feat: unified config.yaml with extensible model presets (A-D data-driven, E/F ready)

- config.yaml canonical for stream/models/counting/batch/output/camera
- models.modes hold engines+class filters only; conf/iou/min_bbox in detection_params
- predict.py derives tracker roles structurally (no per-mode branching)
- dashboard mode switch validates + persists atomically to config.yaml
- .env keeps secrets/deployment only; zones.json geometry; tracker.yaml hyperparams
- batch_mode.json keeps manual/auto batch mode; legacy model_mode ignored w/ warning
- src/detection+tracking gain iou param (default 0.7 = no behavior change)

.env.example

@@ -1,24 +1,34 @@
-# Core variables for Karung Counter
+# .env — SECRETS & deployment-specific values ONLY (never commit real credentials).
+# Canonical config lives in config.yaml (paths, modes, counting knobs, batch, camera).
+# Zone polygons live in zones.json. Tracker hyperparams live in cfg/tracker.yaml.
+#
+# Legacy keys (MODEL_PATH, MODEL_MODE, BATCH_MODE_FILE, *_PATH overrides) are
+# still honoured when set, but deprecated — prefer config.yaml.
+
+# --- Output paths (override config.yaml output.* when set; legacy) ---
 OUTPUT_DIR=/opt/jetson-counter
 DB_PATH=/opt/jetson-counter/jetson_counter.db
 STATE_FILE=/opt/jetson-counter/current_batch.json
-BATCH_MODE_FILE=/opt/jetson-counter/batch_mode.json
 LIVE_STREAM_FRAME_PATH=/dev/shm/jetson-counter/live_frame.jpg
+
+# --- Identity / batch day ---
 CAMERA_NAME=CC1
 OBJECT_LABEL=karung-pakan
 DAILY_CUTOFF_TIME=06:00
+BATCH_MERGE_THRESHOLD_SECONDS=300
 
-# Dashboard variables
+# --- Dashboard (canonical home for these; read from env, NOT config.yaml) ---
 SECRET_KEY=change-me-in-production
 DASHBOARD_HOST=0.0.0.0
 DASHBOARD_PORT=5000
 OFFICE_PORT=5721
 FLASK_DEBUG=false
+SITE_NAME=LIVE
 
-# Stream URL
+# --- Stream (canonical home; read from env, NOT config.yaml) ---
 RTSP_URL=rtsp://user:pass@192.168.192.209:8554/camera_stream_640
 
-# Model pipeline mode: A=combined only; B=v4 truck + yolo11n sack+box;
-# C=A + yolo11n box-only (default); D=v4 truck + best sack-only + yolo11n box-only.
-# Dashboard switches persist here and apply on next service restart.
-MODEL_MODE=C
+# --- Deprecated: model pipeline mode now lives in config.yaml models.active_mode.
+# Dashboard switches persist there (atomic write, comments preserved) and apply
+# on next `karung-counter` restart. Only set this to temporarily override.
+# MODEL_MODE=C

.github/workflows/ci.yml

@@ -12,6 +12,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.10"
-      - run: pip install numpy python-dotenv pytest
+      - run: pip install numpy python-dotenv pytest pyyaml
       - run: python -m pytest tests/ -q
       - run: python -m compileall -q src/ counter_dashboard.py

AGENTS.md

@@ -9,12 +9,14 @@ Indonesian (`karung`=sack, `truk`=truck); YOLO class names are English
 ## Which pipeline to touch
 
 - `predict.py` = **production AND dev CLI** (runs as `karung-counter.service` with
-  zero args). Dev flags: `--source VID --env .env --model X --output-dir D
+  zero args). Dev flags: `--source VID --env .env --config YAML --model X --output-dir D
   --output-json F --sack-conf C --truck-conf C --box-conf C --box-model P
-  --model-mode A|B|C|D --batch-timeout S --max-frames N --no-dashboard --no-db`.
-  Zero flags = systemd behaviour (`MODEL_MODE` env, default C).
-  Model modes: A=combined only; B=v4 truck + yolo11n sack+box;
-  C=A + yolo11n box-only (default); D=v4 truck + best sack-only + yolo11n box-only.
+  --model-mode M --batch-timeout S --max-frames N --no-dashboard --no-db`.
+  Zero flags = systemd behaviour (`config.yaml` + `.env`).
+  Model modes are DATA in `config.yaml` `models.modes` (engines + class filters
+  only; conf/iou/min_bbox in `models.detection_params`):
+  A=combined only; B=v4 truck + yolo11n sack+box; C=A + yolo11n box-only (default);
+  D=v4 truck + best sack-only + yolo11n box-only. New modes need no code change.
   All modes load `.engine` files (2-3 coexist, ~24 MB peak); never mix load
   order assumptions — PyTorch `.pt` must load before TensorRT `.engine`.
 - `src/` = shared library (detection/tracking/counting/batch). `python -m src.main`
@@ -25,10 +27,14 @@ Indonesian (`karung`=sack, `truk`=truck); YOLO class names are English
 
 ## Gotchas
 
-- **Two env-key dialects**: production `.env` uses `RTSP_URL`, `DB_PATH`, `MODEL_PATH`,
-  … (`predict.py`, `counter_dashboard.py`); `src/config.py` reads different keys
-  (`LOCAL_RTSP`, `MODEL_SACK_PATH`, `MODEL_TRUCK_PATH`, …). Check which loader your
-  entry point uses before adding config.
+- **Unified config**: `config.yaml` is canonical (stream/models/counting/batch/
+  output/camera via `src/config_loader.py`); `.env` holds secrets + deployment
+  only (`RTSP_URL`, dashboard host/ports/secret/site); `zones.json` holds
+  geometry (polygons + left/right limits — knob keys there are ignored, warned);
+  `cfg/tracker.yaml` holds tracker hyperparams. `src/config.py` (v3 keys like
+  `LOCAL_RTSP`) is deprecated — don't add keys there. Dashboard mode switches
+  write `config.yaml` `models.active_mode` (atomic, manual restart to apply);
+  `batch_mode.json` keeps only manual/auto batch mode.
 - **Counting filters by class name, not ID**: `SackDetector`/`BoxDetector`/
   `TruckDetector` filter via `BaseDetector(class_filter)` (`src/detection.py`);
   tracker keeps `("sack", "truck", "box")` (`src/tracking.py`); counting uses

config.yaml

@@ -0,0 +1,133 @@
+# config.yaml — canonical configuration for the production pipeline
+# (predict.py + counter_dashboard.py). Single source of truth.
+#
+# Secrets & deployment-specific values (RTSP_URL, dashboard host/ports,
+# SECRET_KEY, SITE_NAME) stay in .env — they are NOT duplicated here.
+# Zone polygons stay in zones.json. Tracker hyperparams stay in cfg/tracker.yaml.
+#
+# Model modes: add new entries under models.modes (E, F, ...) without code changes.
+# A mode only selects engines + class filters; per-class conf/iou/min_bbox_area
+# live in models.detection_params and apply to ALL modes.
+
+stream:
+  resolution: [1280, 720]
+  inference_stride: 2
+
+models:
+  active_mode: "C"   # A|B|C|D (+ future E, F, ...). Dashboard switch writes here (manual restart to apply).
+
+  # Engine file paths (relative to repo root). Mode `engines` entries reference these keys.
+  paths:
+    combined: "models/v4-best.engine"
+    truck_only: "models/v4-best.engine"
+    sack_only: "models/best.engine"
+    box: "models/yolo11n-bbox-100ep-sack+box-20260909-best.engine"
+    truck_detector: "models/truck-detector.engine"
+
+  # Mode presets: engines to load + class filters. NOTHING else per mode.
+  # Each engine declares the classes it contributes, so tracker roles
+  # (shared vs dedicated, separate truck model) derive structurally —
+  # new modes need zero Python changes.
+  modes:
+    A:
+      description: "Combined v4 sack+truck only (legacy, no box counting)"
+      engines:
+        - path: combined
+          classes: [truck, sack]
+      class_filters:
+        truck: ["truck"]
+        sack: ["sack"]
+        box: []
+    B:
+      description: "v4 truck-only + yolo11n sack+box (shared tracker, shared ID space)"
+      engines:
+        - path: truck_only
+          classes: [truck]
+        - path: box
+          classes: [sack, box]
+      class_filters:
+        truck: ["truck"]
+        sack: ["sack"]
+        box: ["box"]
+    C:
+      description: "Combined v4 sack+truck + yolo11n box-only (dedicated tracker)"
+      engines:
+        - path: combined
+          classes: [truck, sack]
+        - path: box
+          classes: [box]
+      class_filters:
+        truck: ["truck"]
+        sack: ["sack"]
+        box: ["box"]
+    D:
+      description: "v4 truck-only + best sack-only + yolo11n box-only (3 engines)"
+      engines:
+        - path: truck_only
+          classes: [truck]
+        - path: sack_only
+          classes: [sack]
+        - path: box
+          classes: [box]
+      class_filters:
+        truck: ["truck"]
+        sack: ["sack"]
+        box: ["box"]
+
+  # Shared per-class detection params (ALL modes). iou default = Ultralytics default.
+  # min_bbox_area is a permissive guardrail in px^2 @1280x720 (existing
+  # perspective min_valid_area filter still applies on top).
+  detection_params:
+    truck:
+      conf: 0.35
+      iou: 0.7
+      min_bbox_area: 5000
+    sack:
+      conf: 0.35
+      iou: 0.7
+      min_bbox_area: 1500
+    box:
+      conf: 0.35
+      iou: 0.7
+      min_bbox_area: 1500
+
+# Polygons live in zones.json (kept separate: site calibration cadence).
+zones:
+  config_file: "zones.json"
+
+# Tracker hyperparams live in cfg/tracker.yaml (kept separate: ML tuning cadence).
+tracker:
+  config_file: "cfg/tracker.yaml"
+
+counting:
+  confirm_delay_sec: 0.5
+  exit_confirm_delay_sec: 6.0
+  entry_overlap_threshold: 0.20
+  exit_overlap_threshold: 0.05
+  camera_noise_deadband: 50
+  # Effective values below match zones.json (which previously overrode predict.py defaults).
+  duplicate_circle_radius: 30
+  min_valid_area: 15000
+  max_reid_transit_distance: 400
+  circle_stay_timeout_sec: 10.0
+  jarak_toleransi_duplikat: 30
+  tolerance_missing_frames: 1200
+
+batch:
+  timeout_seconds: 30.0
+  merge_threshold_seconds: 300
+  daily_cutoff_time: "06:00"
+
+output:
+  dir: "/opt/jetson-counter"
+  db_name: "jetson_counter.db"
+  state_file: "current_batch.json"
+  # batch_mode_file is DEPRECATED: dashboard now persists the mode switch to
+  # models.active_mode in this file. Key kept so old deployments can be detected.
+  batch_mode_file: "batch_mode.json"
+  live_frame_path: "/dev/shm/jetson-counter/live_frame.jpg"
+  live_status_path: "/dev/shm/jetson-counter/live_status.json"
+
+camera:
+  name: "CC1"
+  object_label: "karung-pakan"

counter_dashboard.py

@@ -21,9 +21,26 @@ from werkzeug.serving import WSGIRequestHandler
 from dotenv import load_dotenv
 load_dotenv()
 
+# Unified config (config.yaml canonical; .env supplies secrets/deployment-only
+# values: RTSP_URL, dashboard host/ports/secret/site). Env vars still override
+# file paths when explicitly set (backward compatible with old deployments).
+from src.config_loader import load_config, set_active_mode
+
+_CONFIG_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "config.yaml")
+CFG = load_config(_CONFIG_PATH)
+
+
+def _env_or(cfg_value, *env_keys):
+    """Backward compat: explicit env var wins over config.yaml for file paths."""
+    for k in env_keys:
+        v = os.getenv(k)
+        if v:
+            return v
+    return cfg_value
+
 
 app = Flask(__name__, template_folder="templates")
-app.config["SECRET_KEY"] = os.getenv("SECRET_KEY", "change-me-in-production")
+app.config["SECRET_KEY"] = CFG.dashboard.secret_key
 
 if os.name == "nt":
     _DEFAULT_DIR = "d:/Belajar/menghitung karung"
@@ -31,20 +48,25 @@ if os.name == "nt":
     CURRENT_BATCH_PATH = f"{_DEFAULT_DIR}/current_batch.json"
     BATCH_MODE_PATH = f"{_DEFAULT_DIR}/batch_mode.json"
     LIVE_STREAM_FRAME_PATH = f"{_DEFAULT_DIR}/live_frame.jpg"
+    LIVE_STATUS_FILE = f"{_DEFAULT_DIR}/live_status.json"
 else:
-    _DEFAULT_DIR = "/opt/jetson-counter"
-    DB_PATH = os.getenv("DB_PATH", f"{_DEFAULT_DIR}/jetson_counter.db")
-    CURRENT_BATCH_PATH = os.getenv("STATE_FILE", os.getenv("CURRENT_BATCH_PATH", f"{_DEFAULT_DIR}/current_batch.json"))
-    BATCH_MODE_PATH = os.getenv("BATCH_MODE_FILE", f"{_DEFAULT_DIR}/batch_mode.json")
-    LIVE_STREAM_FRAME_PATH = os.getenv("LIVE_STREAM_FRAME_PATH", "/dev/shm/jetson-counter/live_frame.jpg")
+    _DEFAULT_DIR = CFG.output.dir
+    DB_PATH = _env_or(os.path.join(_DEFAULT_DIR, CFG.output.db_name), "DB_PATH")
+    CURRENT_BATCH_PATH = _env_or(
+        os.path.join(_DEFAULT_DIR, CFG.output.state_file),
+        "STATE_FILE", "CURRENT_BATCH_PATH")
+    BATCH_MODE_PATH = _env_or(
+        os.path.join(_DEFAULT_DIR, CFG.output.batch_mode_file), "BATCH_MODE_FILE")
+    LIVE_STREAM_FRAME_PATH = _env_or(CFG.output.live_frame_path, "LIVE_STREAM_FRAME_PATH")
+    LIVE_STATUS_FILE = _env_or(CFG.output.live_status_path, "LIVE_STATUS_FILE")
 
-CUTOFF_TIME = os.getenv("CUTOFF_TIME", os.getenv("DAILY_CUTOFF_TIME", "20:00"))
+CUTOFF_TIME = _env_or(CFG.batch.daily_cutoff_time, "CUTOFF_TIME", "DAILY_CUTOFF_TIME")
 
-SITE_NAME = os.getenv("SITE_NAME", "LIVE")
+SITE_NAME = CFG.dashboard.site_name
 
-DASHBOARD_PORT = int(os.getenv("DASHBOARD_PORT", "5000"))
-DASHBOARD_HOST = os.getenv("DASHBOARD_HOST", "0.0.0.0")
-FLASK_DEBUG = os.getenv("FLASK_DEBUG", "false").lower() == "true"
+DASHBOARD_PORT = CFG.dashboard.port
+DASHBOARD_HOST = CFG.dashboard.host
+FLASK_DEBUG = CFG.dashboard.debug
 
 @app.route("/api/live-video")
 def api_live_video():
@@ -160,9 +182,9 @@ def get_counting_date(dt=None, cutoff_str=CUTOFF_TIME):
     return dt.date().isoformat()
 
 
-CAMERA_NAME = os.getenv("CAMERA_NAME", "CC1")
-OBJECT_LABEL = os.getenv("OBJECT_LABEL", "karung-pakan")
-OFFICE_PORT = int(os.getenv("OFFICE_PORT", "5721"))
+CAMERA_NAME = _env_or(CFG.camera.name, "CAMERA_NAME")
+OBJECT_LABEL = _env_or(CFG.camera.object_label, "OBJECT_LABEL")
+OFFICE_PORT = CFG.dashboard.office_port
 
 def is_office_request():
     """Check if request comes from office port."""
@@ -359,25 +381,35 @@ def api_batch_stop():
         return jsonify({"success": False, "error": str(e)}), 500
 
 
-MODEL_MODE_CHOICES = ("A", "B", "C", "D")
+# Model modes are DATA in config.yaml models.modes — derived here so future
+# modes (E, F, ...) appear automatically. batch_mode.json keeps ONLY the
+# manual/auto batch "mode"; model_mode lives in config.yaml now.
+MODEL_MODE_CHOICES = tuple(CFG.models.modes.keys())
 MODEL_MODE_DESCRIPTIONS = {
-    "A": "Combined v4 sack+truck only (legacy)",
-    "B": "v4 truck-only + yolo11n sack+box",
-    "C": "Combined v4 sack+truck + yolo11n box-only (default production)",
-    "D": "v4 truck + best.pt sack-only + yolo11n box-only",
+    m: preset.description for m, preset in CFG.models.modes.items()
 }
 
 
+def _active_model_mode() -> str:
+    """Canonical model mode: config.yaml models.active_mode (validated)."""
+    try:
+        return load_config(_CONFIG_PATH).models.active_mode
+    except Exception:
+        return CFG.models.active_mode
+
+
 def _read_batch_mode_file():
-    data = {"mode": "manual", "model_mode": "C"}
+    data = {"mode": "manual", "model_mode": _active_model_mode()}
     if os.path.exists(BATCH_MODE_PATH):
         try:
             with open(BATCH_MODE_PATH, "r") as f:
                 stored = json.load(f)
             data["mode"] = stored.get("mode", "manual")
-            data["model_mode"] = (stored.get("model_mode") or "C").upper()
-            if data["model_mode"] not in MODEL_MODE_CHOICES:
-                data["model_mode"] = "C"
+            # Legacy model_mode in batch_mode.json is IGNORED (config.yaml wins).
+            legacy = (stored.get("model_mode") or "").upper()
+            if legacy and legacy != data["model_mode"]:
+                print(f"[WARN] batch_mode.json model_mode={legacy!r} diabaikan — "
+                      f"config.yaml active_mode={data['model_mode']!r} yang berlaku.")
         except Exception:
             pass
     return data
@@ -395,20 +427,32 @@ def api_batch_mode():
                     return jsonify({"success": False, "error": "Invalid mode. Use 'auto' or 'manual'"}), 400
                 stored["mode"] = mode
             if "model_mode" in req_data:
-                mmode = str(req_data.get("model_mode", "C")).upper()
-                if mmode not in MODEL_MODE_CHOICES:
-                    return jsonify({"success": False, "error": "Invalid model_mode. Use A/B/C/D"}), 400
+                mmode = str(req_data.get("model_mode", "")).upper()
+                try:
+                    # Validates against config.yaml models.modes, then persists
+                    # atomically (tmp+replace, comments preserved). Manual
+                    # `karung-counter` restart still required to apply.
+                    set_active_mode(_CONFIG_PATH, mmode)
+                except (ValueError, RuntimeError) as ve:
+                    return jsonify({
+                        "success": False,
+                        "error": f"Invalid model_mode. Use one of {sorted(MODEL_MODE_CHOICES)}: {ve}",
+                    }), 400
                 stored["model_mode"] = mmode
             stored["updated_at"] = datetime.now().isoformat()
 
-            os.makedirs(os.path.dirname(BATCH_MODE_PATH), exist_ok=True)
-            with open(BATCH_MODE_PATH, "w", encoding="utf-8") as f:
-                json.dump(stored, f, indent=2)
+            # batch_mode.json keeps ONLY the manual/auto batch mode now.
+            batch_state = {"mode": stored["mode"], "updated_at": stored["updated_at"]}
+            os.makedirs(os.path.dirname(BATCH_MODE_PATH) or ".", exist_ok=True)
+            tmp_path = BATCH_MODE_PATH + ".tmp"
+            with open(tmp_path, "w", encoding="utf-8") as f:
+                json.dump(batch_state, f, indent=2)
+            os.replace(tmp_path, BATCH_MODE_PATH)
 
             return jsonify({"success": True, "mode": stored["mode"],
                             "model_mode": stored["model_mode"],
                             "message": f"Batch mode={stored['mode']}, model_mode={stored['model_mode']} "
-                                       f"(model_mode applies on next service restart)"})
+                                        f"(model_mode applies on next karung-counter restart)"})
         except Exception as e:
             return jsonify({"success": False, "error": str(e)}), 500
 
@@ -432,7 +476,7 @@ def api_model_modes():
 @app.route("/api/current-batch")
 def api_current_batch():
     fps_val = 0.0
-    status_file = os.getenv('LIVE_STATUS_FILE', '/dev/shm/jetson-counter/live_status.json' if os.name != 'nt' else 'd:/Belajar/menghitung karung/live_status.json')
+    status_file = LIVE_STATUS_FILE
     try:
         if os.path.exists(status_file):
             with open(status_file, "r") as sf:

deploy_to_jetson.py

@@ -17,6 +17,7 @@ files_to_sync = [
     ("templates/base.html", "/home/jetson/karung/templates/base.html"),
     ("counter_dashboard.py", "/home/jetson/karung/counter_dashboard.py"),
     ("predict.py", "/home/jetson/karung/predict.py"),
+    ("config.yaml", "/home/jetson/karung/config.yaml"),
     (".env", "/home/jetson/karung/.env"),
 ] + [
     (f"models/{f}", f"/home/jetson/karung/models/{f}") for f in MODEL_ENGINES

docs/configuration.md

@@ -1,9 +1,36 @@
 # Configuration
 
+Canonical source: **`config.yaml`** (repo root) — stream, models, counting knobs,
+batch, output paths, camera. Loaded once at startup via `src/config_loader.py`
+(stdlib dataclasses + pyyaml, no heavy deps). Secrets & deployment-only values
+stay in `.env`. Zone polygons stay in `zones.json`. Tracker hyperparams stay in
+`cfg/tracker.yaml`.
+
+```
+config.yaml          canonical: stream/models/counting/batch/output/camera
+.env                 secrets + deployment: RTSP_URL, dashboard host/ports/secret/site
+zones.json           geometry: palet/truck/counting polygons + left/right limits
+cfg/tracker.yaml     tracker hyperparams (FastTrack/ByteTrack tuning)
+```
+
+Model modes are **data** (`config.yaml` → `models.modes`): each preset declares
+only `engines` (path key + contributed classes) and `class_filters`.
+Per-class `conf`/`iou`/`min_bbox_area` live in `models.detection_params` and
+apply to ALL modes. Adding mode E/F/... is a YAML-only change — `predict.py`
+derives tracker roles structurally, and the dashboard `/api/model-modes`
+endpoint lists them automatically.
+
+Mode switch: dashboard `POST /api/batch/mode {"model_mode": "X"}` validates
+against `config.yaml` and persists atomically (tmp+replace, comments preserved)
+to `models.active_mode`. **Manual `karung-counter` restart still required**
+(models load once at startup). `batch_mode.json` keeps only the manual/auto
+batch `mode`; its legacy `model_mode` key is ignored (warned). `MODEL_MODE`
+env var still overrides for one run but is deprecated (warned).
+
 Three layers: environment file → zone polygons → tracker/counter tuning.
 Template: `.env.example`. Production values live in `.env` (git-ignored).
 
-## 1. `.env` (production keys — `predict.py` / `counter_dashboard.py`)
+## 1. `.env` (secrets & deployment — `predict.py` / `counter_dashboard.py`)
 
 | Key | Default | Meaning |
 |---|---|---|
@@ -19,9 +46,9 @@ Template: `.env.example`. Production values live in `.env` (git-ignored).
 | `DASHBOARD_HOST` / `DASHBOARD_PORT` | `0.0.0.0` / `5000` | Dashboard bind |
 | `OFFICE_PORT` | `5721` | Second dashboard port |
 | `FLASK_DEBUG` | `false` | Flask debug |
-| `RTSP_URL` | — | Camera stream URL |
-| `MODEL_PATH` | auto (`models/v4-best.engine` > `.pt` > `v4-best (1).pt` > `models/model_karung_truk.engine` > `.pt`) | Override combined-model weights (`predict.py`) |
-| `MODEL_MODE` | `C` | Model pipeline mode A/B/C/D (see `models/modelREADME.md`); also settable via `--model-mode` or dashboard (applies on restart) |
+| `RTSP_URL` | — | Camera stream URL (env-only, never in YAML) |
+| `MODEL_PATH` | — (deprecated) | Single-file v4 override, folded into `models.paths` |
+| `MODEL_MODE` | — (deprecated) | One-run override of `models.active_mode` (warned) |
 | `BATCH_MERGE_THRESHOLD_SECONDS` | `300` | Merge window for adjacent batches |
 
 On Windows dev machines these resolve to `d:/Belajar/menghitung karung/...`.

models/modelREADME.md

@@ -18,7 +18,13 @@ Class names below are read directly from each checkpoint (`YOLO(path).names`).
 | `best.{pt,onnx,engine}` | `{0: sack}` | seg | Sack-only specialist (Mode D sack) |
 | `karung-dimuat-detection-di-feedmill-yolo26n-seg-200e.{pt,onnx,engine}` | `{0: person, 1: sack}` | **seg** | Person-exclusion seg model (legacy `predict_new.py`) |
 
-## Model Modes (`predict.py --model-mode`, default **C**)
+## Model Modes (`predict.py --model-mode` / `config.yaml models.active_mode`, default **C**)
+
+Modes are DATA in `config.yaml` `models.modes` — each preset declares engines
+(path key + contributed classes) and class filters only. Per-class
+conf/iou/min_bbox live in `models.detection_params` (shared across modes).
+Add E/F/... in YAML with no code change; `predict.py` derives tracker roles
+structurally and the dashboard lists new modes automatically.
 
 Counting filters by **class name, not ID** (`BaseDetector(class_filter)` in
 `src/detection.py`; tracker allow-list in `src/tracking.py`; dual counters in

predict.py

@@ -24,6 +24,14 @@ from src.truck_roi import TruckROITracker
 from src.counting import LineCrossCounter, MultiClassLineCounter
 from src.batch import BatchLifecycleManager, BatchRecord
 from src.dashboard import DashboardOverlay
+from src.config_loader import (
+    Config,
+    check_legacy_batch_mode,
+    load_config,
+    read_zone_polygons,
+    resolve_active_mode,
+    LEGACY_ZONE_KNOBS,
+)
 
 # --- SQLite Database & State Configuration ---
 if os.name == 'nt':
@@ -75,10 +83,12 @@ def parse_args():
     p.add_argument("--box-model", type=str, default=None,
                    help="Override yolo11n sack+box model path (.pt/.engine)")
     p.add_argument("--model-mode", type=str, default=None,
-                   choices=["A", "B", "C", "D"],
-                    help="Model pipeline mode (default: MODEL_MODE env or C). "
-                        "A=combined only; B=v4 truck + yolo11n sack+box; "
-                        "C=A + yolo11n box-only; D=v4 truck + best sack + yolo11n box")
+                    help="Model pipeline mode (default: config.yaml models.active_mode). "
+                         "A=combined only; B=v4 truck + yolo11n sack+box; "
+                         "C=A + yolo11n box-only; D=v4 truck + best sack + yolo11n box. "
+                         "New modes from config.yaml need no code change.")
+    p.add_argument("--config", type=str, default=None,
+                    help="Path to config.yaml (default: config.yaml next to predict.py)")
     p.add_argument("--batch-timeout", type=float, default=None,
                    help="Override sack-idle + truck-gone timeouts (seconds)")
     p.add_argument("--max-frames", type=int, default=None,
@@ -443,15 +453,22 @@ all_counted_sacks_map = {}
 last_seen_near_person_frame = {}
 blocked_due_to_duplicate = {}
 
-# --- Model Paths ---
-# All weights live in models/ (see models/modelREADME.md for mode/filter matrix).
-# Model gabungan karung + truk terbaru (v4-best TensorRT / PyTorch)
+# --- Model paths & config ---
+# All weights live in models/ (see models/modelREADME.md for mode/filter matrix
+# and config.yaml models.* for the canonical paths + per-mode presets).
 _BASE_DIR = os.path.dirname(os.path.abspath(__file__))
 _MODELS_DIR = os.path.join(_BASE_DIR, "models")
-_env_model = os.getenv("MODEL_PATH")
-if _env_model and os.path.exists(_env_model):
-    COMBINED_MODEL_PATH = _env_model
-else:
+_DEFAULT_CONFIG_PATH = os.path.join(_BASE_DIR, "config.yaml")
+
+# Active Config object (set in __main__/run_prediction before the pipeline starts).
+_CFG: Config | None = None
+
+
+def _legacy_combined_fallback() -> str:
+    """Pre-YAML combined-model auto-pick (kept for --model/MODEL_PATH handling)."""
+    _env_model = os.getenv("MODEL_PATH")
+    if _env_model and os.path.exists(_env_model):
+        return _env_model
     _candidates = [
         os.path.join(_MODELS_DIR, "v4-best.engine"),
         os.path.join(_MODELS_DIR, "v4-best.pt"),
@@ -459,123 +476,134 @@ else:
         os.path.join(_MODELS_DIR, "model_karung_truk.engine"),
         os.path.join(_MODELS_DIR, "model_karung_truk.pt"),
     ]
-    COMBINED_MODEL_PATH = next((p for p in _candidates if os.path.exists(p)), os.path.join(_MODELS_DIR, "v4-best.pt"))
+    return next((p for p in _candidates if os.path.exists(p)),
+                os.path.join(_MODELS_DIR, "v4-best.pt"))
 
 
-def _pick_box_model(explicit: str | None) -> str:
-    """Resolve yolo11n sack+box weights: explicit > .engine > .pt."""
-    if explicit:
-        return explicit
-    for cand in (
-        os.path.join(_MODELS_DIR, "yolo11n-bbox-100ep-sack+box-20260909-best.engine"),
-        os.path.join(_MODELS_DIR, "yolo11n-bbox-100ep-sack+box-20260909-best.pt"),
-    ):
-        if os.path.exists(cand):
-            return cand
-    return os.path.join(_MODELS_DIR, "yolo11n-bbox-100ep-sack+box-20260909-best.engine")
+def apply_cli_overrides_to_config(cfg: Config, args) -> Config:
+    """Fold legacy CLI/env model overrides into the Config object (in place).
 
-
-def _pick_sack_only_model() -> str:
-    """Resolve best sack-only weights: .engine > .pt."""
-    for cand in (
-        os.path.join(_MODELS_DIR, "best.engine"),
-        os.path.join(_MODELS_DIR, "best.pt"),
-    ):
-        if os.path.exists(cand):
-            return cand
-    return os.path.join(_MODELS_DIR, "best.engine")
+    --model/--box-model/MODEL_PATH keep their pre-YAML meaning:
+    a single v4 file used wherever a v4 engine is needed.
+    --sack-conf/--truck-conf/--box-conf override detection_params conf.
+    """
+    v4_override = args.model or os.getenv("MODEL_PATH")
+    if v4_override:
+        if not os.path.exists(v4_override):
+            print(f"[WARN] --model/MODEL_PATH {v4_override} tidak ditemukan, "
+                  f"dipakai langsung (YOLO bisa resolve).")
+        for key in ("combined", "truck_only"):
+            if key in cfg.models.paths:
+                cfg.models.paths[key] = v4_override
+    if getattr(args, "box_model", None):
+        cfg.models.paths["box"] = args.box_model
+    for cls_name, val in (("sack", args.sack_conf), ("truck", args.truck_conf),
+                          ("box", args.box_conf)):
+        if val is not None and cls_name in cfg.models.detection_params:
+            cfg.models.detection_params[cls_name].conf = float(val)
+    if getattr(args, "batch_timeout", None) is not None:
+        cfg.batch.timeout_seconds = float(args.batch_timeout)
+    return cfg
 
 
 # --- Model pipeline modes -------------------------------------------------
-# A: combined v4 sack+truck only (legacy production).
-# B: v4 truck-only + yolo11n sack+box.
-# C: A + yolo11n box-only (default production).
-# D: v4 truck + best.pt sack-only + yolo11n box-only.
+# Modes are DATA in config.yaml models.modes (engines + class filters only).
+# A=combined only; B=v4 truck + yolo11n sack+box; C=A + yolo11n box-only (default);
+# D=v4 truck + best sack-only + yolo11n box-only. New modes need no code change.
 # All engines verified to coexist (~16-24 MB peak of 7.6 GB).
-MODEL_MODES = ("A", "B", "C", "D")
 
 
 def resolve_model_mode(explicit: str | None) -> str:
-    """Precedence: --model-mode > MODEL_MODE env > batch_mode.json model_mode > C.
+    """Precedence: --model-mode > MODEL_MODE env (deprecated) > config.yaml.
 
-    Dashboard-driven switches persist to batch_mode.json and take effect
+    Dashboard switches persist to config.yaml models.active_mode and take effect
     on next service restart (models are loaded once at startup).
+    batch_mode.json is legacy and no longer consulted.
     """
-    mode = (explicit or os.getenv("MODEL_MODE") or "").upper()
-    if not mode:
-        try:
-            _bm = os.path.join(
-                os.getenv("OUTPUT_DIR", "/opt/jetson-counter") if os.name != "nt"
-                else "d:/Belajar/menghitung karung",
-                "batch_mode.json",
-            )
-            _bm = os.getenv("BATCH_MODE_FILE", _bm)
-            with open(_bm, "r", encoding="utf-8") as f:
-                mode = (json.load(f).get("model_mode") or "").upper()
-        except Exception:
-            pass
-    mode = mode or "C"
-    if mode not in MODEL_MODES:
-        print(f"[WARN] MODEL_MODE '{mode}' tidak dikenal, pakai C.")
-        return "C"
-    return mode
+    cfg = _CFG or load_config(_DEFAULT_CONFIG_PATH)
+    return resolve_active_mode(explicit, cfg)
 
 
-def build_model_pipeline(mode, combined_path, box_model_path, sack_only_path,
-                         sack_conf, truck_conf, box_conf, device):
-    """Instantiate YOLO handles + detectors/trackers for a model mode.
+def _bbox_area_ok(d, min_area: float) -> bool:
+    """Permissive per-class guardrail from config.yaml detection_params."""
+    x1, y1, x2, y2 = d.bbox
+    return (x2 - x1) * (y2 - y1) >= min_area
 
-    Returns dict with keys: mode, truck_model, sack_model, box_model,
-    truck_detector, tracker, box_tracker (None when sack tracker covers boxes).
+
+def build_model_pipeline(mode, cfg, device, base_dir=None):
+    """Instantiate YOLO handles + detectors/trackers for a mode preset.
+
+    Roles derive structurally from the preset's per-engine `classes`
+    (config.yaml models.modes) — no per-mode-letter branching, so new
+    modes (E, F, ...) work without code changes:
+    - tracker (primary): first engine covering `sack`
+    - truck_detector: first engine covering `truck`
+      (separate_truck_model = truck engine is not the primary)
+    - box_tracker: dedicated iff a non-primary engine covers `box`
+      (shared with the primary tracker otherwise, e.g. mode B)
+
+    Returns dict with keys: mode, truck_detector, tracker, box_tracker
+    (None when sack tracker covers boxes or no box engine), separate_truck_model,
+    class_filters, min_areas.
     """
-    print(f"[INFO] Model mode: {mode}")
+    base_dir = base_dir or _BASE_DIR
+    preset = cfg.models.modes[mode]
+    filters = {k: tuple(v) for k, v in preset.class_filters.items()}
+    dp_truck = cfg.detection_params_for("truck")
+    dp_sack = cfg.detection_params_for("sack")
+    dp_box = cfg.detection_params_for("box")
+    print(f"[INFO] Model mode: {mode} ({preset.description})")
     dummy = np.zeros((720, 1280, 3), dtype=np.uint8)
 
-    def _load(path, label):
-        print(f"[INFO] Memuat {label}: {path}")
-        m = YOLO(path)
-        _ = m(dummy, imgsz=640, device=device, verbose=False)  # warm-up CUDA/TRT ctx
-        return m
+    handles: dict[str, object] = {}
+
+    def _load(path_key: str):
+        if path_key not in handles:
+            path = cfg.engine_path(path_key, base_dir)
+            print(f"[INFO] Memuat {path_key}: {path}")
+            m = YOLO(path)
+            _ = m(dummy, imgsz=640, device=device, verbose=False)  # warm-up CUDA/TRT ctx
+            handles[path_key] = m
+        return handles[path_key]
+
+    primary_key = next((e.path for e in preset.engines if "sack" in e.classes), None)
+    truck_key = next((e.path for e in preset.engines if "truck" in e.classes), None)
+    box_keys = [e.path for e in preset.engines if "box" in e.classes]
+
+    if primary_key is None:
+        raise ValueError(
+            f"Mode {mode!r} has no engine covering class 'sack' — "
+            f"at least one engines[].classes must include sack."
+        )
+
+    tracker = ByteTrackTracker(_load(primary_key), dp_sack.conf, dp_sack.iou)
+    truck_detector = (
+        TruckDetector(_load(truck_key), dp_truck.conf,
+                      class_filter=filters.get("truck") or ("truck",),
+                      iou=dp_truck.iou)
+        if truck_key is not None else None
+    )
+    separate_truck_model = truck_key is not None and truck_key != primary_key
+
+    box_tracker = None
+    if box_keys and filters.get("box"):
+        dedicated = next((k for k in box_keys if k != primary_key), None)
+        if dedicated is not None:
+            box_tracker = ByteTrackTracker(_load(dedicated), dp_box.conf, dp_box.iou)
+        # else: boxes share the primary tracker (single ID space, e.g. mode B)
 
-    if mode == "A":
-        shared = _load(combined_path, "model gabungan sack+truck")
-        return {
-            "mode": mode,
-            "truck_detector": TruckDetector(shared, truck_conf),
-            "tracker": ByteTrackTracker(shared, sack_conf),
-            "box_tracker": None,
-            "separate_truck_model": False,
-        }
-    if mode == "B":
-        v4 = _load(combined_path, "model v4 (truck-only)")
-        yb = _load(box_model_path, "model yolo11n (sack+box)")
-        return {
-            "mode": mode,
-            "truck_detector": TruckDetector(v4, truck_conf, class_filter=("truck",)),
-            "tracker": ByteTrackTracker(yb, sack_conf),
-            "box_tracker": None,  # boxes share the yolo11n tracker
-            "separate_truck_model": True,
-        }
-    if mode == "C":
-        shared = _load(combined_path, "model gabungan sack+truck")
-        yb = _load(box_model_path, "model yolo11n (box-only)")
-        return {
-            "mode": mode,
-            "truck_detector": TruckDetector(shared, truck_conf),
-            "tracker": ByteTrackTracker(shared, sack_conf),
-            "box_tracker": ByteTrackTracker(yb, box_conf),
-            "separate_truck_model": False,
-        }
-    # mode == "D"
-    v4 = _load(combined_path, "model v4 (truck-only)")
-    sb = _load(sack_only_path, "model best (sack-only)")
-    yb = _load(box_model_path, "model yolo11n (box-only)")
     return {
         "mode": mode,
-        "truck_detector": TruckDetector(v4, truck_conf, class_filter=("truck",)),
-        "tracker": ByteTrackTracker(sb, sack_conf),
-        "box_tracker": ByteTrackTracker(yb, box_conf),
-        "separate_truck_model": True,
+        "truck_detector": truck_detector,
+        "tracker": tracker,
+        "box_tracker": box_tracker,
+        "separate_truck_model": separate_truck_model,
+        "class_filters": filters,
+        "min_areas": {
+            "truck": dp_truck.min_bbox_area,
+            "sack": dp_sack.min_bbox_area,
+            "box": dp_box.min_bbox_area,
+        },
     }
 
 # =====================================================================
@@ -646,6 +674,13 @@ def load_zones():
                 EXTERNAL_STREAM_URL_REF = data.get('external_stream_url', 'http://192.168.192.96:8888/cam/')
                 if not EXTERNAL_STREAM_URL_REF:
                     EXTERNAL_STREAM_URL_REF = 'http://192.168.192.96:8888/cam/'
+                # Knob keys below are LEGACY: config.yaml counting.* is canonical and
+                # run_prediction overrides these globals from it. Kept reading here
+                # only so import-time defaults stay sane; warn once to guide migration.
+                _legacy_knobs = sorted(LEGACY_ZONE_KNOBS & set(data.keys()))
+                if _legacy_knobs:
+                    print(f"[WARN] zones.json knob keys {_legacy_knobs} diabaikan — "
+                          f"pakai config.yaml (counting.*) sebagai gantinya.")
                 print("[INFO] Berhasil memuat koordinat zona dan parameter kalibrasi dari zones.json")
                 return
         except Exception as e:
@@ -1200,9 +1235,10 @@ def _filter_sacks_in_roi(detections, roi):
 
 def run_prediction(model_path, source_path,
                     output_json_path="hasil_perhitungan.json", max_frames=None,
-                    inference_stride=2, sack_conf=0.35, truck_conf=0.35,
-                    box_conf=0.35, box_model_path=None, model_mode=None,
-                    output_dir=None, batch_timeout=None):
+                    inference_stride=None, sack_conf=None, truck_conf=None,
+                    box_conf=None, box_model_path=None, model_mode=None,
+                    output_dir=None, batch_timeout=None,
+                    config_path=None, cfg=None):
     global prev_active_track_ids, lost_tracks, metrics, track_positions, counted_at_frame
     global track_confirmed_state, already_counted, is_locked, has_crossed_line, exit_crossed_line, track_areas
     global pending_enter_since, pending_exit_since, track_started_in_truck, outside_truck_frames
@@ -1212,25 +1248,65 @@ def run_prediction(model_path, source_path,
     global active_batch_info, system_state
     global DUPLICATE_CIRCLE_RADIUS, MIN_VALID_AREA, JARAK_TOLERANSI_DUPLIKAT, MAX_REID_TRANSIT_DISTANCE
     global DB_PATH, STATE_FILE, BATCH_MODE_FILE, LIVE_STREAM_FRAME_PATH
+    global CAMERA_NAME, OBJECT_LABEL, DAILY_CUTOFF_TIME, BATCH_MERGE_THRESHOLD_SECONDS
+    global CIRCLE_STAY_TIMEOUT_SEC, _CFG
 
-    mode = resolve_model_mode(model_mode)
-    if box_model_path is None:
-        box_model_path = _pick_box_model(None)
+    # --- Unified config (config.yaml canonical, .env for secrets/deployment) ---
+    global _CFG
+    _CFG = cfg or load_config(config_path or _DEFAULT_CONFIG_PATH)
+    cfg = _CFG
 
+    # Legacy path overrides keep their pre-YAML meaning (folded into cfg).
+    if model_path and model_path != _legacy_combined_fallback():
+        for _k in ("combined", "truck_only"):
+            if _k in cfg.models.paths:
+                cfg.models.paths[_k] = model_path
+    if box_model_path:
+        cfg.models.paths["box"] = box_model_path
+    for _cls, _v in (("sack", sack_conf), ("truck", truck_conf), ("box", box_conf)):
+        if _v is not None and _cls in cfg.models.detection_params:
+            cfg.models.detection_params[_cls].conf = float(_v)
+
+    # --- Path / identity / timing globals from config (output_dir wins) ---
     if output_dir:
-        # Cross-platform: plain join, keep .env layout when output_dir is None
+        # Cross-platform: plain join, keep config layout when output_dir is None
         DB_PATH = os.path.join(output_dir, "jetson_counter.db")
         STATE_FILE = os.path.join(output_dir, "current_batch.json")
         BATCH_MODE_FILE = os.path.join(output_dir, "batch_mode.json")
         LIVE_STREAM_FRAME_PATH = os.path.join(output_dir, "live_frame.jpg")
         print(f"[INFO] Output dir override: {output_dir}")
+    else:
+        DB_PATH = os.path.join(cfg.output.dir, cfg.output.db_name)
+        STATE_FILE = os.path.join(cfg.output.dir, cfg.output.state_file)
+        BATCH_MODE_FILE = os.path.join(cfg.output.dir, cfg.output.batch_mode_file)
+        LIVE_STREAM_FRAME_PATH = cfg.output.live_frame_path
+    CAMERA_NAME = cfg.camera.name
+    OBJECT_LABEL = cfg.camera.object_label
+    DAILY_CUTOFF_TIME = cfg.batch.daily_cutoff_time
+    BATCH_MERGE_THRESHOLD_SECONDS = cfg.batch.merge_threshold_seconds
+
+    # --- Counting-knob globals from config (values match legacy zones.json) ---
+    CONFIRM_DELAY_SEC = cfg.counting.confirm_delay_sec
+    EXIT_CONFIRM_DELAY_SEC = cfg.counting.exit_confirm_delay_sec
+    DUPLICATE_CIRCLE_RADIUS = cfg.counting.duplicate_circle_radius
+    MIN_VALID_AREA = cfg.counting.min_valid_area
+    JARAK_TOLERANSI_DUPLIKAT = cfg.counting.jarak_toleransi_duplikat
+    MAX_REID_TRANSIT_DISTANCE = cfg.counting.max_reid_transit_distance
+    CIRCLE_STAY_TIMEOUT_SEC = cfg.counting.circle_stay_timeout_sec
+
+    # Legacy batch_mode.json no longer drives the mode — nudge once if stale.
+    _legacy_warn = check_legacy_batch_mode(cfg, BATCH_MODE_FILE)
+    if _legacy_warn:
+        print(f"[WARN] {_legacy_warn}")
+
+    mode = resolve_model_mode(model_mode)
 
     # 1. Silencing YOLO logs
     from ultralytics.utils import LOGGER
     import logging
     LOGGER.setLevel(logging.WARNING)
 
-    INFERENCE_STRIDE = inference_stride
+    INFERENCE_STRIDE = inference_stride if inference_stride is not None else cfg.stream.inference_stride
 
     saver = None
     saver_thread = None
@@ -1265,17 +1341,15 @@ def run_prediction(model_path, source_path,
 
     print(f"[INFO] Resolusi Asli: {int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))}x{int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))} @ {fps:.1f} FPS (Diresize ke 1280x720 untuk koordinat tetap)")
 
-    # Initialize components per model mode (engines verified to coexist)
-    pipe = build_model_pipeline(
-        mode, model_path, box_model_path, _pick_sack_only_model(),
-        sack_conf, truck_conf, box_conf, device,
-    )
+    # Initialize components per mode preset (engines verified to coexist)
+    pipe = build_model_pipeline(mode, cfg, device, _BASE_DIR)
     print("[INFO] Warm-up model selesai.")
 
     truck_detector = pipe["truck_detector"]
     tracker = pipe["tracker"]
     box_tracker = pipe["box_tracker"]
     separate_truck_model = pipe["separate_truck_model"]
+    min_areas = pipe["min_areas"]
     stabilizer = BboxStabilizer(
         ema_alpha=0.35,
         max_hold_frames=10,
@@ -1334,10 +1408,10 @@ def run_prediction(model_path, source_path,
         confidence=1.0
     )
 
-    DUPLICATE_CIRCLE_RADIUS = DUPLICATE_CIRCLE_RADIUS_REF
-    MIN_VALID_AREA = MIN_VALID_AREA_REF
-    JARAK_TOLERANSI_DUPLIKAT = JARAK_TOLERANSI_DUPLIKAT_REF
-    MAX_REID_TRANSIT_DISTANCE = MAX_REID_TRANSIT_DISTANCE_REF
+    # NOTE: counting-knob globals (DUPLICATE_CIRCLE_RADIUS, MIN_VALID_AREA,
+    # JARAK_TOLERANSI_DUPLIKAT, MAX_REID_TRANSIT_DISTANCE, ...) were set from
+    # config.yaml at the top of run_prediction — intentionally NOT reset to
+    # zones.json REFs here (config.yaml is canonical now).
 
     counter = MultiClassLineCounter(
         line_y=static_line_y,
@@ -1346,7 +1420,7 @@ def run_prediction(model_path, source_path,
         margin=20,
         dedup_radius=float(DUPLICATE_CIRCLE_RADIUS),
     )
-    _idle_timeout = batch_timeout if batch_timeout is not None else 30.0
+    _idle_timeout = batch_timeout if batch_timeout is not None else cfg.batch.timeout_seconds
     batch_mgr = BatchLifecycleManager(
         stabilize_seconds=0.0,         # Start batch instantly when triggered by crossing
         stabilize_threshold_px=9999.0,  # Disable displacement threshold check
@@ -1426,11 +1500,19 @@ def run_prediction(model_path, source_path,
             # Filter sack (+box, modes B-D) detections (confidence >= 0.50).
             # Counter splits by class_name downstream; MultiClassLineCounter
             # ignores anything that is not sack/box.
-            raw_tracked_sacks = [d for d in raw_tracked_all if d.class_name in ("sack", "box") and d.confidence >= 0.50]
+            # min_bbox_area guardrails come from config.yaml detection_params.
+            raw_tracked_sacks = [
+                d for d in raw_tracked_all
+                if d.class_name in ("sack", "box") and d.confidence >= 0.50
+                and _bbox_area_ok(d, min_areas.get(d.class_name, 0))
+            ]
 
             # Truck candidates: from shared tracker (modes A/C) and/or the
             # separate v4 truck model (modes B/D, every 5th frame, cached).
-            truck_candidates = [d for d in raw_tracked_all if d.class_name == "truck"]
+            truck_candidates = [
+                d for d in raw_tracked_all
+                if d.class_name == "truck" and _bbox_area_ok(d, min_areas.get("truck", 0))
+            ]
             if separate_truck_model:
                 if frame_idx % 5 == 0 or 'last_detected_trucks' not in locals():
                     last_detected_trucks = truck_detector.detect(frame)
@@ -1465,7 +1547,9 @@ def run_prediction(model_path, source_path,
                     last_raw_tracked_boxes = raw_tracked_boxes
                 else:
                     raw_tracked_boxes = last_raw_tracked_boxes
-                raw_boxes = [d for d in raw_tracked_boxes if d.class_name == "box"]
+                raw_boxes = [d for d in raw_tracked_boxes
+                             if d.class_name == "box"
+                             and _bbox_area_ok(d, min_areas.get("box", 0))]
                 stable_boxes = box_stabilizer.update(raw_boxes)
                 stable_boxes = [
                     d for d in stable_boxes
@@ -1777,35 +1861,36 @@ if __name__ == "__main__":
     NO_DASHBOARD = args.no_dashboard
     NO_DB = args.no_db
 
-    MODEL_FILE = args.model or COMBINED_MODEL_PATH
-    if args.model and not os.path.exists(args.model):
-        print(f"[WARN] --model {args.model} tidak ditemukan, dipakai langsung (YOLO bisa resolve).")
+    # Unified config first (config.yaml canonical; .env supplies secrets/RTSP).
+    _CFG = load_config(args.config or _DEFAULT_CONFIG_PATH)
+    _CFG = apply_cli_overrides_to_config(_CFG, args)
 
     if args.source:
         SOURCE_INPUT = args.source
     else:
-        _env_url = os.getenv("RTSP_URL")
+        _env_url = _CFG.stream.rtsp_url
         SOURCE_INPUT = _env_url if _env_url else ("anomali.mp4" if (os.path.exists("anomali.mp4") and os.name == 'nt') else "rtsp://192.168.192.96:8554/cam")
     OUTPUT_JSON = args.output_json or "hasil_perhitungan.json"
 
-    print(f"[INFO] source={SOURCE_INPUT} model={MODEL_FILE} "
-          f"sack_conf={args.sack_conf or 0.35} truck_conf={args.truck_conf or 0.35} "
-           f"box_conf={args.box_conf or 0.35} model_mode={args.model_mode or os.getenv('MODEL_MODE', 'C')} "
+    print(f"[INFO] config={args.config or _DEFAULT_CONFIG_PATH} source={SOURCE_INPUT} "
+          f"model_mode={resolve_model_mode(args.model_mode)} "
           f"no_dashboard={NO_DASHBOARD} no_db={NO_DB}")
 
     try:
         run_prediction(
-            model_path=MODEL_FILE,
+            model_path=args.model,
             source_path=SOURCE_INPUT,
             output_json_path=OUTPUT_JSON,
             max_frames=args.max_frames,
-            sack_conf=args.sack_conf or 0.35,
-            truck_conf=args.truck_conf or 0.35,
-            box_conf=args.box_conf or 0.35,
+            sack_conf=args.sack_conf,
+            truck_conf=args.truck_conf,
+            box_conf=args.box_conf,
             box_model_path=args.box_model,
             model_mode=args.model_mode,
             output_dir=args.output_dir,
             batch_timeout=args.batch_timeout,
+            config_path=args.config or _DEFAULT_CONFIG_PATH,
+            cfg=_CFG,
         )
     except KeyboardInterrupt:
         print("\n" + "=" * 50)

requirements.txt

@@ -7,6 +7,7 @@ numpy
 shapely
 flask
 python-dotenv
+pyyaml  # also pulled by ultralytics; used directly by src/config_loader.py
 openpyxl
 paramiko  # deploy_to_jetson.py only
 

src/config_loader.py

@@ -0,0 +1,609 @@
+"""Unified config loader — single source of truth for the production pipeline.
+
+Reads `config.yaml` (canonical) + `.env` (secrets & deployment-only values:
+RTSP_URL, dashboard host/ports, SECRET_KEY, SITE_NAME).
+
+Design rules (see docs/configuration.md):
+- `config.yaml` owns stream/models/zones-knobs/counting/batch/output/camera.
+- `.env` owns RTSP_URL + dashboard host/ports/secret/site (never in YAML).
+- `zones.json` owns zone polygons only (knob keys there are ignored, warned).
+- `cfg/tracker.yaml` owns tracker hyperparams (referenced, not parsed here).
+- Model modes are data: `models.modes` maps mode id -> engines + class filters.
+  Adding mode E/F/... is a YAML-only change (no Python edits).
+- No heavy dependencies: stdlib dataclasses + pyyaml only (pyyaml already
+  comes with ultralytics). Import-safe for CI smoke tests.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import warnings
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+try:
+    import yaml
+except ImportError:  # pragma: no cover
+    yaml = None  # type: ignore
+
+
+# --------------------------------------------------------------------------- #
+# Dataclasses (one per config.yaml section)
+# --------------------------------------------------------------------------- #
+
+@dataclass
+class StreamConfig:
+    rtsp_url: str = ""  # always from .env RTSP_URL, never YAML
+    resolution: Tuple[int, int] = (1280, 720)
+    inference_stride: int = 2
+
+
+@dataclass
+class ModeEngine:
+    """One engine in a mode preset: which file + which classes it contributes."""
+
+    path: str = ""  # key into ModelsConfig.paths
+    classes: List[str] = field(default_factory=list)
+
+
+@dataclass
+class ModelMode:
+    """A mode preset: ONLY engine selection + class filters (nothing else)."""
+
+    description: str = ""
+    engines: List[ModeEngine] = field(default_factory=list)
+    class_filters: Dict[str, List[str]] = field(default_factory=dict)
+
+
+@dataclass
+class DetectionParams:
+    conf: float = 0.35
+    iou: float = 0.7  # Ultralytics default; matches pre-YAML behaviour
+    min_bbox_area: int = 1500  # px^2 @1280x720; permissive guardrail
+
+
+@dataclass
+class ModelsConfig:
+    active_mode: str = "C"
+    paths: Dict[str, str] = field(default_factory=dict)
+    modes: Dict[str, ModelMode] = field(default_factory=dict)
+    detection_params: Dict[str, DetectionParams] = field(default_factory=dict)
+
+
+@dataclass
+class ZonesConfig:
+    config_file: str = "zones.json"
+
+
+@dataclass
+class TrackerConfig:
+    config_file: str = "cfg/tracker.yaml"
+
+
+@dataclass
+class CountingConfig:
+    confirm_delay_sec: float = 0.5
+    exit_confirm_delay_sec: float = 6.0
+    entry_overlap_threshold: float = 0.20
+    exit_overlap_threshold: float = 0.05
+    camera_noise_deadband: int = 50
+    duplicate_circle_radius: int = 30
+    min_valid_area: int = 15000
+    max_reid_transit_distance: int = 400
+    circle_stay_timeout_sec: float = 10.0
+    jarak_toleransi_duplikat: int = 30
+    tolerance_missing_frames: int = 1200
+
+
+@dataclass
+class BatchConfig:
+    timeout_seconds: float = 30.0
+    merge_threshold_seconds: int = 300
+    daily_cutoff_time: str = "06:00"
+
+
+@dataclass
+class OutputConfig:
+    dir: str = "/opt/jetson-counter"
+    db_name: str = "jetson_counter.db"
+    state_file: str = "current_batch.json"
+    batch_mode_file: str = "batch_mode.json"  # DEPRECATED (mode now in config.yaml)
+    live_frame_path: str = "/dev/shm/jetson-counter/live_frame.jpg"
+    live_status_path: str = "/dev/shm/jetson-counter/live_status.json"
+
+
+@dataclass
+class DashboardConfig:
+    """Deployment-only values — always from .env, never YAML."""
+
+    host: str = "0.0.0.0"
+    port: int = 5000
+    office_port: int = 5721
+    secret_key: str = "change-me-in-production"
+    site_name: str = "LIVE"
+    debug: bool = False
+
+
+@dataclass
+class CameraConfig:
+    name: str = "CC1"
+    object_label: str = "karung-pakan"
+
+
+@dataclass
+class Config:
+    stream: StreamConfig = field(default_factory=StreamConfig)
+    models: ModelsConfig = field(default_factory=ModelsConfig)
+    zones: ZonesConfig = field(default_factory=ZonesConfig)
+    tracker: TrackerConfig = field(default_factory=TrackerConfig)
+    counting: CountingConfig = field(default_factory=CountingConfig)
+    batch: BatchConfig = field(default_factory=BatchConfig)
+    output: OutputConfig = field(default_factory=OutputConfig)
+    dashboard: DashboardConfig = field(default_factory=DashboardConfig)
+    camera: CameraConfig = field(default_factory=CameraConfig)
+
+    # -- mode helpers ------------------------------------------------------ #
+    def get_active_mode(self) -> ModelMode:
+        try:
+            return self.models.modes[self.models.active_mode]
+        except KeyError:
+            raise ValueError(
+                f"config.yaml models.active_mode={self.models.active_mode!r} "
+                f"not defined in models.modes (available: {sorted(self.models.modes)})"
+            )
+
+    def detection_params_for(self, class_name: str) -> DetectionParams:
+        try:
+            return self.models.detection_params[class_name]
+        except KeyError:
+            raise ValueError(
+                f"No detection_params for class {class_name!r} "
+                f"(available: {sorted(self.models.detection_params)})"
+            )
+
+    def engine_path(self, key: str, base_dir: str | Path = ".") -> str:
+        """Resolve a mode engine key via models.paths (relative to base_dir).
+
+        Sibling fallback (dev convenience, mirrors pre-YAML behaviour):
+        missing `.engine` -> existing `.pt` next to it, with a WARNING.
+        """
+        try:
+            rel = self.models.paths[key]
+        except KeyError:
+            raise ValueError(
+                f"Engine key {key!r} not defined in models.paths "
+                f"(available: {sorted(self.models.paths)})"
+            )
+        p = Path(base_dir) / rel
+        if not p.exists() and p.suffix == ".engine":
+            pt_sibling = p.with_suffix(".pt")
+            if pt_sibling.exists():
+                warnings.warn(
+                    f"Engine {p} missing — falling back to {pt_sibling} "
+                    f"(export the .engine for production).",
+                    UserWarning,
+                    stacklevel=2,
+                )
+                return str(pt_sibling)
+        return str(p)
+
+
+# --------------------------------------------------------------------------- #
+# Parsing
+# --------------------------------------------------------------------------- #
+
+def _require_yaml() -> None:
+    if yaml is None:
+        raise RuntimeError(
+            "pyyaml is required by src/config_loader.py "
+            "(pip install pyyaml — already a dependency of ultralytics)"
+        )
+
+
+def _parse_detection_params(raw: Dict[str, Any]) -> Dict[str, DetectionParams]:
+    out: Dict[str, DetectionParams] = {}
+    for cls_name, vals in (raw or {}).items():
+        vals = vals or {}
+        out[cls_name] = DetectionParams(
+            conf=float(vals.get("conf", 0.35)),
+            iou=float(vals.get("iou", 0.7)),
+            min_bbox_area=int(vals.get("min_bbox_area", 1500)),
+        )
+    return out
+
+
+def _parse_models(raw: Dict[str, Any]) -> ModelsConfig:
+    raw = raw or {}
+    modes: Dict[str, ModelMode] = {}
+    for mode_id, m in (raw.get("modes") or {}).items():
+        m = m or {}
+        engines: List[ModeEngine] = []
+        for e in m.get("engines") or []:
+            if isinstance(e, str):  # short form: bare path key, classes unknown
+                engines.append(ModeEngine(path=e, classes=[]))
+            else:
+                engines.append(ModeEngine(
+                    path=str(e.get("path", "")),
+                    classes=list(e.get("classes") or []),
+                ))
+        modes[str(mode_id).upper()] = ModelMode(
+            description=str(m.get("description", "")),
+            engines=engines,
+            class_filters={k: list(v or []) for k, v in (m.get("class_filters") or {}).items()},
+        )
+    return ModelsConfig(
+        active_mode=str(raw.get("active_mode", "C")).upper(),
+        paths={k: str(v) for k, v in (raw.get("paths") or {}).items()},
+        modes=modes,
+        detection_params=_parse_detection_params(raw.get("detection_params")),
+    )
+
+
+def _dashboard_from_env() -> DashboardConfig:
+    return DashboardConfig(
+        host=os.getenv("DASHBOARD_HOST", "0.0.0.0"),
+        port=int(os.getenv("DASHBOARD_PORT", "5000")),
+        office_port=int(os.getenv("OFFICE_PORT", "5721")),
+        secret_key=os.getenv("SECRET_KEY", "change-me-in-production"),
+        site_name=os.getenv("SITE_NAME", "LIVE"),
+        debug=os.getenv("FLASK_DEBUG", "false").lower() == "true",
+    )
+
+
+def _default_models() -> ModelsConfig:
+    """Legacy models section (modes A-D + paths + params) for the .env-backup path.
+
+    Mirrors config.yaml defaults so a missing config.yaml behaves identically.
+    """
+    paths = {
+        "combined": "models/v4-best.engine",
+        "truck_only": "models/v4-best.engine",
+        "sack_only": "models/best.engine",
+        "box": "models/yolo11n-bbox-100ep-sack+box-20260909-best.engine",
+        "truck_detector": "models/truck-detector.engine",
+    }
+    params = {
+        "truck": DetectionParams(conf=0.35, iou=0.7, min_bbox_area=5000),
+        "sack": DetectionParams(conf=0.35, iou=0.7, min_bbox_area=1500),
+        "box": DetectionParams(conf=0.35, iou=0.7, min_bbox_area=1500),
+    }
+    truck_f = {"truck": ["truck"]}
+    sack_f = {"sack": ["sack"]}
+    box_f = {"box": ["box"]}
+    modes = {
+        "A": ModelMode(
+            description="Combined v4 sack+truck only (legacy, no box counting)",
+            engines=[ModeEngine(path="combined", classes=["truck", "sack"])],
+            class_filters={**truck_f, **sack_f, "box": []},
+        ),
+        "B": ModelMode(
+            description="v4 truck-only + yolo11n sack+box (shared tracker)",
+            engines=[
+                ModeEngine(path="truck_only", classes=["truck"]),
+                ModeEngine(path="box", classes=["sack", "box"]),
+            ],
+            class_filters={**truck_f, **sack_f, **box_f},
+        ),
+        "C": ModelMode(
+            description="Combined v4 sack+truck + yolo11n box-only (dedicated tracker)",
+            engines=[
+                ModeEngine(path="combined", classes=["truck", "sack"]),
+                ModeEngine(path="box", classes=["box"]),
+            ],
+            class_filters={**truck_f, **sack_f, **box_f},
+        ),
+        "D": ModelMode(
+            description="v4 truck-only + best sack-only + yolo11n box-only",
+            engines=[
+                ModeEngine(path="truck_only", classes=["truck"]),
+                ModeEngine(path="sack_only", classes=["sack"]),
+                ModeEngine(path="box", classes=["box"]),
+            ],
+            class_filters={**truck_f, **sack_f, **box_f},
+        ),
+    }
+    return ModelsConfig(active_mode="C", paths=paths, modes=modes,
+                        detection_params=params)
+
+
+def _defaults_for_platform() -> Config:
+    """Fallback Config when config.yaml is missing: .env values + code defaults.
+
+    Mirrors the pre-YAML behaviour of predict.py / counter_dashboard.py so
+    nothing breaks (emits WARNING to nudge migration to config.yaml).
+    """
+    if os.name == "nt":
+        out_dir = "d:/Belajar/menghitung karung"
+        live_frame = f"{out_dir}/live_frame.jpg"
+        live_status = f"{out_dir}/live_status.json"
+    else:
+        out_dir = os.getenv("OUTPUT_DIR", "/opt/jetson-counter")
+        live_frame = os.getenv("LIVE_STREAM_FRAME_PATH", "/dev/shm/jetson-counter/live_frame.jpg")
+        live_status = os.getenv(
+            "LIVE_STATUS_FILE", "/dev/shm/jetson-counter/live_status.json"
+        )
+    cfg = Config(
+        stream=StreamConfig(
+            rtsp_url=os.getenv("RTSP_URL", ""),
+            resolution=(1280, 720),
+            inference_stride=2,
+        ),
+        models=_default_models(),
+        output=OutputConfig(
+            dir=out_dir,
+            db_name="jetson_counter.db",
+            state_file="current_batch.json",
+            batch_mode_file="batch_mode.json",
+            live_frame_path=live_frame,
+            live_status_path=live_status,
+        ),
+        camera=CameraConfig(
+            name=os.getenv("CAMERA_NAME", "CC1"),
+            object_label=os.getenv("OBJECT_LABEL", "karung-pakan"),
+        ),
+        batch=BatchConfig(
+            timeout_seconds=30.0,
+            merge_threshold_seconds=int(os.getenv("BATCH_MERGE_THRESHOLD_SECONDS", "300")),
+            daily_cutoff_time=os.getenv("DAILY_CUTOFF_TIME", "06:00"),
+        ),
+        dashboard=_dashboard_from_env(),
+    )
+    return cfg
+
+
+def load_config(path: str | Path = "config.yaml") -> Config:
+    """Load Config from YAML + .env. Missing file -> .env/defaults + WARNING."""
+    _require_yaml()
+    p = Path(path)
+    if not p.exists():
+        warnings.warn(
+            f"config file {p} not found — falling back to .env + built-in defaults. "
+            f"Create {p} (see config.yaml in repo) to silence this warning.",
+            UserWarning,
+            stacklevel=2,
+        )
+        return _defaults_for_platform()
+
+    with open(p, "r", encoding="utf-8") as f:
+        raw = yaml.safe_load(f) or {}
+
+    stream_raw = raw.get("stream") or {}
+    zones_raw = raw.get("zones") or {}
+    tracker_raw = raw.get("tracker") or {}
+    counting_raw = raw.get("counting") or {}
+    batch_raw = raw.get("batch") or {}
+    output_raw = raw.get("output") or {}
+    camera_raw = raw.get("camera") or {}
+
+    res = stream_raw.get("resolution", [1280, 720])
+
+    # Output paths: explicit legacy env vars win (pre-YAML behaviour),
+    # otherwise config.yaml. dir+name split reproduces the resolved absolute
+    # path via os.path.join (all production files share one dir in practice).
+    if os.name == "nt":
+        out_dir = os.getenv("OUTPUT_DIR", "d:/Belajar/menghitung karung")
+        _db = os.getenv("DB_PATH", f"{out_dir}/jetson_counter.db")
+        _st = os.getenv("STATE_FILE", f"{out_dir}/current_batch.json")
+        _bm = os.getenv("BATCH_MODE_FILE", f"{out_dir}/batch_mode.json")
+        _lf = os.getenv("LIVE_STREAM_FRAME_PATH", f"{out_dir}/live_frame.jpg")
+        _ls = os.getenv("LIVE_STATUS_FILE", f"{out_dir}/live_status.json")
+    else:
+        out_dir = os.getenv("OUTPUT_DIR", str(output_raw.get("dir", "/opt/jetson-counter")))
+        _db = os.getenv("DB_PATH", os.path.join(out_dir, str(output_raw.get("db_name", "jetson_counter.db"))))
+        _st = os.getenv("STATE_FILE", os.path.join(out_dir, str(output_raw.get("state_file", "current_batch.json"))))
+        _bm = os.getenv("BATCH_MODE_FILE", os.path.join(out_dir, str(output_raw.get("batch_mode_file", "batch_mode.json"))))
+        _lf = os.getenv("LIVE_STREAM_FRAME_PATH", str(output_raw.get(
+            "live_frame_path", "/dev/shm/jetson-counter/live_frame.jpg")))
+        _ls = os.getenv("LIVE_STATUS_FILE", str(output_raw.get(
+            "live_status_path", "/dev/shm/jetson-counter/live_status.json")))
+    output = OutputConfig(
+        dir=os.path.dirname(_db) if os.getenv("DB_PATH") else out_dir,
+        db_name=os.path.basename(_db),
+        state_file=os.path.basename(_st),
+        batch_mode_file=os.path.basename(_bm),
+        live_frame_path=_lf,
+        live_status_path=_ls,
+    )
+
+    cfg = Config(
+        stream=StreamConfig(
+            rtsp_url=os.getenv("RTSP_URL", ""),
+            resolution=(int(res[0]), int(res[1])),
+            inference_stride=int(stream_raw.get("inference_stride", 2)),
+        ),
+        models=_parse_models(raw.get("models")),
+        zones=ZonesConfig(config_file=str(zones_raw.get("config_file", "zones.json"))),
+        tracker=TrackerConfig(config_file=str(tracker_raw.get("config_file", "cfg/tracker.yaml"))),
+        counting=CountingConfig(
+            confirm_delay_sec=float(counting_raw.get("confirm_delay_sec", 0.5)),
+            exit_confirm_delay_sec=float(counting_raw.get("exit_confirm_delay_sec", 6.0)),
+            entry_overlap_threshold=float(counting_raw.get("entry_overlap_threshold", 0.20)),
+            exit_overlap_threshold=float(counting_raw.get("exit_overlap_threshold", 0.05)),
+            camera_noise_deadband=int(counting_raw.get("camera_noise_deadband", 50)),
+            duplicate_circle_radius=int(counting_raw.get("duplicate_circle_radius", 30)),
+            min_valid_area=int(counting_raw.get("min_valid_area", 15000)),
+            max_reid_transit_distance=int(counting_raw.get("max_reid_transit_distance", 400)),
+            circle_stay_timeout_sec=float(counting_raw.get("circle_stay_timeout_sec", 10.0)),
+            jarak_toleransi_duplikat=int(counting_raw.get("jarak_toleransi_duplikat", 30)),
+            tolerance_missing_frames=int(counting_raw.get("tolerance_missing_frames", 1200)),
+        ),
+        batch=BatchConfig(
+            timeout_seconds=float(batch_raw.get("timeout_seconds", 30.0)),
+            merge_threshold_seconds=int(os.getenv(
+                "BATCH_MERGE_THRESHOLD_SECONDS",
+                batch_raw.get("merge_threshold_seconds", 300))),
+            daily_cutoff_time=str(os.getenv(
+                "DAILY_CUTOFF_TIME",
+                batch_raw.get("daily_cutoff_time", "06:00"))),
+        ),
+        output=output,
+        dashboard=_dashboard_from_env(),
+        camera=CameraConfig(
+            name=str(camera_raw.get("name", os.getenv("CAMERA_NAME", "CC1"))),
+            object_label=str(camera_raw.get("object_label", os.getenv("OBJECT_LABEL", "karung-pakan"))),
+        ),
+    )
+    validate_config(cfg)
+    return cfg
+
+
+def validate_config(cfg: Config) -> None:
+    """Raise ValueError on inconsistent mode/paths/params (fail fast at startup)."""
+    if not cfg.models.modes:
+        raise ValueError("config.yaml models.modes is empty — define at least one mode")
+    if cfg.models.active_mode not in cfg.models.modes:
+        raise ValueError(
+            f"models.active_mode={cfg.models.active_mode!r} not in models.modes "
+            f"(available: {sorted(cfg.models.modes)})"
+        )
+    for mode_id, mode in cfg.models.modes.items():
+        for e in mode.engines:
+            if e.path not in cfg.models.paths:
+                raise ValueError(
+                    f"Mode {mode_id!r} references engine {e.path!r} "
+                    f"missing from models.paths (available: {sorted(cfg.models.paths)})"
+                )
+    for cls_name, dp in cfg.models.detection_params.items():
+        if not (0.0 < dp.conf <= 1.0):
+            raise ValueError(f"detection_params.{cls_name}.conf={dp.conf} must be in (0, 1]")
+        if not (0.0 < dp.iou <= 1.0):
+            raise ValueError(f"detection_params.{cls_name}.iou={dp.iou} must be in (0, 1]")
+        if dp.min_bbox_area < 0:
+            raise ValueError(
+                f"detection_params.{cls_name}.min_bbox_area={dp.min_bbox_area} must be >= 0"
+            )
+
+
+# --------------------------------------------------------------------------- #
+# Mode resolution (precedence) + legacy migration helpers
+# --------------------------------------------------------------------------- #
+
+def resolve_active_mode(explicit: Optional[str], cfg: Config) -> str:
+    """Precedence: explicit --model-mode > MODEL_MODE env (deprecated) > config.yaml.
+
+    Returns the validated (uppercased) mode id. Unknown ids fall back to
+    config.yaml active_mode with a WARNING (never crash production).
+    """
+    if explicit:
+        mode = explicit.upper()
+    else:
+        env_mode = os.getenv("MODEL_MODE", "")
+        if env_mode:
+            warnings.warn(
+                "MODEL_MODE env var is deprecated — set models.active_mode in "
+                "config.yaml instead. Env value still honoured for this run.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+            mode = env_mode.upper()
+        else:
+            mode = cfg.models.active_mode
+    if mode not in cfg.models.modes:
+        warnings.warn(
+            f"Model mode {mode!r} not defined in config.yaml models.modes — "
+            f"falling back to {cfg.models.active_mode!r}.",
+            UserWarning,
+            stacklevel=2,
+        )
+        return cfg.models.active_mode
+    return mode
+
+
+def set_active_mode(path: str | Path, mode: str) -> str:
+    """Persist a mode switch to config.yaml (atomic tmp+replace, comments preserved).
+
+    Uses a targeted line edit (not yaml.dump) so human comments/formatting survive.
+    Validates the mode exists before touching the file. Manual service restart
+    still required to apply (models load once at startup).
+    """
+    mode = mode.upper()
+    cfg = load_config(path)  # validates file + all modes
+    if mode not in cfg.models.modes:
+        raise ValueError(
+            f"Invalid model mode {mode!r} (available: {sorted(cfg.models.modes)})"
+        )
+    p = Path(path)
+    text = p.read_text(encoding="utf-8")
+    new_text, n = re.subn(
+        r'^(\s*active_mode\s*:\s*)["\']?\w+["\']?',
+        lambda m: f'{m.group(1)}"{mode}"',
+        text,
+        count=1,
+        flags=re.M,
+    )
+    if n != 1:
+        raise RuntimeError(
+            f"Could not locate a single 'active_mode:' line in {p} — file left untouched"
+        )
+    tmp = p.with_suffix(p.suffix + ".tmp")
+    tmp.write_text(new_text, encoding="utf-8")
+    os.replace(tmp, p)
+    return mode
+
+
+def check_legacy_batch_mode(cfg: Config, batch_mode_path: str | Path) -> Optional[str]:
+    """One-time migration nudge: warn if legacy batch_mode.json disagrees with config.yaml.
+
+    Returns the warning message (or None). Caller decides how to surface it.
+    """
+    import json
+
+    try:
+        with open(batch_mode_path, "r", encoding="utf-8") as f:
+            stored = json.load(f)
+        legacy = (stored.get("model_mode") or "").upper()
+    except (OSError, ValueError):
+        return None
+    if legacy and legacy != cfg.models.active_mode:
+        return (
+            f"Legacy {batch_mode_path} pins model_mode={legacy!r} but config.yaml "
+            f"active_mode={cfg.models.active_mode!r} wins. Delete {batch_mode_path} "
+            f"(or align it) to silence this warning."
+        )
+    return None
+
+
+# --------------------------------------------------------------------------- #
+# zones.json polygons (polygons only — knob keys there are ignored)
+# --------------------------------------------------------------------------- #
+
+# Knob keys that used to live in zones.json and are now owned by config.yaml.
+# If present, they are IGNORED (config.yaml wins) — warn once to guide migration.
+# NOTE: palet/truck/counting polygons + left_limit/right_limit + external_stream_url
+# stay in zones.json (geometry ownership); they are NOT in this set.
+LEGACY_ZONE_KNOBS = {
+    "duplicate_circle_radius", "min_valid_area",
+    "jarak_toleransi_duplikat", "max_reid_transit_distance",
+    "circle_stay_timeout_sec", "inference_stride",
+    "confirm_delay_sec", "exit_confirm_delay_sec",
+}
+
+
+def read_zone_polygons(path: str | Path) -> Dict[str, Any]:
+    """Read palet/truck/counting polygons from zones.json.
+
+    Returns {"palet": [...], "truck": [...], "counting": [...]}.
+    Warns if legacy knob keys are present (they are ignored — config.yaml owns them).
+    """
+    import json
+
+    with open(path, "r", encoding="utf-8") as f:
+        data = json.load(f)
+    ignored = sorted(LEGACY_ZONE_KNOBS & set(data.keys()))
+    if ignored:
+        warnings.warn(
+            f"{path} contains legacy knob keys {ignored} — IGNORED, "
+            f"config.yaml owns these values now. Keep only palet/truck/counting polygons.",
+            UserWarning,
+            stacklevel=2,
+        )
+    return {
+        "palet": data.get("palet", []),
+        "truck": data.get("truck", []),
+        "counting": data.get("counting", []),
+        "left_limit": float(data.get("left_limit", 0.27578)),
+        "right_limit": float(data.get("right_limit", 0.72578)),
+        "external_stream_url": data.get("external_stream_url", ""),
+    }

src/detection.py

@@ -23,14 +23,16 @@ class BaseDetector:
         model_path: str | YOLO,
         conf: float = 0.35,
         class_filter: tuple[str, ...] | list[str] | None = None,
+        iou: float = 0.7,
     ) -> None:
         self._model = model_path if isinstance(model_path, YOLO) else YOLO(model_path)
         self._conf = conf
         self._class_filter = set(class_filter) if class_filter else None
+        self._iou = iou
 
     def detect(self, frame: np.ndarray) -> list[Detection]:
         results = self._model.predict(
-            frame, conf=self._conf, verbose=False
+            frame, conf=self._conf, iou=self._iou, verbose=False
         )
         return self._parse(results[0])
 
@@ -69,8 +71,8 @@ class BaseDetector:
 class SackDetector(BaseDetector):
     """Detects sacks (drops persons/boxes/trucks from multi-class models)."""
 
-    def __init__(self, model_path: str | YOLO, conf: float = 0.35) -> None:
-        super().__init__(model_path, conf, class_filter=("sack",))
+    def __init__(self, model_path: str | YOLO, conf: float = 0.35, iou: float = 0.7) -> None:
+        super().__init__(model_path, conf, class_filter=("sack",), iou=iou)
 
 
 class TruckDetector(BaseDetector):
@@ -81,12 +83,13 @@ class TruckDetector(BaseDetector):
         model_path: str | YOLO,
         conf: float = 0.35,
         class_filter: tuple[str, ...] | list[str] | None = None,
+        iou: float = 0.7,
     ) -> None:
-        super().__init__(model_path, conf, class_filter=class_filter)
+        super().__init__(model_path, conf, class_filter=class_filter, iou=iou)
 
 
 class BoxDetector(BaseDetector):
     """Detects boxes (drops sacks/persons from the sack+box model)."""
 
-    def __init__(self, model_path: str | YOLO, conf: float = 0.35) -> None:
-        super().__init__(model_path, conf, class_filter=("box",))
+    def __init__(self, model_path: str | YOLO, conf: float = 0.35, iou: float = 0.7) -> None:
+        super().__init__(model_path, conf, class_filter=("box",), iou=iou)

src/tracking.py

@@ -28,7 +28,7 @@ _TRACKER_CFG = os.path.join(
 class ByteTrackTracker:
     """Tracks sacks across frames using FastTrack (occlusion-aware)."""
 
-    def __init__(self, model_path: str | YOLO, conf: float = 0.35) -> None:
+    def __init__(self, model_path: str | YOLO, conf: float = 0.35, iou: float = 0.7) -> None:
         if isinstance(model_path, YOLO):
             self._model = model_path
             self._model_path = getattr(model_path, "ckpt_path", str(model_path))
@@ -36,6 +36,7 @@ class ByteTrackTracker:
             self._model = YOLO(model_path)
             self._model_path = model_path
         self._conf = conf
+        self._iou = iou
         self._tracker_cfg = _TRACKER_CFG if os.path.exists(_TRACKER_CFG) else "bytetrack.yaml"
 
     def update(
@@ -45,6 +46,7 @@ class ByteTrackTracker:
         results = self._model.track(
             frame,
             conf=self._conf,
+            iou=self._iou,
             persist=True,
             tracker=self._tracker_cfg,
             verbose=False,

tests/test_config_loader.py

@@ -0,0 +1,128 @@
+"""Smoke tests for src/config_loader.py — pure-Python, no cv2/ultralytics needed."""
+
+import json
+import os
+
+import pytest
+import yaml
+
+from src.config_loader import (
+    check_legacy_batch_mode,
+    load_config,
+    read_zone_polygons,
+    resolve_active_mode,
+    set_active_mode,
+    validate_config,
+)
+
+
+@pytest.fixture()
+def repo_config_path():
+    p = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.yaml")
+    assert os.path.exists(p), "repo config.yaml missing"
+    return p
+
+
+def test_repo_config_loads_and_validates(repo_config_path):
+    cfg = load_config(repo_config_path)
+    assert cfg.models.active_mode == "C"
+    assert set(cfg.models.modes) >= {"A", "B", "C", "D"}
+    for mode_id, mode in cfg.models.modes.items():
+        assert mode.engines, f"mode {mode_id} has no engines"
+        for e in mode.engines:
+            assert e.path in cfg.models.paths, f"mode {mode_id}: unknown engine {e.path}"
+            assert e.classes, f"mode {mode_id}: engine {e.path} declares no classes"
+    for cls_name in ("truck", "sack", "box"):
+        dp = cfg.detection_params_for(cls_name)
+        assert 0.0 < dp.conf <= 1.0
+        assert 0.0 < dp.iou <= 1.0
+        assert dp.min_bbox_area >= 0
+
+
+def test_mode_preset_only_filters_and_engines(repo_config_path):
+    """Modes must not carry per-class conf/iou — those live in detection_params."""
+    with open(repo_config_path) as f:
+        raw = yaml.safe_load(f)
+    for mode_id, mode in raw["models"]["modes"].items():
+        assert "conf" not in mode, f"mode {mode_id}: conf belongs in detection_params"
+        assert "iou" not in mode, f"mode {mode_id}: iou belongs in detection_params"
+        assert "min_bbox_area" not in mode, f"mode {mode_id}"
+
+
+def test_resolve_active_mode_precedence(repo_config_path, monkeypatch):
+    cfg = load_config(repo_config_path)
+    monkeypatch.delenv("MODEL_MODE", raising=False)
+    assert resolve_active_mode(None, cfg) == "C"
+    assert resolve_active_mode("d", cfg) == "D"
+    with pytest.warns(UserWarning):  # unknown -> fallback to active
+        assert resolve_active_mode("Z", cfg) == "C"
+    monkeypatch.setenv("MODEL_MODE", "B")
+    with pytest.warns(DeprecationWarning):  # env still honoured, deprecated
+        assert resolve_active_mode(None, cfg) == "B"
+
+
+def test_missing_file_falls_back_with_warning(tmp_path, monkeypatch):
+    monkeypatch.delenv("MODEL_MODE", raising=False)
+    with pytest.warns(UserWarning, match="not found"):
+        cfg = load_config(tmp_path / "nope.yaml")
+    assert cfg.models.active_mode == "C"
+    assert cfg.stream.inference_stride == 2
+
+
+def test_set_active_mode_roundtrip_and_validation(tmp_path, repo_config_path):
+    import shutil
+
+    dst = tmp_path / "config.yaml"
+    shutil.copy(repo_config_path, dst)
+    assert set_active_mode(dst, "d") == "D"
+    assert load_config(dst).models.active_mode == "D"
+    text = dst.read_text()
+    assert 'active_mode: "D"' in text
+    assert "# config.yaml" in text  # comments preserved (no yaml.dump reformat)
+    with pytest.raises(ValueError, match="Invalid model mode"):
+        set_active_mode(dst, "Z")
+    assert load_config(dst).models.active_mode == "D"  # untouched on failure
+
+
+def test_future_mode_extensible_without_code(tmp_path, repo_config_path):
+    """Adding mode E is a YAML-only change: loader accepts it, no code edits."""
+    import shutil
+
+    dst = tmp_path / "config.yaml"
+    shutil.copy(repo_config_path, dst)
+    with open(dst) as f:
+        raw = yaml.safe_load(f)
+    raw["models"]["modes"]["E"] = {
+        "description": "hypothetical future preset",
+        "engines": [{"path": "combined", "classes": ["truck", "sack"]}],
+        "class_filters": {"truck": ["truck"], "sack": ["sack"], "box": []},
+    }
+    with open(dst, "w") as f:
+        yaml.safe_dump(raw, f)
+    cfg = load_config(dst)
+    assert resolve_active_mode("E", cfg) == "E"
+    assert set_active_mode(dst, "E") == "E"
+
+
+def test_legacy_batch_mode_check(tmp_path, repo_config_path):
+    cfg = load_config(repo_config_path)
+    legacy = tmp_path / "batch_mode.json"
+    legacy.write_text(json.dumps({"mode": "auto", "model_mode": "B"}))
+    msg = check_legacy_batch_mode(cfg, legacy)
+    assert msg is not None and "'B'" in msg and "'C'" in msg
+    legacy.write_text(json.dumps({"mode": "auto", "model_mode": "C"}))
+    assert check_legacy_batch_mode(cfg, legacy) is None
+    assert check_legacy_batch_mode(cfg, tmp_path / "missing.json") is None
+
+
+def test_zone_polygons_and_legacy_knob_warning(tmp_path):
+    zf = tmp_path / "zones.json"
+    zf.write_text(json.dumps({
+        "palet": [[0, 0]], "truck": [[1, 1]], "counting": [[2, 2]],
+        "left_limit": 0.27, "right_limit": 0.72,
+        "duplicate_circle_radius": 30,  # legacy knob -> ignored + warned
+    }))
+    with pytest.warns(UserWarning, match="legacy knob"):
+        zones = read_zone_polygons(zf)
+    assert zones["palet"] == [[0, 0]]
+    assert zones["left_limit"] == 0.27  # geometry stays in zones.json