feat: live count worker, tracking, and page updates

backend/counting_bench.py

@@ -25,6 +25,7 @@ DEFAULTS = {
     "entry_travel_min": 60.0, "handoff_radius": 100.0,
     "unload_confirm_frames": 3, "min_area_scale": 1.0,
     "dedup_radius": 60.0, "spatial_dedup": False,
+    "count_classes": ["sack"],
 }
 
 
@@ -176,7 +177,8 @@ def count_video(path: str, model, params: dict, should_cancel=None,
         raise CountingBenchError(f"Could not open {path}")
     total_frames = int(capture.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
 
-    tracker = ByteTrackTracker(model, settings["conf"])
+    tracker = ByteTrackTracker(model, settings["conf"],
+                                 class_filter=tuple(settings["count_classes"]))
     stabilizer = BboxStabilizer(ema_alpha=0.35, max_hold_frames=10,
                                 max_height_ratio=1.5, min_height_ratio=0.70)
     counter = LineCrossCounter(
@@ -199,7 +201,8 @@ def count_video(path: str, model, params: dict, should_cancel=None,
             if not ok or frame is None:
                 break
             frame = cv2.resize(frame, (1280, 720))
-            detections = [d for d in tracker.update(frame, []) if d.class_name == "sack"]
+            # The tracker only emits count_classes, so no post-filter is needed.
+            detections = tracker.update(frame, [])
             stable = stabilizer.update(detections)
             inside = [
                 d for d in stable