
+    def test_confidence_bounded(self):
+        cluster = _make_cluster(cx=2.0)
+        img = _blank_depth(val=2000)
+        est = estimate_cluster_size(cluster, img, CameraParams())
+        assert 0.0 <= est.confidence <= 1.0
diff --git a/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt b/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt
index 9db8a9a..ec57f1d 100644
--- a/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/CMakeLists.txt
@@ -27,6 +27,9 @@ rosidl_generate_interfaces(${PROJECT_NAME}
   "msg/ObstacleVelocityArray.msg"
   # Issue #339 — lane/path edge detector
   "msg/PathEdges.msg"
+  # Issue #348 — depth-based obstacle size estimator
+  "msg/ObstacleSize.msg"
+  "msg/ObstacleSizeArray.msg"
   DEPENDENCIES std_msgs geometry_msgs vision_msgs builtin_interfaces
 )
 
diff --git a/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSize.msg b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSize.msg
new file mode 100644
index 0000000..a8aa872
--- /dev/null
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSize.msg
@@ -0,0 +1,27 @@
+# ObstacleSize.msg — Depth-projected LIDAR cluster size estimate (Issue #348)
+#
+# Fuses a 2-D LIDAR cluster with the D435i depth image to estimate the full
+# 3-D size (width × height) of each detected obstacle.
+#
+# obstacle_id : matches the obstacle_id in ObstacleVelocity (same LIDAR cluster)
+# centroid_x  : LIDAR-frame forward distance to centroid (metres)
+# centroid_y  : LIDAR-frame lateral distance to centroid (metres, +Y = left)
+# depth_z     : sampled D435i depth at projected centroid (metres, 0 = unknown)
+# width_m     : horizontal size from LIDAR bbox (metres)
+# height_m    : vertical size from D435i depth strip (metres, 0 = unknown)
+# pixel_u     : projected centroid column in depth image (pixels)
+# pixel_v     : projected centroid row    in depth image (pixels)
+# lidar_range : range from LIDAR origin to centroid (metres)
+# confidence  : 0.0–1.0; based on depth sample quality and cluster track age
+#
+std_msgs/Header header
+uint32  obstacle_id
+float32 centroid_x
+float32 centroid_y
+float32 depth_z
+float32 width_m
+float32 height_m
+int32   pixel_u
+int32   pixel_v
+float32 lidar_range
+float32 confidence
diff --git a/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSizeArray.msg b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSizeArray.msg
new file mode 100644
index 0000000..43ef0bd
--- /dev/null
+++ b/jetson/ros2_ws/src/saltybot_scene_msgs/msg/ObstacleSizeArray.msg
@@ -0,0 +1,4 @@
+# ObstacleSizeArray.msg — All depth-projected obstacle size estimates (Issue #348)
+
+std_msgs/Header  header
+ObstacleSize[]   obstacles