Karung Counter

Repo: https://git.proit.id/andrew/karung-counting-feedmill-semarang

AI video-analytics system that counts feed sacks (karung-pakan) being loaded onto / unloaded from trucks at a feedmill loading gate. It watches an RTSP camera, detects the truck bed, tracks sacks with persistent IDs, counts line-crossings per truck (batch), persists results to SQLite/CSV, and serves a live Flask dashboard.

Language note: UI strings, logs and comments are largely in Indonesian (karung = sack/bag, truk = truck, muat = load). Class names inside the YOLO models are English (sack, box, truck, person).

Pipeline (production, predict.py)

Frame → TruckDetect → ROI → SackTrack → Stabilize → Count → Dashboard / DB

Stage	Module	What it does
Stream	src/streaming.py	RTSP (RTSPSource) or video file (VideoFileSource)
Truck detect	src/detection.py (TruckDetector)	Dedicated truck model, run every 15 frames
ROI	src/truck_roi.py (TruckROITracker)	Picks main truck in lane, EMA-smooths bbox, derives counting line
Sack track	src/tracking.py (ByteTrackTracker)	YOLO + FastTrack/ByteTrack, tuned via cfg/tracker.yaml
Stabilize	src/stabilizer.py (BboxStabilizer)	EMA bbox smoothing + occlusion hold + height clamps
Count	src/counting.py (LineCrossCounter)	Zone-based line crossing: loading / unloading / net, 3-layer dedup
Batch	src/batch.py (BatchLifecycleManager)	4-state FSM: IDLE → TRUCK_STABILIZING → COUNTING_SACKS ⇄ WAITING_FOR_ACTIVITY
Output	src/logger.py, src/dashboard.py	CSV logs + on-frame overlay

Full detail: docs/architecture.md.

Quickstart

pip install -r requirements.txt   # Jetson: never pip-install torch from PyPI, use NVIDIA wheels
cp .env.example .env              # then edit RTSP_URL / paths

# Dev / offline (any flags omitted = production defaults):
python predict.py --source path/to/video.mp4 --env .env
python predict.py --source vid.mp4 --no-dashboard --no-db --output-dir /tmp/out --max-frames 500
python predict.py --help          # all flags: --config/--model/--model-mode/--output-json/--sack-conf/--truck-conf/--box-conf/--box-model/--batch-timeout

# Production (Jetson systemd, zero flags):
sudo systemctl restart karung-counter karung-counter-dashboard

Production on the Jetson runs predict.py + counter_dashboard.py as systemd services (see docs/deployment.md) — predict.py is the production pipeline (SQLite persistence, live-frame publishing to /dev/shm, zone polygons); the src/ package is the shared detection/tracking/counting library it builds on (the old v3 src/main.py loop is deprecated). Retired experiments live in archive/.

Models

All weights live in models/ (see models/modelREADME.md for the per-mode detector/filter matrix).

File	Classes	Role
models/truck-detector.{pt,engine}	truck	Specialized truck detector (ROI, batch lifecycle)
models/v4-best.{pt,onnx,engine} / models/model_karung_truk.{pt,onnx,engine}	sack, truck	Combined sack+truck (Modes A/B/C/D truck; Modes A/C sack)
models/karung-dimuat-…-seg-200e.{pt,onnx,engine}	person, sack	Segmentation model: sacks counted, persons drawn/skipped
models/yolo11n-bbox-100ep-sack+box-20260909-best.{pt,onnx,engine}	sack, box	Unified sack+box (Modes B/C/D)
models/best.{pt,onnx,engine}	sack	Sack-only specialist (Mode D sack)

.pt = PyTorch, .engine = TensorRT FP16 (Jetson GPU, what production loads), .onnx = ONNX export artifact. See docs/models.md for the multi-model / multi-class / hybrid support matrix.

Configuration

File	Purpose
config.yaml	Canonical config: stream, model modes/presets, counting knobs, batch, output paths, camera
.env (see .env.example)	Secrets + deployment only: RTSP URL, dashboard host/ports/secret/site
zones.json	Calibrated pallet / truck / counting polygons + left/right limits (geometry only)
cfg/tracker.yaml	FastTrack/ByteTrack occlusion tuning (buffer 60, re-ID windows)

See docs/configuration.md for every variable.

Dashboard (counter_dashboard.py, port 5000 / office 5721)

Pages in templates/: monitoring.html, operator.html (batch start/stop; DO panel in do_manual), history.html, analytics.html. JSON APIs under /api/* (live video MJPEG, current/previous batch, summary, daily data, batch mode, DO upload/staged/settings, CSV/Excel export). Data source: jetson_counter.db + current_batch.json + batch_mode.json (batch flow: auto | do_manual | manual; model mode in config.yaml).

Batch modes

Mode	Who opens/closes batch	DO gate
auto (default)	AI truck FSM	—
do_manual	Operator + DO photo scan	yes
manual	Operator buttons (legacy)	no

Office (:5721) switches mode / model mode / require_plate / ocr_engine (403 on operator POST). DO manual flow: open http://<jetson>:5000/operator on a phone → Ambil Foto DO → review/edit OCR fields → Mulai Batch → count → Selesai (soft-warn if zone busy). Discard only when both sack and box nets are 0. ERD: ERD.md.

Repository layout

src/                  Shared library (streaming, detection, tracking, stabilizer,
                      truck_roi, counting, batch, dashboard, logger, config_loader, main)
config.yaml           Canonical config (stream/models/counting/batch/output/camera)
cfg/tracker.yaml      Tracker tuning
predict.py            Single entrypoint: production service + dev CLI (see --help)
counter_dashboard.py  Flask dashboard + APIs  |  templates/*.html  pages
tests/                Pytest smoke tests (counting, batch, config, config_loader)
archive/              Retired experiments (predict_new.py, rpo_iki/, simple_predict.py,
                      check/merge/test scripts) — history preserved, not imported
export_model.py       Export .pt → TensorRT .engine (FP16)  |  export_v4.py  variant
deploy_to_jetson.py   Paramiko sync + service restart
*.service             systemd units (counter, dashboard, mediamtx)
zones.json            Calibrated zone geometry
batch_history_folder/ Per-batch JSON reports
CHANGELOG.md          Release history (dated entries, Keep-a-Changelog style)

Script-by-script reference: docs/scripts.md.

Docs

• docs/architecture.md — pipeline stages, modules, batch FSM
• docs/models.md — model inventory & detection-mode support matrix
• docs/configuration.md — all config files/variables
• docs/deployment.md — Jetson services, TensorRT, deploy flow
• ERD.md — SQLite ERD (Mermaid): batches / delivery_orders / daily_summaries
• docs/operator-tiles-history-plan.md — operator tiles + history cleanup plan
• docs/scripts.md — entry points & utility scripts
• CHANGELOG.md — release history