👁️ FirstBreath Vision

The Computer Vision and AI subsystem for the FirstBreath platform. This system hosts individual microservices responsible for RTSP stream management, high-performance batch inference (YOLOv11 + TensorRT), and motion metrics processing.

🏗 Architecture

The system is designed as a pipeline of independent microservices communicating via Redis:

1. 🎥 Camera Manager (services/camera-manager):

   • Connects to RTSP streams using OpenCV (HW accelerated).
   • Publishes raw frames to Redis via a binary protocol.
   • Manages camera lifecycles and tracking scripts.

2. 🧠 Batch Inference (services/batch-inference):

   • GPU Optimized: Uses NVIDIA TensorRT (FP32) / CUDA.
   • Batch Processing: Aggregates frames from multiple cameras into batches for efficient GPU usage.
   • Auto-Export: Automatically converts .pt models to .engine optimized for the host hardware.
   • Publishes detection results back to Redis.

3. ⚙️ Redis Worker (services/redis-worker):

   • Consumes detection results and calculates motion metrics (velocity, agitation).
   • Inserts metrics into MySQL.
   • Triggers alerts (e.g., "STRESSED" state) via Redis Pub/Sub.

🚀 Getting Started

Prerequisites

• Docker & Docker Compose
• NVIDIA GPU (Optional, but recommended for inference performance)
• NVIDIA Container Toolkit (for GPU support)

Installation

1. Clone the repository:

   git clone git@github.com:FirstBreath/firstbreath-vision.git
   cd firstbreath-vision

2. Run with Docker Compose:

   docker-compose up --build

   This launches:

   • camera-manager
   • batch-inference (Compiles TensorRT engine on first run ~2-5 mins)
   • redis-worker

📊 Observability

All services expose Prometheus Metrics (scraped via monitor-net).

• Camera Manager: Active scripts, CPU/RAM per camera.
• Batch Inference: Inference Latency (P95), Batch Size, FPS.
• Redis Worker: DB Insert Rate, Queue Lag, Alert counts.