Hardware integration

Cognitive Companion is designed to work with affordable, readily available edge hardware. This page covers the supported devices and how they integrate with the system.

Inference hardware

Continuous Tracking can keep its stateful services on a home server while moving detector, pose, body ReID, and face inference to a separate accelerator.

• Run CTS inference on Jetson Orin Nano Super covers the six-camera recommendation, eight-camera qualification gates, the three CTS models, the complete five-model Buffalo_L pack, and production metrics.
• Model quantization and accelerator portability explains calibration, PTQ, QAT, sparse INT8, Triton internals, and paths for Intel or AMD discrete GPUs.
• Camera setup and tuning guide covers exposure, resolution, mount angle, privacy zones, and a validation checklist for each newly mounted camera.

Supported devices

Seeed reCamera

Type

Compact Linux camera module

Purpose

Image capture and upload

Interface

POST /api/v1/device/recamera

Auth

8-character device key in JSON body

The reCamera is a compact Linux-based camera module from Seeed Studio. It captures images and uploads them to the Cognitive Companion backend for processing.

Integration flow:

1. reCamera runs its on-device YOLO11 model and produces a JSON payload containing the JPEG image and detection results
2. The payload is POSTed to /api/v1/device/recamera with the device key as the ?api_key= query parameter
3. The backend optionally rotates the image and applies a label filter before uploading to MinIO
4. The event aggregator batches the frame with others from the same sensor
5. When a batch is ready, matching rules are evaluated and pipelines execute

Payload format:

The reCamera posts a JSON object with this structure:

{
  "code": 0,
  "data": {
    "image": "<base64-encoded JPEG>",
    "labels": ["person"],
    "boxes": [[x1, y1, x2, y2, score, class_id]],
    "count": 287,
    "perf": [[model_id, preprocess_ms, inference_ms]],
    "resolution": [1280, 720]
  },
  "name": "invoke",
  "type": 1
}

data.labels lists the object classes detected by the YOLO11 model and can be used to filter which images are forwarded to the pipeline.

Device key configuration:

# In config/auth.yaml
device_keys:
  - key: "RCAM0001"
    name: "Kitchen reCamera"
    device_type: recamera
    sensor_id: recamera_kitchen
    permissions:
      - "device:recamera"

Per-camera options:

Configure per-camera behavior in config/settings.yaml under the cameras key, using the sensor_id as the key:

cameras:
  recamera_kitchen:
    rotate: 90           # clockwise rotation before storage (90, 180, 270)
    label_filter:
      labels: ["person"] # labels to match against payload.data.labels
      mode: "any"        # "any": at least one match; "all": every label must match

Option	Description
rotate	Rotates the JPEG clockwise before uploading to MinIO. Useful for cameras mounted at non-standard angles. Accepted values: 90, 180, 270. Omit to skip rotation.
label_filter.labels	List of YOLO11 label strings. Images are only forwarded when the detected labels satisfy the filter.
label_filter.mode	"any" (default): pass if at least one label matches. "all": pass only when every configured label is detected.

When a label filter is configured and the image does not match, the endpoint returns {"status": "filtered", "reason": "label_filter"} and the image is not saved or forwarded to the pipeline.

Placement tips:

• Mount at doorways to leverage motion direction detection (entering vs. leaving)
• Ensure adequate lighting for face recognition. ArcFace works best with even illumination.
• Consider the field of view. Wider angles capture more context but reduce face resolution.
• For multi-camera setups, assign each camera to a room for location tracking
• Use rotate when the camera must be mounted sideways or upside-down due to physical constraints

Seeed reTerminal

Type

Color e-ink display module

Chip

ESP32

Purpose

Notification display and button input

Image poll

GET /api/v1/image/active (device key auth)

Button input

POST /api/v1/device/reterminal

The reTerminal is an ESP32-based color e-ink display module from Seeed Studio. It serves as the household's notification display and physical interaction point.

E-ink display integration:

1. The reTerminal polls GET /api/v1/image/active at regular intervals
2. The backend identifies the device by its device key and returns its specific active image
3. When a notification is sent to the e-ink channel, the image updates on the next poll
4. Expired images automatically fall back to the default template

Button integration:

1. Physical button presses on the reTerminal are sent to POST /api/v1/device/reterminal
2. Button events can trigger rules (e.g., "request assistance" button)
3. Button presses can also acknowledge/dismiss active alerts

Configuration:

# In config/auth.yaml
device_keys:
  RTRM0001:
    sensor_id: hallway_display
    device_type: reterminal

Home Assistant Sensors

Type

Various smart home sensors

Purpose

Presence detection, light levels, audio playback

Interface

Home Assistant REST API (polled)

The backend polls Home Assistant entities at a configurable interval (default: 30 seconds). Supported sensor types:

Presence Sensors (PIR/mmWave)

Binary sensors that detect room occupancy. Used for:

• Room occupancy tracking: determine which rooms are occupied
• Person location inference: correlate presence sensor data with camera sightings to track people in rooms without cameras (e.g., bathrooms)
• Bathroom monitoring: configurable time limit triggers alerts for extended bathroom occupancy

Light Sensors

Illuminance sensors used for context-aware rules:

• Available as context data in pipeline steps
• Queryable via the get_light_level MCP tool
• Can inform vision analysis prompts (e.g., "analyze this low-light image")

Media Players

Smart speakers, tablets, and other audio devices used for:

• TTS announcements: speak notification messages aloud
• Home Assistant announce service: broadcast to multiple rooms

Network requirements

All devices must be on the same local network as the Cognitive Companion backend. There is no cloud relay; communication is direct.

Recommended network setup:

• Dedicated VLAN or subnet for IoT/camera devices
• Backend server on the same subnet (or with routing configured)
• Home Assistant accessible from the backend via its REST API
• MinIO accessible from the backend for media storage

Add new hardware

The system is designed to accommodate new edge devices. Any device that can make HTTP requests can integrate:

1. Define a device key in config/auth.yaml with the appropriate sensor ID and type
2. Create a sensor in the admin console matching the device
3. Implement the upload. POST images to /api/v1/device/recamera or a custom endpoint.
4. For display devices, poll GET /api/v1/image/active with the device key

For truly custom integrations, you can also create a new router endpoint in backend/routers/ following the adding a new API endpoint guide.