karung-counting-feedmill-semarang/docs/deployment.md

Deployment (Jetson)

Repo: https://git.proit.id/andrew/karung-counting-feedmill-semarang (git remote set-url origin <url> after the ervan → andrew transfer).

Systemd services

Unit	Runs	After
karung-counter.service	/usr/bin/python3 predict.py (cwd /home/jetson/karung, QT_QPA_PLATFORM=offscreen)	network.target
karung-counter-dashboard.service	/usr/bin/python3 counter_dashboard.py	network.target + counter
mediamtx.service	MediaMTX restream (see zones.json:external_stream_url)	—

Both app units: Restart=always, RestartSec=5, load EnvironmentFile=/home/jetson/karung/.env.

sudo systemctl enable --now karung-counter karung-counter-dashboard
sudo systemctl restart karung-counter karung-counter-dashboard
systemctl status karung-counter karung-counter-dashboard --no-pager

Deploy flow (deploy_to_jetson.py)

Paramiko sync of templates/{operator,monitoring,base}.html, counter_dashboard.py, predict.py, config.yaml, .env plus models/*.engine (v4-best, yolo11n-sack+box, best, truck-detector, model_karung_truk) → 192.168.192.96:/home/jetson/karung/ (creates remote models/ if missing, skips missing local files), then restarts both services and checks status + ports (5000/5721). Run from the dev machine. .pt/.onnx stay local-only (dev/export).

TensorRT export

On the Jetson (needs CUDA): python3 export_model.py models/<name>.pt exports to FP16 .engine next to the .pt (default: karung-dimuat seg model). Production loads .engine only — see models/modelREADME.md for which weights each mode uses.

Runtime data files

• SQLite jetson_counter.db: batches(counting_date, batch_number, camera_name, object_label, count, start/end_time, box_*, plate, do_numbers, expected_*, net_sack, net_box), daily_summaries(...), delivery_orders(...) (DO photos).
• current_batch.json (crash recovery), batch_mode.json (batch flow mode only: auto|do_manual|manual — model mode lives in config.yaml), do_settings.json (require_plate/do + ocr_engine), $OUTPUT_DIR/do_photos/YYYY-MM-DD/ (7-day retention), batch_history_folder/batch_<ts>.json + hasil_perhitungan.json (per-batch reports).
• Live frame: /dev/shm/jetson-counter/live_frame.jpg (written every 2nd frame, consumed by /api/live-video MJPEG).
• Helpers: check_jetson_db.py (root); retired ops scripts in archive/ (backup.py, dump_db.py, migrate_jetson_db.py, merge_batches_*.py, update_batches.py, diagnose_truck_jetson.py).

DO OCR packages

Default engine is RapidOCR (PP-OCR via onnxruntime, bundled models — offline-friendly; installed with requirements.txt):

pip install rapidocr_onnxruntime

Backups (ocr_engine — office monitoring UI toggle, no restart): Tesseract (Backup, Light):

# lighter; system package + Indonesian traineddata
sudo apt install tesseract-ocr tesseract-ocr-ind
pip install pytesseract

PaddleOCR (Accuracy, Heavy to run):

# optional; heavier — see paddleocr docs for Jetson wheels
pip install paddleocr

Missing deps → upload returns explicit error (never silent fallback); flip engine back to RapidOCR/Tesseract from the office dashboard (:5721). Operator page: http://<jetson>:5000/operator (smartphone camera capture for DO photos). Photo dir under output.dir with 7-day retention (hourly purge in dashboard process).

Dashboard (counter_dashboard.py)

Pages: / + /monitoring, /operator (manual start/stop; DO panel in do_manual), /history, /analytics. Key APIs: /api/live-video, /api/current-batch, /api/previous-batch, /api/batch/{start,stop,stop-preview,mode}, /api/model-modes (mode list is derived from config.yaml, so future modes appear automatically), /api/do/{upload,photo/<id>,staged,settings,retention} + PUT/DELETE /api/do/<id>, /api/summary, /api/daily-data, /api/day-detail/<date>, /api/recent-batches, /api/available-dates, /api/export-daily-csv, /api/export-day-csv/<date> (Excel via openpyxl), /api/batch-clip/<date>/<batch_number> + /status, /file, /api/batch-clip/pending (office-only, see below).

Port split: mode / model_mode / require_plate / require_do / ocr_engine POSTs → office 5721 only (403 on 5000). Smartphones open http://<host>:5000/operator for camera capture.

Batch clip download (motionEye)

History page Klip button runs an async job (all endpoints office port 5721 only, 403 on 5000):

• POST /api/batch-clip/<date>/<batch_number> — idempotent: starts the job or returns the existing one → 200 {"success": true, "status": "running"|"ready", "job": "<date>/<n>"}.
• GET …/<date>/<batch_number>/status → {"success": true, "status": "running"|"ready"|"error", "error": "<msg|empty>", "elapsed": <s>} (404 if no job).
• GET …/<date>/<batch_number>/file → mp4 attachment (409 while running, 404 if error/none).
• GET /api/batch-clip/pending → all jobs, newest first — restores button state after reload/navigation and feeds the cross-page toast.

Button states: Klip → Memproses… <s> dtk (disabled; status polled every 1.5 s, pending refreshed every 3 s while a job on the page runs) → Siap — unduh → Unduh lagi after the first download (instant, file already exists); failure → Gagal (click shows the error). Jobs for other dates are left to the toast.

Toast: every office page (templates/base.html) polls pending every 3 s — bootstrap once, a 403 on the operator port disables polling for that tab — and shows a non-blocking bottom-right Klip siap / Klip gagal card (auto-dismiss 8 s, click → /history). Shown once per job per tab (sessionStorage clipNotified, capped at 50 keys).

Rendering itself is unchanged (clip_batch in src/clip.py, ffmpeg under the hood). The finished temp file is kept ≤1 h so Unduh lagi re-downloads without re-rendering; older clip_*.mp4 files are swept then. Jobs live in dashboard memory — a karung-counter-dashboard restart drops them (status 404 → button falls back to Klip, press again).

• URL keys (.env): MOTIONEYE_URL (base URL; empty → 404 motionEye belum dikonfigurasi (MOTIONEYE_URL)), MOTIONEYE_CAMERA_ID (default 2).
• MOTIONEYE_CLIP_PAD — float seconds padded before/after the batch window at both cut points (default 3 → ±3 s; empty/invalid falls back to 3.0).
• MOTIONEYE_OSD_ALIGN — OSD-clock alignment, default on (unset = on; 0 / false / no / empty = off). Why: the camera's burned-in OSD clock and the recording clock drift — up to ~+22 s inside a long clip — because of dropped frames, so filename-based offsets cut at the wrong media time. When on, clip_batch OCRs 1 frame at each cut point (~2–6 s extra per clip, result cached per clip+second) and corrects the cut times. 0 → filename-based offsets only. If OCR fails, it silently falls back to filename-based offsets; both cases log one [CLIP] batch=... pad=... align=... line at request start (plus any [CLIP] alignment detail lines) — check with journalctl -u karung-counter-dashboard.
• motionEye HTTP API used: GET /movie/<id>/list/ (recording segments in the window) and GET /movie/<id>/download<path> (segment bytes).
• Requires ffmpeg / ffprobe on PATH — both already installed on the Jetson.
• Recording retention: motionEye preserve_movies: 3 → only the last 3 days of recordings exist. Older batches return 404 (ClipError, no recording in window).
• Errors are surfaced on the job: …/status reports status: "error" with the error message (…/file → 404), and a failing POST answers non-2xx {"success": false, "error": "<msg>"}. Finished file lives at $(dirname DB)/tmp/clip_<uuid>.mp4 (kept ≤1 h, then swept).
• Output filename built from the batch start_time: b{batch}-{YYYY-MM-DD}-{HH-MM-SS}.mp4 (e.g. batch 21 starting 2026-09-26T14:25:09.123456 → b21-2026-09-26-14-25-09.mp4).