How danikchernov/sentinel-vision works
Sentinel Vision Console is a local-first video intelligence application. A user configures camera, network-stream, or file sources through the dashboard/API; the backend starts per-source workers that ingest frames, detect and track objects, maintain persistent identities, optionally refine labels, and publish previews and summaries. Source definitions persist in a JSON file at runtime. The graph includes the documented optional QLoRA labeling workflow; the excerpts do not show the frontend’s API calls or the training scripts’ internal handoff, so those links are omitted.
Main entry points: source_manager.py Source manager · video.py Video source worker · main.py FastAPI application · tracker.py DeepSORT tracker
Components 18
Dashboard and API
- React dashboard
frontend/src/main.js - FastAPI application
backend/app/main.py - API and source schemas
backend/app/schemas.py
Source management
- Source manager
backend/app/source_manager.py - Runtime settings
backend/app/config.py - Persisted sources
- Runtime readiness
backend/app/runtime.py
Video perception
- Video source worker
backend/app/video.py - YOLO detector
backend/app/detector.py - DeepSORT tracker
backend/app/tracker.py - Shared inference stack
backend/app/video.py
Identity and labeling
- Appearance encoder
backend/app/embeddings.py - Identity registry
backend/app/identity.py - Vision labeler
backend/app/vlm.py
Labeler training
- Instruction dataset builder
train/build_instruction_dataset.py - QLoRA trainer
train/train_vlm_qlora.py
Other components
- Operator
- Video sources
Connections 21
- Operator → React dashboard: uses
README.md - FastAPI application → React dashboard: serves
backend/app/main.py - FastAPI application → Source manager: routes source operations
backend/app/main.py - FastAPI application → Runtime readiness: reports readiness
backend/app/main.py - FastAPI application → API and source schemas: validates requests
backend/app/main.py - Source manager → API and source schemas: uses source records
backend/app/source_manager.py - Source manager → Runtime settings: uses settings
backend/app/source_manager.py - Source manager → Persisted sources: loads and saves
backend/app/source_manager.py - Source manager → Video source worker: starts workers
backend/app/source_manager.py - Source manager → Shared inference stack: constructs stack
backend/app/source_manager.py - Source manager → Identity registry: creates registry
backend/app/source_manager.py - Source manager → Runtime readiness: checks dependencies
backend/app/source_manager.py - Video sources → Video source worker: provides video
README.md - Video source worker → Shared inference stack: processes frames
backend/app/video.py - Video source worker → DeepSORT tracker: updates tracks
backend/app/video.py - Video source worker → Appearance encoder: encodes crops
backend/app/video.py - Video source worker → Identity registry: registers observations
backend/app/video.py - Video source worker → Vision labeler: labels crops
backend/app/video.py - Shared inference stack → YOLO detector: runs detection
backend/app/video.py - DeepSORT tracker → YOLO detector: consumes detections
backend/app/tracker.py - QLoRA trainer → Vision labeler: fine-tunes model
README.md