@blueharford/scrypted-spatial-awareness 0.4.8-beta.1 → 0.5.0-beta

package/README.md

@@ -51,6 +51,7 @@ Done! Your camera topology is configured.
 ### Visual Editor
 - **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
 
 ### AI Features (optional)
@@ -58,6 +59,7 @@ Done! Your camera topology is configured.
 - **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
 
 ### Integrations
 - **MQTT** - Home Assistant integration
@@ -170,6 +172,98 @@ Base URL: `/endpoint/@blueharford/scrypted-spatial-awareness`
 | `/api/training/apply` | POST | Apply results to topology |
 | `/api/training/status` | GET | Current training status |
 
+### Discovery API
+
+| 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.
+
+### How It Works
+
+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
+
+### Using Discovery
+
+**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
+
+### Discovery Settings
+
+| 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.
+
+### Requirements
+
+- **Vision-capable LLM** - Install @scrypted/llm with a vision model (OpenAI GPT-4V, Claude, etc.)
+- **Camera access** - Plugin needs camera.takePicture() capability
+
+### What Gets Discovered
+
+- **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.
+
+### How to Draw Zones
+
+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
+
+### Zone Types
+
+| 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 |
+
+### Using Zones
+
+- 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`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blueharford/scrypted-spatial-awareness",
-  "version": "0.4.8-beta.1",
+  "version": "0.5.0-beta",
   "description": "Cross-camera object tracking for Scrypted NVR with spatial awareness",
   "author": "Joshua Seidel <blueharford>",
   "license": "Apache-2.0",

package/src/core/spatial-reasoning.ts

@@ -10,6 +10,7 @@ import sdk, {
   Camera,
   MediaObject,
   ScryptedDevice,
+  ScryptedMimeTypes,
 } from '@scrypted/sdk';
 import {
   CameraTopology,
@@ -26,7 +27,7 @@ import {
 } from '../models/topology';
 import { TrackedObject, ObjectSighting } from '../models/tracked-object';
 
-const { systemManager } = sdk;
+const { systemManager, mediaManager } = sdk;
 
 /** Configuration for the spatial reasoning engine */
 export interface SpatialReasoningConfig {
@@ -68,6 +69,29 @@ interface ChatCompletionDevice extends ScryptedDevice {
   streamChatCompletion?(params: any): AsyncGenerator<any>;
 }
 
+/**
+ * Convert a MediaObject to a base64 data URL for vision LLM consumption
+ * @param mediaObject - MediaObject from camera.takePicture()
+ * @returns Base64 data URL (data:image/jpeg;base64,...) or null if conversion fails
+ */
+export async function mediaObjectToBase64(mediaObject: MediaObject): Promise<string | null> {
+  try {
+    // Convert MediaObject to Buffer using mediaManager
+    const buffer = await mediaManager.convertMediaObjectToBuffer(mediaObject, ScryptedMimeTypes.Image);
+
+    // Convert buffer to base64
+    const base64 = buffer.toString('base64');
+
+    // Determine MIME type - default to JPEG for camera images
+    const mimeType = mediaObject.mimeType?.split(';')[0] || 'image/jpeg';
+
+    return `data:${mimeType};base64,${base64}`;
+  } catch (e) {
+    console.warn('Failed to convert MediaObject to base64:', e);
+    return null;
+  }
+}
+
 export class SpatialReasoningEngine {
   private config: SpatialReasoningConfig;
   private console: Console;
@@ -712,7 +736,7 @@ export class SpatialReasoningEngine {
     return connection.name || undefined;
   }
 
-  /** Get LLM-enhanced description using ChatCompletion interface */
+  /** Get LLM-enhanced description using ChatCompletion interface with vision support */
   private async getLlmEnhancedDescription(
     tracked: TrackedObject,
     fromCamera: CameraNode,
@@ -726,6 +750,9 @@ export class SpatialReasoningEngine {
     if (!llm || !llm.getChatCompletion) return null;
 
     try {
+      // Convert image to base64 for vision LLM
+      const imageBase64 = await mediaObjectToBase64(mediaObject);
+
       // Retrieve relevant context for RAG
       const relevantChunks = this.retrieveRelevantContext(
         fromCamera.deviceId,
@@ -746,12 +773,25 @@ export class SpatialReasoningEngine {
         ragContext
       );
 
+      // Build message content - use multimodal format if we have an image
+      let messageContent: any;
+      if (imageBase64) {
+        // Vision-capable multimodal message format (OpenAI compatible)
+        messageContent = [
+          { type: 'text', text: prompt },
+          { type: 'image_url', image_url: { url: imageBase64 } },
+        ];
+      } else {
+        // Fallback to text-only if image conversion failed
+        messageContent = prompt;
+      }
+
       // Call LLM using ChatCompletion interface
       const result = await llm.getChatCompletion({
         messages: [
           {
             role: 'user',
-            content: prompt,
+            content: messageContent,
           },
         ],
         max_tokens: 150,
@@ -809,7 +849,7 @@ Examples of good descriptions:
 Generate ONLY the description, nothing else:`;
   }
 
-  /** Suggest a new landmark based on AI analysis using ChatCompletion */
+  /** Suggest a new landmark based on AI analysis using ChatCompletion with vision */
   async suggestLandmark(
     cameraId: string,
     mediaObject: MediaObject,
@@ -822,6 +862,9 @@ Generate ONLY the description, nothing else:`;
     if (!llm || !llm.getChatCompletion) return null;
 
     try {
+      // Convert image to base64 for vision LLM
+      const imageBase64 = await mediaObjectToBase64(mediaObject);
+
       const prompt = `Analyze this security camera image. A ${objectClass} was detected.
 
 Looking at the surroundings and environment, identify any notable landmarks or features visible that could help describe this location. Consider:
@@ -835,12 +878,25 @@ If you can identify a clear landmark feature, respond with ONLY a JSON object:
 
 If no clear landmark is identifiable, respond with: {"name": null}`;
 
+      // Build message content - use multimodal format if we have an image
+      let messageContent: any;
+      if (imageBase64) {
+        // Vision-capable multimodal message format (OpenAI compatible)
+        messageContent = [
+          { type: 'text', text: prompt },
+          { type: 'image_url', image_url: { url: imageBase64 } },
+        ];
+      } else {
+        // Fallback to text-only if image conversion failed
+        messageContent = prompt;
+      }
+
       // Call LLM using ChatCompletion interface
       const result = await llm.getChatCompletion({
         messages: [
           {
             role: 'user',
-            content: prompt,
+            content: messageContent,
           },
         ],
         max_tokens: 100,