Camera Forensics enables investigators to query large fleets of camera streams using natural language and object filters, then collect frames and case artifacts.
Capabilities:
- Natural language search ("red car turning left") over indexed footage
- Object/person/plate detection with temporal tracking
- Cross‑camera correlation and geo‑spatial filtering
- Frame capture, evidence timelines, and chain‑of‑custody export
- Alerts on re‑identification and watchlists
Architecture:
- Edge: RTSP ingestion → FFmpeg → frames → Kafka
- AI: YOLO/Detectron + ReID embeddings + action recognition
- Index: Vector DB (FAISS/PGVector), metadata in Postgres
- Orchestration: Kafka + workers, object storage in S3
- UI: Next.js dashboard for search, review, and export
Tech stack: Python, PyTorch, OpenCV, FFmpeg, Kafka, PostgreSQL, PGVector, Next.js, S3.