👽 Sipeed MaxiCam

Edge computer vision at the source. Run local CV models on the Sipeed MaixCam and stream real-time visual logs securely to your private memory layer.

Overview

The CI-MaixCam integration turns a Sipeed MaixCam into a dedicated edge computer vision sensor attached to your Companion Hub. Instead of streaming raw video to the cloud for analysis, vision models run directly on the MaixCam’s NPU — object detection, scene classification, face recognition (for personal use), OCR, and more — and only the resulting structured events are sent to your Digital Memory Server.

This design keeps raw video private while still building a rich, structured log of what the camera observes.

Key Features

• On-device inference — models run on the MaixCam K230 NPU; no raw video leaves the device
• Structured event stream — detections (objects, text, faces, activities) are streamed to Digital Memory as typed events
• Model management — deploy, swap, and update CV models from the Hub dashboard
• Snapshot capture — periodic or event-triggered still captures sent to Hub for memory enrichment
• Multi-camera — manage multiple MaixCam devices from a single Hub instance
• RTSP fallback — optionally stream raw video to internal Hub services (e.g. Frigate) when needed

Supported Hardware

Pre-installed Models

The integration ships with several ready-to-run models:

Custom ONNX models can be converted and deployed via the Hub model manager.

Use Cases

• Detect when specific objects (keys, packages, vehicles) appear in a room or driveway
• OCR whiteboards and documents in real time and ingest into Digital Memory
• Build a local occupancy log for home automation triggers in Home Assistant
• Attach to a 3D printer to detect print failures automatically
• Log what’s visible on your desk or workspace as a continuous ambient memory enrichment

Architecture

Sipeed MaixCam
├── Camera sensor
└── K230 NPU ──▶ On-device inference
       │  structured events (JSON over MQTT/HTTP)
       ▼
CI-MaixCam Hub App
├── Event receiver
├── Model manager (OTA model updates)
└── Memory bridge
       │
       ▼
Digital Memory Server
(object sightings, text captures, scene events)

Setup

pip install maixpy
maixpy flash --device /dev/ttyUSB0 --firmware ci-maixcam-v1.0.0.img

Usage

Viewing the Detection Feed

The Hub dashboard at http://maixcam.ci.localhost shows a live event stream from all registered cameras, including bounding boxes overlaid on JPEG snapshots.

Querying Detections

All detections are stored in Digital Memory. Ask your agent: “When was the last time a person was detected by the front camera?” or search the timeline for “package detected”.

Custom Automation Triggers

Send MaixCam detections to Home Assistant via MQTT for automation triggers:

# In your CI-MaixCam config
outputs:
  - type: mqtt
    topic: maixcam/detections
    broker: mqtt://homeassistant.ci.localhost

Troubleshooting

MaixCam not appearing in device discovery Ensure it is on the same LAN as your Hub. Check its Wi-Fi connection by opening its local UI at its assigned IP. Verify firewall rules allow traffic on port 8888 (CI-MaixCam’s event port).

Inference very slow Confirm the model is compiled for K230 KPU (.kmodel format). Unoptimized ONNX models fall back to CPU and are significantly slower.

Memory events missing object labels Check the model’s output class list in the Hub model manager. Ensure class labels are mapped correctly in the model configuration.

Links

• Sipeed MaixCam Hardware 
• Digital Memory Server