Condor AI Detection

The Notification project includes functions that poll the Condor AI detection API for smoke and fire events, persist structured detection data, and broadcast updates to the frontend in real time.

Overview

The Condor integration runs on a 1-minute timer and processes detection events through the following pipeline:

1. Poll — PollCondorApi requests recent smoke/fire detections from the Condor API
2. Queue — Events are grouped by camera and queued to the ai-detection-processor queue
3. Process — ProcessCameraEventsDequeue picks up queued events and runs the processing pipeline
4. Persist — Events are stored in both Cosmos DB (for real-time state) and SQL Server (for historical queries)
5. Broadcast — The latest detection is sent to connected frontend clients via SignalR

Detection Pipeline

Camera Matching

Each Condor event includes a camera_id that may correspond to either a device's DeviceId or a stream's StreamId. The function attempts both matches against the camera configuration stored in Cosmos DB.

Events are filtered by:

• Environment — Camera must be enabled in the current deployment environment (dev, stg, prd)
• Public visibility — Camera must be marked as IsAlertCaPublic

Detection Types

Type	Identifier	Description
60 Detection	Output image path starts with /app/output/detections/	Standard single-image detection
Pano Detection	Output image path starts with detections/panoramas/	Panoramic multi-frame detection

COCO Data Persistence

Each Condor detection event may include structured COCO-format detection and annotation data. After the CondorCameraEvent is upserted to SQL Server, its generated Id is used as the foreign key for bulk-inserting the COCO data.

COCODetection

Stores bounding-box detections with confidence scores:

Column	Type	Source
CondorCameraEventId	INT (FK)	From upserted CondorCameraEvent.Id
DetectionId	INT	CocoDetectionDTO.ImageId
CategoryId	INT	Detection category (smoke, fire)
Score	DECIMAL(3,2)	Confidence score (0.00–1.00)
BBox	NVARCHAR(255)	Bounding box coordinates (comma-separated)
Lat, Lon	NVARCHAR(25)	Optional geolocation (nullable, for future use)
EstimatedAzimuth	DECIMAL(10,2)	Optional azimuth estimate (nullable, for future use)

COCOAnnotation

Stores richer annotation data including segmentation masks:

Column	Type	Source
CondorCameraEventId	INT (FK)	From upserted CondorCameraEvent.Id
AnnotationId	INT	CocoAnnotationDTO.Id
CategoryId	INT	Annotation category
Score	DECIMAL(3,2)	Confidence score
BBox	NVARCHAR(255)	Bounding box coordinates
Segmentation	NVARCHAR(MAX)	Segmentation mask data (semicolon-separated)
IsCrowd	INT	Whether the annotation represents a crowd region
Area	BIGINT	Area of the annotation in pixels

Bulk Insert Pattern

Both detection and annotation data are inserted using table-valued parameters (TVPs) for performance. The stored procedures COCODetectionBulkInsert and COCOAnnotationBulkInsert delete existing rows for the event before inserting, ensuring idempotent upsert behavior.

Cosmos DB State

Active detection events are tracked in Cosmos DB as CondorEventEntity documents:

• TTL: 3 hours (auto-expire after detection activity stops)
• Event IDs: Accumulates all Condor event IDs associated with the camera
• Categories: Tracks all detected categories (smoke, fire) across updates

Camera Detection Flags

After processing, the camera's operational document is patched with detection flags:

Flag	Description
HasAiDetectionHit	Set to true on any AI detection
HasRecent60DetectionHit	Set to true for standard 60-second detections
HasRecentPanoDetectionHit	Set to true for panoramic detections

These flags are reset by the CleanupSmokeEvents function when no active events remain.

SignalR Broadcasting

After processing, the CondorCameraEvent (including PTZ values) is broadcast to the frontend via SignalR:

Target: "BroadcastLatestAiDetectionToGroup"
Group:  {cameraId}
Data:   CondorCameraEvent (with Pan, Tilt, Zoom, image URLs, category, detection type)