fix: gpu memory leaks

services/tracking_controller.py

@@ -272,12 +272,14 @@ class ObjectTracker:
         for tid in stale_track_ids:
             del self._tracks[tid]
 
-    def update(self, detections: List[Detection]) -> List[TrackedObject]:
+    def update(self, detections: List[Detection], frame_shape: tuple = None, model_input_size: int = 640) -> List[TrackedObject]:
         """
         Update tracker with new detections (decoupled from inference).
 
         Args:
             detections: List of Detection objects from model inference
+            frame_shape: Original frame shape (C, H, W) for scaling bboxes back from model space
+            model_input_size: Model input size (default: 640 for YOLOv8)
 
         Returns:
             List of currently tracked objects
@@ -291,6 +293,22 @@ class ObjectTracker:
                 self._cleanup_stale_tracks()
                 return list(self._tracks.values())
 
+            # Scale detections from model space (640x640) to frame space (H x W)
+            if frame_shape is not None:
+                _, frame_h, frame_w = frame_shape
+                scale_x = frame_w / model_input_size
+                scale_y = frame_h / model_input_size
+
+                # Scale all detection bboxes
+                for det in detections:
+                    x1, y1, x2, y2 = det.bbox
+                    det.bbox = [
+                        x1 * scale_x,
+                        y1 * scale_y,
+                        x2 * scale_x,
+                        y2 * scale_y
+                    ]
+
             # Convert detections to tensor for GPU processing
             det_tensor = torch.tensor(
                 [[*det.bbox, det.confidence, det.class_id] for det in detections],