TriNetra: Real-Time Violence Detection System
A CCTV monitoring system that watches live video streams and raises an alert the moment physical violence is detected, with a mobile app for security staff.
Overview
CCTV operators cannot watch every camera at once. TriNetra turns an ordinary RTSP camera feed into an active monitoring system: frames are analysed by a hybrid deep learning model, and anything that looks like physical violence is pushed to a mobile client within seconds.
How it works
- Stream ingestion. Cameras connect over RTSP and are pulled with FFmpeg, so no extra hardware is needed on site.
- Frame buffering. Extracted frames are queued in Redis, which decouples video ingestion from inference and keeps the pipeline responsive under load.
- Spatial features. MobileNetV2 encodes each frame into a compact feature vector.
- Temporal modelling. A Bi-LSTM reads the sequence of frame features in both directions and classifies the clip as violent or non-violent.
- Alerting. Detections are stored in PostgreSQL and delivered to the React Native app so staff can verify the event.
Architecture
The system is split into four independent microservices: ingestor, inference worker, API and notification service. Each one is packaged as a Docker container, so inference can be scaled horizontally without touching the camera-facing side.
Tech stack
- ML: TensorFlow, MobileNetV2 (transfer learning), Bi-LSTM
- Backend: Python, FastAPI, Redis queues, PostgreSQL
- Video: RTSP, FFmpeg
- Client: React Native
- Infrastructure: Docker
Highlights
- Real-time detection instead of after-the-fact review of recordings.
- Works with existing IP cameras over RTSP.
- Queue-based design absorbs bursts of frames without dropping alerts.
TriNetra: Real-Time Violence Detection System