VioMobileNet - Real-time AI Violence Monitoring System

A high-performance system designed to analyze violent behavior from RTSP camera streams and video uploads. Powered by MoViNet, optimized for ultra-high throughput and minimal latency.

Key Features

• Dual Inference Pipeline:

• Operates two parallel inference processes: HIGH_SERVER and LOW_MED_SERVER.

• Prioritized Queuing: High-risk (high) cameras are isolated in a dedicated stream to ensure the highest possible monitoring frequency.

• Instant Score Response:

• Boost Logic: Uses an exponential function to spike scores when the threshold exceeds 0.2, triggering alarms the moment violence begins.

• Quick Drop Logic: Force-resets scores to 0 as soon as the scene stabilizes, eliminating the "accumulation lag" common in AI temporal analysis.

• LIFO (Last-In-First-Out) & Interleaving:

• Frame Interleaving: Distributes frame processing across cameras to prevent any single stream from hogging the queue.

• Aggressive LIFO: Automatically purges stale frames to ensure system-wide latency stays under 100ms.

• Non-blocking Background Upload:

• Evidence images are uploaded to MinIO via dedicated background threads. This prevents the primary AI inference loop from stuttering during data transmission.

• Smart Redis Caching:

• Dynamic TTL (Time-to-Live) management (up to 60 seconds) for violent events, ensuring API/CURL queries receive stable, reliable data.

---

Directory Structure

VioMobileNet/
├── app.py              # API Gateway: Routes cameras to the correct queue (High vs. Low/Med).
├── core.py             # AI Core: GPU Memory Growth and TensorFlow thread management.
├── rtsp_worker.py      # The Brain: Parallel inference logic and multi-cadence streaming.
├── worker.py           # Celery Worker: Offline video processing for file uploads.
├── requirements.txt    # Dependencies: TensorFlow 2.15, OpenCV, Redis, MinIO...
└── docker-compose.yml  # Orchestration: Links API, Workers, Redis, and MinIO.

---

Operational Workflow

1. Ingestion: The Streamer reads RTSP feeds, resizes frames to 256x256, and pushes them into risk-based queues using interleaving.
2. Dual Inference:

• HIGH_SERVER processes priority cameras at high FPS.
• LOW_MED_SERVER handles standard monitoring to conserve resources.

3. Score Processing: Applies Max Pooling and exponential scaling to make the model "sensory-sharp" to aggressive actions.
4. Result Distribution:

• Status data is cached in Redis with flexible TTL.
• Visual evidence is offloaded via Threaded Upload to MinIO.

---

Getting Started

1. System Requirements

• GPU: NVIDIA RTX 30 Series or higher (CUDA 12.x support).
• RAM: 16GB Minimum (System consumes ~3GB for two inference pipelines).
• Environment: Docker & NVIDIA Container Toolkit installed.

2. Deployment

# Spin up the entire stack
docker-compose up -d --build

# Monitor logs for parallel inference streams
docker logs -f viomobilenet_api

3. UI Controls (Monitoring Client)

Toggle between camera feeds instantly using keyboard shortcuts:

• Keys [1-8]: Switch between Camera 01 and Camera 08.
• Key [Q]: Exit the monitoring system.

---

Performance Benchmarks (Tested on RTX 4060 Ti)

• AI Inference Speed: ~30ms - 50ms per frame.
• Capacity: Supports 8-10 simultaneous streams with high stability.
• Pipeline Latency: Consistently maintained < 100ms via LIFO and Background Uploading.