diff --git a/src/backend/services/camera.py b/src/backend/services/camera.py
index e853be3..54a7fa3 100644
--- a/src/backend/services/camera.py
+++ b/src/backend/services/camera.py
@@ -47,6 +47,11 @@ class CameraFrame:
     timestamp: float
 
 
+def _owned(array: np.ndarray) -> np.ndarray:
+    """`array` if it owns its memory, else a copy (a view on an SDK buffer has a `base`)."""
+    return array if array.base is None else array.copy()
+
+
 @dataclass
 class CameraIntrinsics:
     """Per-sensor intrinsics. `*_ir` = depth, `*_rgb` = color."""
@@ -389,10 +394,11 @@ class RealSenseD435:
     # Frame capture
     # ------------------------------------------------------------------
     def capture(self, timeout_ms: int = 5000) -> Optional[CameraFrame]:
-        """The next frame acquired after this call (None on timeout or when closed).
+        """The next frame COMPLETED after this call (None on timeout or when closed).
 
-        Does not touch the pipeline: it waits for the acquisition thread.
-        Requests waiting at the same moment get the same frame.
+        A grab already in progress when the call arrives may have started
+        before it. Does not touch the pipeline: it waits for the acquisition
+        thread. Requests waiting at the same moment get the same frame.
         """
         acquisition = self._acquisition
         if not self._open or acquisition is None:
@@ -423,7 +429,11 @@ class RealSenseD435:
             return None
 
         with timer.phase("convert"):
-            return self._frame_from(color_frame, depth_frame)
+            frame = self._frame_from(color_frame, depth_frame)
+        # Own copies: frames now live on after the grab (latest frame, 100 ms cache,
+        # shared encodings), and views would keep the SDK's frame pool busy.
+        with timer.phase("copy"):
+            return CameraFrame(rgb=_owned(frame.rgb), depth=_owned(frame.depth), timestamp=frame.timestamp)
 
     def _frame_from(self, color_frame, depth_frame) -> CameraFrame:
         """numpy BGR + depth in mm from the SDK frames (YUYV decode, USB 2 resize, depth scale)."""