Spatial Awareness - Scrypted Plugin

Cross-camera object tracking for Scrypted NVR. Track people, vehicles, and animals as they move across your property.

What It Does

Instead of getting separate "person detected" alerts from each camera, get one coherent narrative:

> "Man in blue jacket entered via Driveway, walked to Front Door, left package, exited via Driveway" (2 min on property)

| Traditional Alerts | Spatial Awareness |
|-------------------|-------------------|
| Alert per camera | One alert per movement |
| "Person on Camera X" | "Person moving from X to Y" |
| No identity tracking | Same object tracked across cameras |
| Basic detection | Movement patterns, dwell time, unusual paths |

Quick Start

$3

``bash
npx scrypted install @blueharford/scrypted-spatial-awareness
`

$3

The fastest setup is Training Mode - walk your property while the system learns:

1. Open plugin settings → Click Training Mode
2. Tap Start Training on your phone
3. Walk between cameras naturally
4. System auto-detects you and records transit times
5. Tap End Training → Apply Results

Done! Your camera topology is configured.

$3

- Select notifiers (Pushover, email, etc.)
- Set loitering threshold (default: 3s)
- Set per-object cooldown (default: 30s)

Features

$3

- Cross-Camera Tracking - Correlate objects across cameras using timing, visual similarity, and spatial position
- Journey History - Complete path for each tracked object
- Entry/Exit Detection - Know when objects enter or leave your property
- Smart Alerts - Loitering thresholds and per-object cooldowns prevent spam

$3

- Floor Plan - Upload image or draw with built-in tools
- Drag & Drop - Place cameras, draw connections
- Polygon Zone Drawing - Draw custom zones (yards, driveways, patios, etc.)
- Live Tracking - Watch objects move in real-time

$3

- LLM Descriptions - "Woman with stroller" instead of just "Person"
- Auto-Learning - Transit times adjust based on observations
- Connection Suggestions - System suggests new camera paths
- Landmark Discovery - AI identifies landmarks from footage
- Auto-Topology Discovery - Vision LLM analyzes camera views to build topology

$3

- MQTT - Home Assistant integration
- REST API - Query tracked objects programmatically

Topology Configuration

The plugin uses topology data (landmarks, zones, connections) to generate meaningful alerts. Camera names are not used for location descriptions - only topology landmarks matter.

$3

For best alert quality, configure landmarks in the topology editor:

1. Entry/Exit Points - Mark where people enter/exit your property
- Examples: Driveway, Front Gate, Side Gate, Street
- Set isEntryPoint: true or isExitPoint: true

2. Access Points - Paths and walkways
- Examples: Front Walkway, Back Path, Garage Door
- Type: access

3. Zones - Areas of your property
- Examples: Front Yard, Back Yard, Side Yard, Patio
- Type: zone

4. Structures - Buildings and fixed features
- Examples: Garage, Shed, Front Porch, Deck
- Type: structure

5. Features - Other notable landmarks
- Examples: Mailbox, Pool, Garden, Trash Cans
- Type: feature

$3

Each camera should have visibleLandmarks configured - the landmarks visible in that camera's view:

`json
{
"cameras": [{
"deviceId": "abc123",
"name": "Front Camera",
"context": {
"visibleLandmarks": ["front-door", "driveway", "mailbox"]
}
}]
}
`

$3

`json
{
"landmarks": [
{ "id": "driveway", "name": "Driveway", "type": "access", "isEntryPoint": true },
{ "id": "front-door", "name": "Front Door", "type": "access" },
{ "id": "backyard", "name": "Back Yard", "type": "zone" },
{ "id": "garage", "name": "Garage", "type": "structure" }
]
}
`

With proper landmarks, alerts become rich and contextual:
- "Person arrived at the Driveway from Main Street"
- "Person moved from the Front Porch heading towards the Back Yard"
- "Person left the Garage towards Driveway after 2m on property - visited Driveway > Front Door > Garage"

Configuration

$3

| Setting | Default | Description |
|---------|---------|-------------|
| Correlation Window | 30s | Max time for cross-camera matching |
| Correlation Threshold | 0.35 | Min confidence for auto-correlation |
| Loitering Threshold | 3s | Time before triggering alerts |
| Minimum Detection Confidence | 0.5 | Ignore detections below this score |
| Per-Object Cooldown | 30s | Min time between alerts for same object |
| Notify on Alert Updates | Off | Send notifications when an existing alert updates |
| Alert Update Cooldown | 60s | Min time between update notifications |
| LLM Rate Limit | 5s | Min time between LLM API calls |

$3

| Alert | Description |
|-------|-------------|
| Property Entry | Object entered via entry point |
| Property Exit | Object exited via exit point |
| Movement | Object moved between cameras |
| Unusual Path | Object took unexpected route |
| Dwell Time | Object lingered >5 minutes |

API

Base URL: /endpoint/@blueharford/scrypted-spatial-awareness

| Endpoint | Method | Description |
|----------|--------|-------------|
| /api/tracked-objects | GET | List tracked objects |
| /api/journey/{id} | GET | Get object journey |
| /api/topology | GET/PUT | Camera topology |
| /api/alerts | GET | Recent alerts |
| /api/live-tracking | GET | Real-time object positions |
| /ui/editor | GET | Visual topology editor |
| /ui/training | GET | Training mode UI |

$3

| Endpoint | Method | Description |
|----------|--------|-------------|
| /api/training/start | POST | Start training session |
| /api/training/end | POST | End session, get results |
| /api/training/apply | POST | Apply results to topology |
| /api/training/status | GET | Current training status |

$3

| Endpoint | Method | Description |
|----------|--------|-------------|
| /api/discovery/scan | POST | Run full discovery scan |
| /api/discovery/status | GET | Current discovery status |
| /api/discovery/suggestions | GET | Pending suggestions |
| /api/discovery/camera/{id} | GET | Analyze single camera |

Auto-Topology Discovery

The plugin can automatically analyze camera views using a vision-capable LLM to discover landmarks, zones, and camera connections.

If a configured LLM provider does not support vision, it will be automatically excluded from discovery and shown in plugin settings under Excluded LLMs (No Vision).

$3

1. Capture Snapshots - System takes a picture from each camera
2. Scene Analysis - Vision LLM identifies landmarks, zones, and edges in each view
3. Cross-Camera Correlation - LLM correlates findings across cameras to identify shared landmarks and connections
4. Suggestions - Discoveries are presented as suggestions you can accept or reject

$3

Manual Scan:
1. Open the topology editor (/ui/editor)
2. Find the "Auto-Discovery" section in the sidebar
3. Click "Scan Now"
4. Review and accept/reject suggestions

Automatic Scan:
- Set Auto-Discovery Interval (hours) in plugin settings
- System will periodically scan and generate suggestions
- Set to 0 to disable automatic scanning

$3

| Setting | Default | Description |
|---------|---------|-------------|
| Auto-Discovery Interval | 0 (disabled) | Hours between automatic scans (0 = disabled) |
| Min Landmark Confidence | 0.6 | Minimum confidence for landmark suggestions |
| Min Connection Confidence | 0.5 | Minimum confidence for connection suggestions |
| Auto-Accept Threshold | 0.85 | Auto-accept suggestions above this confidence |

> Rate Limiting Note: If you set the interval to less than 1 hour, a warning will appear in the discovery status. Frequent scans can consume significant LLM API quota and may be rate-limited by your provider.

$3

- Vision-capable LLM - Install @scrypted/llm with a vision model (OpenAI GPT-4V, Claude, etc.)
- Camera access - Plugin needs camera.takePicture() capability

$3

- Landmarks: Doors, gates, mailbox, garage, structures, fences
- Zones: Front yard, driveway, patio, street, walkways
- Connections: Suggested camera paths with transit time estimates
- Edges: What's visible at frame boundaries (for correlation)

Zone Drawing

The visual editor includes a polygon zone drawing tool for marking areas on your floor plan.

$3

1. Click the Draw Zone button in the toolbar (green)
2. Enter a zone name and select the type (yard, driveway, patio, etc.)
3. Click Start Drawing
4. Click on the canvas to add polygon points
5. Double-click or press Enter to finish the zone
6. Press Escape to cancel, Backspace to undo last point

$3

| Type | Color | Description |
|------|-------|-------------|
| Yard | Green | Front yard, backyard, side yard |
| Driveway | Gray | Driveway, parking area |
| Street | Dark Gray | Street, sidewalk |
| Patio | Orange | Patio, deck |
| Walkway | Brown | Walkways, paths |
| Parking | Light Gray | Parking lot, parking space |
| Garden | Light Green | Garden, landscaped area |
| Pool | Blue | Pool area |
| Garage | Medium Gray | Garage area |
| Entrance | Pink | Entry areas |
| Custom | Purple | Custom zone type |

$3

- Click on a zone to select it and edit its properties
- Zones are color-coded by type for easy identification
- Zones help provide context for object movement descriptions
- Auto-Discovery can suggest zones based on camera analysis

MQTT Topics

Base: scrypted/spatial-awareness

| Topic | Description |
|-------|-------------|
| /occupancy/state | ON/OFF property occupancy |
| /count/state | Active object count |
| /person_count/state | People on property |
| /alerts | Alert events |
| /events/entry | Entry events |
| /events/exit | Exit events |

How Correlation Works

When an object is detected on a new camera, the system scores potential matches:

| Factor | Weight | Description |
|--------|--------|-------------|
| Timing | 30% | Transit time within expected range |
| Visual | 35% | Embedding similarity (if available) |
| Spatial | 25% | Exit zone → Entry zone coherence |
| Class | 10% | Object type match |

Objects are correlated if total score exceeds threshold (default: 0.6).

Requirements

- Scrypted with NVR plugin
- Cameras with object detection (NVR, OpenVINO, CoreML, etc.)
- Optional: LLM plugin for rich descriptions
- Optional: MQTT broker for Home Assistant

Development

`bash
npm install
npm run build
npm run scrypted-deploy
``

License

Apache-2.0

Author

Joshua Seidel (@blueharford)