Phantom Vision VMS / Ghost Protocol

Appearance

About this article ​

  • Audience: Firmware authors, integrators, and operators deploying on Shadow 1
  • Goal: Internalize hardware guardrails, imaging standards, and cloud integration contracts before changing pipelines
  • Type: Concept

Summary ​

Ghost Protocol edge firmware targets Shadow 1 (RVC4 / Qualcomm QCS8550) running device OS 1.30.1+ with the Edge AI v3 vision pipeline. Requirements align with OpenIPC modular design, OWASP IoT ISVS hardening, ONVIF interoperability, and Ghost Protocol pipeline best practices.

Prerequisites ​

Platform baseline ​

AttributeRequirement
HardwareShadow 1, IP67 enclosure, 48 MP IMX586 sensor, 52 TOPS AI (48 INT8 DSP + 4 FP16 GPU)
OSDevice OS 1.30.1+, Linux kernel 5.15
FrameworkEdge AI v3 unified dai.node.Camera pipelines
CloudGoogle Cloud VMS integration via outbound MQTT and optional WebRTC/SRTSP archival

Connectivity ​

  • Preferred: LAN over PoE+ (IEEE 802.3at) on the industrial Ethernet port — power + data for production fleets.
  • Supported: WLAN via M8 connector + Wi‑Fi adapter for failover or sites without hardline.
  • M8 accessories: I/O relay controller and CANbus adapter — see Shadow connectivity & M8 accessories.

Power and thermal ​

  • PoE+ (IEEE 802.3at) required for peak 4K@60 FPS encode plus inference; standard 15 W PoE forces downscaled inference and throttled encoding. Size budget for M8 accessories too.
  • Firmware monitors power delivery and degrades gracefully when budget is constrained.
  • Thermal guardrails via SystemLogger:
    • Optimal: die < 70 °C — full pipeline
    • Warning: die > 75 °C — reduce inference FPS, prioritize video
    • Critical: die > 85 °C — suspend AI, maintain RTSP/MQTT, publish THERMAL_CRITICAL
  • Memory budget (8 GB RAM): ~2 GB video buffer, ~4 GB AI execution, ~2 GB OS and services. eMMC wear managed through rolling retention and prune policies.

Imaging and spatial metadata ​

  • Evidence mode — 4K (3840×2160) up to 60 FPS H.265 for NVR ingestion
  • Action mode — 1080p high frame rate downscale for tracking workloads
  • Spatial fusion — Shadow 1 is single-lens; use dai.node.NeuralDepth to produce XYZ coordinates (RDF frame) alongside detections
  • Tracking — On-device dai.node.ObjectTracker with device-side sync; avoid host-only timestamp matching
  • Bounding boxes alone are insufficient for fusion-aware analytics

Deployment and OTA ​

  • Custom CLI (vms-cli) replaces legacy oakctl workflows for deploy, model hot-swap, and status telemetry.
  • A/B rollback — New builds land on partition B; watchdog reverts to partition A if pipeline init fails.
  • Model format — Convert models through the Ghost Protocol ML Pipeline / Qualcomm AI Hub to .superblob or Qualcomm SNPE .dlc; raw ONNX falls back to CPU and violates thermal budgets.
  • Container DNS must propagate without restart (systemd-resolved on OS 1.30.1+).

Cloud and security integration ​

  • Telemetry — Local MQTT client publishes JSON metadata to Google Cloud Pub/Sub; video and metadata paths stay decoupled.
  • Live viewing — WebRTC signaling from the edge for low-latency preview.
  • Archival — Secure RTSP or ONVIF Profile S over VPN (for example WireGuard) into object storage ingestion.
  • Zero trust — Outbound-only control plane (MQTT or WebSocket); mTLS or OIDC for GCP authentication; inbound SSH/HTTP disabled by default.

Payloads include timestamp, camera_id (role label), detection list with optional spatial fields, and telemetry block (die temp, fps).

Reference standards ​

StandardApplication
OpenIPCModular firmware and human-readable stream configs
OWASP IoT ISVSDevice hardening and secure update mechanics
ONVIF (use-go/onvif)Discovery, Profile S video, Profile M metadata
Edge AI architectureCamera node and pipeline patterns

Next steps ​

Operator depth ​

Hardware-specific calibration, thermal probe logs, and deployment evidence live in private operator handoff (not published).