Pipeline
- RTSP / ONVIF / file ingest via OpenCV
- VLM inference with parallel box decoding
- Kalman tracking + geofence breach logic
- Live overlays & alerts over WebSocket
Massive Scale Capabilities
Engineered for border security command centers, fusing cutting-edge AI with ultra-low latency streaming protocols.
Hardware-Agnostic Edge
Integrates instantly with existing IP (RTSP), ONVIF, and drone links without requiring dedicated proprietary smart cameras. Works on standard mini-PCs.
Zero-Shot Grounding
Powered by the LocateAnything-3B VLM. Execute natural language queries like "unattended cargo" or "armed trespasser" without retraining models.
Parallel Box Decoding
Bypasses sequential text generation bottlenecks. Our PBD engine extracts bounding boxes 2.5× faster, achieving 12.7+ targets per second on edge hardware.
Interactive Geofencing
Draw virtual tripwires and exclusion zones directly on the video feed. Real-time vector math triggers alerts instantly upon polygon intersection.
Persistent Tracking
ByteTrack & Kalman filtering ensure targets maintain their identities even through dense dust, foliage, or temporary camera occlusions.
Unified C2 Dashboard
A single React/Vite command center. Replaces disparate pipelines (YOLO, OCR, FaceRec) with one unified threat timeline and telemetry HUD.
Pitch Presentation
Review the end-to-end architecture, technical methodologies, and SIH impact statement.
📄 Slide Deck View
Live Command Interface
Interact with the SENTINEL-3B tactical frontend. Explore geofencing, threat logs, and telemetry.
🔴 Live Terminal Preview
Field Demonstrations
Live captures and decoding demos from the tactical operations center.
Our Team
Meet the people behind this project.
×
![]()
Technical Documentation
Essential architecture, deployment, and integration reference — condensed for quick review.
Quick deploy
- Clone repo & pull model weights
uvicorn main:app --port 8000in backendnpm run devin frontend- Windows: use
start_project.bat
Performance profiles
| Profile | Latency | Resolution | VRAM |
|---|---|---|---|
| Fast | ~14–18 ms | 640×640 | 7.8 GB |
| Hybrid (default) | ~22–32 ms | 1024×1024 | 11.7 GB |
| Slow / forensic | ~60–85 ms | 4K | 18.2 GB |
API
Stream: ws://host:8000/ws/stream
JSON payloads include camera ID, prompts, geofence polygons, frame tracks, and severity alerts.
SIH PS 26187
- Human & vehicle localization
- Perimeter violation detection
- Night / thermal-ready ingest
- Unified C2 integration
NVIDIA GPU (RTX 3080+)