Inicio · Capítulos · 14. Señales para visión artificial
14.9.28 Perfil concreto
OAK-D Pro AF PoE + Jetson Orin Nano
camera |
DepthAI v3 pipeline on RVC2 for sensor-side
acquisition/preprocessing; host receives bounded synchronized
outputs. |
host |
Jetson Orin Nano 8GB with NVMe; pin tested
JetPack/DepthAI/TensorRT/container matrix rather than following latest
at runtime. |
supervision |
systemd owns core host services/containers, restart budgets,
dependencies and watchdog integration. |
containers |
OCI images with NVIDIA runtime where GPU access is required;
explicit CPU/memory/PID/storage quotas. |
inference |
Prefer a serialized stage pipeline on 8GB until benchmark evidence
proves safe concurrency; TensorRT engines are exact-profile
artifacts. |
metadata |
SQLite WAL is acceptable for one edge node with disciplined
backup/checkpoint and local storage; migrate to PostgreSQL only when
concurrency/remote topology requires it. |
artifacts |
Content-addressed filesystem/object layout on NVMe; NAS receives
verified replicas instead of hosting active scratch/DB. |
telemetry |
Local OpenTelemetry Collector or equivalent agent receives OTLP,
host metrics and logs, then exports locally/remotely with bounded
queues. |
powerThermal |
Record nvpmodel/power mode, cooling and tegrastats-derived state as
deployment and benchmark evidence. |
Secuencia operacional de
una sesión
- Boot y verificación del release activo.
- Health de almacenamiento, cámara, clocks, modelos y
calibración.
- Carga de pipeline DepthAI y warm-up registrado.
- Inicio de sesión y scene revision.
- Captura de intentos con commit durable.
- Análisis por jobs reanudables según perfil de recursos.
- Publicación progresiva: preview, resultado parcial y resultado
final.
- Revisión del entrenador como nueva revisión.
- Export/backup/sync en segundo plano y dentro de presupuesto.
- Cierre limpio, checkpoint y reporte de sesión.