feat: merged detection output + parallel model execution

- Multiple models run simultaneously on each frame via ThreadPoolExecutor
- Detections merged with IoU deduplication (same class, high overlap → keep highest conf)
- Single merged_annotated.mp4 output (not per-model videos)
- Single model still uses existing run_pipeline for backward compat
- Active jobs banner on upload page (polls every 5s)
- GPU warning when selecting 3+ models
- 62/62 tests pass

app.py

@@ -255,6 +255,16 @@ def api_cancel_job(job_id):
     return jsonify({"error": "cannot cancel job in state " + job.status.name}), 400
 
 
+@app.route("/api/jobs/active")
+def api_active_jobs():
+    """Return count of running/pending jobs."""
+    active = [j for j in job_queue.list_jobs() if j.status.name in ('RUNNING', 'PENDING')]
+    return jsonify({
+        "count": len(active),
+        "jobs": [{"job_id": j.job_id, "status": j.status.name, "video": os.path.basename(j.video_path)} for j in active],
+    })
+
+
 @app.route("/api/jobs/<job_id>", methods=["DELETE"])
 def api_delete_job(job_id):
     """Delete a completed/failed/cancelled job."""

src/job.py

@@ -12,7 +12,7 @@ from enum import Enum, auto
 from pathlib import Path
 
 from src.model_registry import ModelConfig, find_truck_detector
-from src.pipeline import run_pipeline, PipelineResult
+from src.pipeline import run_pipeline, run_merged_pipeline, PipelineResult, MergedPipelineResult
 
 
 class JobStatus(Enum):
@@ -172,26 +172,15 @@ class JobQueue:
             return
 
         try:
-            for i, model_cfg in enumerate(job.model_configs):
-                with self._lock:
-                    if job.status == JobStatus.CANCELLED:
-                        break
-                    job.current_model = model_cfg.filename
-                    job.progress = i / total_models
+            if total_models == 1:
+                # Single model: use existing run_pipeline
+                model_cfg = job.model_configs[0]
                 class_filter = job.class_filters.get(model_cfg.filename)
-
-                # Auto-select truck detector if model lacks "truck" class
                 truck_model_config = None
                 if "truck" not in (model_cfg.known_classes or []):
-                    truck_model_config = find_truck_detector(os.path.join(os.path.dirname(model_cfg.path), "..", "models"))
-                    # If not found relative, try absolute models dir
-                    if truck_model_config is None:
                     truck_model_config = find_truck_detector("./models")
 
-                output_path = os.path.join(
-                    job.output_dir,
-                    f"{model_cfg.stem}_annotated.mp4",
-                )
+                output_path = os.path.join(job.output_dir, f"{model_cfg.stem}_annotated.mp4")
 
                 def _on_live_stats(stats):
                     with self._lock:
@@ -200,7 +189,7 @@ class JobQueue:
                         job.live_unloading = stats["unloading"]
                         job.live_frame = stats["frame"]
 
-                result: PipelineResult = run_pipeline(
+                result = run_pipeline(
                     video_path=job.video_path,
                     model_config=model_cfg,
                     output_path=output_path,
@@ -214,8 +203,7 @@ class JobQueue:
                 )
 
                 with self._lock:
-                    job.results.append(
-                        JobResult(
+                    job.results.append(JobResult(
                         model_name=model_cfg.filename,
                         output_path=result.output_path,
                         loading_count=result.loading_count,
@@ -224,17 +212,50 @@ class JobQueue:
                         batch_count=result.batch_count,
                         frame_count=result.frame_count,
                         duration_seconds=result.duration_seconds,
-                        )
+                    ))
+            else:
+                # Multiple models: run merged pipeline
+                with self._lock:
+                    job.current_model = f"Processing {total_models} models (merged)..."
+
+                output_path = os.path.join(job.output_dir, "merged_annotated.mp4")
+
+                def _on_live_stats_merged(stats):
+                    with self._lock:
+                        job.live_fps = stats["fps"]
+                        job.live_loading = stats["loading"]
+                        job.live_unloading = stats["unloading"]
+                        job.live_frame = stats["frame"]
+
+                # Auto-select truck detector
+                truck_model_config = None
+                if not any("truck" in (c.known_classes or []) for c in job.model_configs):
+                    truck_model_config = find_truck_detector("./models")
+
+                result = run_merged_pipeline(
+                    video_path=job.video_path,
+                    model_configs=job.model_configs,
+                    output_path=output_path,
+                    class_filters=job.class_filters,
+                    truck_model_config=truck_model_config,
+                    preview_path=job.preview_path,
+                    preview_enabled=job.preview_enabled,
+                    preview_every_n=2,
+                    cancel_check=lambda: job.status == JobStatus.CANCELLED,
+                    live_stats_callback=_on_live_stats_merged,
                 )
 
-                # Force release GPU memory between models
-                import gc
-                gc.collect()
-                try:
-                    import torch
-                    torch.cuda.empty_cache()
-                except Exception:
-                    pass
+                with self._lock:
+                    job.results.append(JobResult(
+                        model_name="Merged (" + ", ".join(c.stem for c in job.model_configs) + ")",
+                        output_path=result.output_path,
+                        loading_count=result.loading_count,
+                        unloading_count=result.unloading_count,
+                        net_count=result.net_count,
+                        batch_count=result.batch_count,
+                        frame_count=result.frame_count,
+                        duration_seconds=result.duration_seconds,
+                    ))
 
             with self._lock:
                 if job.status != JobStatus.CANCELLED:

src/pipeline.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import os
 import time
 from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass
 
 import cv2
@@ -41,6 +42,24 @@ class PipelineResult:
         return self.loading_count - self.unloading_count
 
 
+@dataclass
+class MergedPipelineResult:
+    """Summary of a completed merged pipeline run (multiple models)."""
+
+    output_path: str
+    frame_count: int
+    loading_count: int
+    unloading_count: int
+    batch_count: int
+    duration_seconds: float
+    model_names: list[str]
+    class_filter: list[str] | None
+
+    @property
+    def net_count(self) -> int:
+        return self.loading_count - self.unloading_count
+
+
 def apply_class_filter(
     detections: list[Detection], class_filter: list[str] | None
 ) -> list[Detection]:
@@ -306,3 +325,247 @@ def run_pipeline(
         model_name=model_config.filename,
         class_filter=effective_filter,
     )
+
+
+def _compute_iou(box1: tuple, box2: tuple) -> float:
+    x1 = max(box1[0], box2[0])
+    y1 = max(box1[1], box2[1])
+    x2 = min(box1[2], box2[2])
+    y2 = min(box1[3], box2[3])
+    inter = max(0, x2 - x1) * max(0, y2 - y1)
+    area1 = (box1[2] - box1[0]) * (box1[3] - box1[1])
+    area2 = (box2[2] - box2[0]) * (box2[3] - box2[1])
+    union = area1 + area2 - inter
+    return inter / union if union > 0 else 0.0
+
+
+def _deduplicate_detections(dets: list[Detection], iou_threshold: float = 0.5) -> list[Detection]:
+    if not dets:
+        return dets
+    dets = sorted(dets, key=lambda d: d.confidence, reverse=True)
+    kept: list[Detection] = []
+    for det in dets:
+        is_dup = False
+        for existing in kept:
+            if existing.class_name == det.class_name:
+                if _compute_iou(existing.bbox, det.bbox) > iou_threshold:
+                    is_dup = True
+                    break
+        if not is_dup:
+            kept.append(det)
+    return kept
+
+
+def run_merged_pipeline(
+    video_path: str,
+    model_configs: list[ModelConfig],
+    output_path: str,
+    class_filters: dict[str, list[str] | None] | None = None,
+    truck_model_config: ModelConfig | None = None,
+    sack_conf: float = 0.4,
+    truck_conf: float = 0.5,
+    truck_det_interval: int = 15,
+    progress_callback: Callable[[int, int], None] | None = None,
+    cancel_check: Callable[[], bool] | None = None,
+    preview_path: str | None = None,
+    preview_enabled: bool = True,
+    preview_every_n: int = 2,
+    preview_max_dim: int = 480,
+    preview_jpeg_quality: int = 60,
+    live_stats_callback: Callable[[dict], None] | None = None,
+) -> MergedPipelineResult:
+    """Run multiple models on the same video, merge detections into one output."""
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise RuntimeError(f"Cannot open video: {video_path}")
+
+    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+    # Load all detectors
+    detectors = []
+    for cfg in model_configs:
+        if cfg.path.endswith(".onnx"):
+            try:
+                import onnxruntime
+            except ImportError:
+                raise RuntimeError(
+                    f"onnxruntime not installed. Cannot load '{cfg.filename}'. "
+                    f"Use the .engine version instead."
+                )
+        model = YOLO(cfg.path)
+        eff_filter = (class_filters or {}).get(cfg.stem) or cfg.known_classes or None
+        detectors.append(BaseDetector(model, conf=sack_conf, class_filter=eff_filter))
+
+    # Truck detector
+    truck_det = None
+    if truck_model_config is not None:
+        truck_shared = YOLO(truck_model_config.path)
+        truck_det = BaseDetector(truck_shared, conf=truck_conf, class_filter=("truck",))
+    elif len(model_configs) == 1 and "truck" in (model_configs[0].known_classes or []):
+        truck_det = detectors[0]
+
+    # Single tracker using first model's weights
+    tracker_model = YOLO(model_configs[0].path)
+    tracker = ByteTrackTracker(tracker_model, conf=sack_conf)
+    stabilizer = BboxStabilizer()
+    roi_tracker = TruckROITracker(frame_width=w, frame_height=h)
+    counter = LineCrossCounter(
+        line_y=int(h * 0.50),
+        line_x_start=int(w * 0.38),
+        line_x_end=int(w * 0.72),
+        margin=20,
+    )
+    batch_mgr = BatchLifecycleManager()
+    dashboard = DashboardOverlay()
+    writer = AnnotatedVideoWriter(output_path, fps=fps, frame_size=(w, h))
+
+    start_time = time.time()
+    frame_idx = 0
+    completed_batches = 0
+    _frame_times: list[float] = []
+
+    def on_batch_end(record):
+        nonlocal completed_batches
+        completed_batches += 1
+    batch_mgr.on_batch_end(on_batch_end)
+
+    try:
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+
+            frame_idx += 1
+            timestamp = time.time()
+            _frame_times.append(timestamp)
+            if len(_frame_times) > 30:
+                _frame_times.pop(0)
+
+            if cancel_check and frame_idx % 30 == 0 and cancel_check():
+                break
+
+            # Truck detection
+            roi = roi_tracker.roi
+            if truck_det is not None and frame_idx % truck_det_interval == 0:
+                trucks = truck_det.detect(frame)
+                roi = roi_tracker.update(trucks)
+
+            truck_present = roi is not None and roi.confidence > 0
+            if roi is not None:
+                counter.line_y = roi.line_y
+                counter.line_x_start = roi.x1
+                counter.line_x_end = roi.x2
+
+            truck_centroid = None
+            if roi is not None:
+                truck_centroid = ((roi.x1 + roi.x2) / 2.0, (roi.y1 + roi.y2) / 2.0)
+
+            if frame_idx % truck_det_interval == 0:
+                batch_mgr.update(
+                    truck_detected=truck_present,
+                    timestamp=timestamp,
+                    loading_count=counter.loading_count,
+                    unloading_count=counter.unloading_count,
+                    truck_centroid=truck_centroid,
+                )
+
+            # Run all detectors in parallel
+            all_detections: list[Detection] = []
+            if batch_mgr.is_active:
+                with ThreadPoolExecutor(max_workers=len(detectors)) as pool:
+                    futures = {pool.submit(d.detect, frame): d for d in detectors}
+                    for future in as_completed(futures):
+                        try:
+                            all_detections.extend(future.result())
+                        except Exception:
+                            pass
+
+                # Deduplicate
+                merged = _deduplicate_detections(all_detections)
+
+                # Stabilize and filter to ROI
+                stable = stabilizer.update(merged)
+                if roi is not None:
+                    tracked = [d for d in stable if roi.contains_x((d.bbox[0] + d.bbox[2]) / 2.0)]
+                else:
+                    tracked = stable
+
+                counter.update(tracked)
+            else:
+                tracked = []
+
+            # Annotate
+            viz = dashboard.draw(
+                frame=frame,
+                detections=tracked,
+                roi=roi,
+                loading_count=counter.loading_count,
+                unloading_count=counter.unloading_count,
+                batch_id=batch_mgr.current_batch_id,
+                history=batch_mgr.history,
+                system_state=batch_mgr.state,
+                batch_duration=batch_mgr.batch_duration,
+                stabilize_progress=batch_mgr.stabilize_progress,
+                waiting_duration=batch_mgr.waiting_duration,
+            )
+
+            # Model info overlay
+            model_names = ", ".join(c.stem for c in model_configs)
+            cv2.putText(viz, f"Models: {model_names}", (10, h - 50), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (200, 200, 200), 1)
+
+            writer.write_frame(viz)
+
+            # Live stats
+            if live_stats_callback and frame_idx % 10 == 0:
+                if len(_frame_times) >= 2:
+                    dt = _frame_times[-1] - _frame_times[0]
+                    processing_fps = (len(_frame_times) - 1) / dt if dt > 0 else 0.0
+                else:
+                    processing_fps = 0.0
+                live_stats_callback({
+                    "fps": round(processing_fps, 1),
+                    "loading": counter.loading_count,
+                    "unloading": counter.unloading_count,
+                    "frame": frame_idx,
+                })
+
+            # Preview
+            if preview_enabled and preview_path is not None and frame_idx % max(1, preview_every_n) == 0:
+                ph, pw = viz.shape[:2]
+                if max(ph, pw) > preview_max_dim:
+                    scale = preview_max_dim / max(ph, pw)
+                    preview_viz = cv2.resize(viz, (int(pw * scale), int(ph * scale)))
+                else:
+                    preview_viz = viz
+                tmp_path = preview_path + ".tmp.jpg"
+                cv2.imwrite(tmp_path, preview_viz, [cv2.IMWRITE_JPEG_QUALITY, preview_jpeg_quality])
+                os.replace(tmp_path, preview_path)
+
+            if progress_callback:
+                progress_callback(frame_idx, total_frames)
+
+    finally:
+        cap.release()
+        writer.finish()
+        import gc
+        gc.collect()
+        try:
+            import torch
+            torch.cuda.empty_cache()
+        except Exception:
+            pass
+
+    duration = time.time() - start_time
+    return MergedPipelineResult(
+        output_path=output_path,
+        frame_count=frame_idx,
+        loading_count=counter.loading_count,
+        unloading_count=counter.unloading_count,
+        batch_count=completed_batches,
+        duration_seconds=duration,
+        model_names=[c.filename for c in model_configs],
+        class_filter=None,
+    )

static/app.js

@@ -270,7 +270,14 @@
      Submit — enable/disable, loading state
   ------------------------------------------------ */
   function initFormSubmit() {
-    form.addEventListener('submit', function () {
+    form.addEventListener('submit', function (e) {
+      var checked = modelGrid.querySelectorAll('.model-card-check:checked');
+      if (checked.length >= 3) {
+        if (!confirm('Running ' + checked.length + ' models simultaneously may use significant GPU memory on Jetson. Continue?')) {
+          e.preventDefault();
+          return;
+        }
+      }
       if (videoPathInput.value) {
         form.action = '/upload/reuse';
       } else {
@@ -304,6 +311,27 @@
   initModelCards();
   initFormSubmit();
   loadExistingVideos();
+
+  // Poll active jobs
+  function checkActiveJobs() {
+    fetch('/api/jobs/active')
+      .then(function(res) { return res.json(); })
+      .then(function(data) {
+        var banner = document.getElementById('active-jobs-banner');
+        var text = document.getElementById('active-jobs-text');
+        if (banner && text) {
+          if (data.count > 0) {
+            banner.style.display = '';
+            text.textContent = data.count + ' job' + (data.count > 1 ? 's' : '') + ' in progress';
+          } else {
+            banner.style.display = 'none';
+          }
+        }
+      })
+      .catch(function() {});
+  }
+  checkActiveJobs();
+  setInterval(checkActiveJobs, 5000);
 })();
 
 function formatRelativeTime(timestamp) {

templates/index.html

@@ -257,6 +257,19 @@
   }
   .empty-state p { margin-bottom: var(--space-2); }
 
+  /* Active jobs banner */
+  .active-jobs-banner {
+    display: flex;
+    align-items: center;
+    gap: var(--space-3);
+    padding: var(--space-3) var(--space-4);
+    background: #f0f9ff;
+    border: 1px solid #0EA5E9;
+    border-radius: var(--radius-md);
+    margin-bottom: var(--space-6);
+    font-size: var(--text-sm);
+  }
+
   /* Video delete button */
   .video-delete-btn {
     flex-shrink: 0;
@@ -291,6 +304,12 @@
   <p>Upload a video and select detection models to run analysis.</p>
 </div>
 
+<div class="active-jobs-banner" id="active-jobs-banner" style="display:none">
+  <span class="badge badge-info">&#9654;</span>
+  <span id="active-jobs-text"></span>
+  <a href="/jobs" class="btn btn-ghost btn-sm">View All</a>
+</div>
+
 <form action="/upload" method="post" enctype="multipart/form-data" id="upload-form">
   <input type="hidden" name="video_path" id="video-path-input" value="">