Real-Time Object Detection Pipeline

A real-time video analytics system built to self-host on a Raspberry Pi 5 (8GB). Runs YOLOv8 inference in C++ via ONNX Runtime, streams detection events through Apache Kafka, processes data with Python batch aggregation, and serves a live dashboard via React.

Architecture

The system is split into five services that communicate through Kafka and shared data stores:

Service	Language	Description
capture-inference	C++	Captures webcam frames via OpenCV, runs YOLOv8 inference via ONNX Runtime, publishes detection events to Kafka, serves annotated frames via MJPEG stream
stream-processor	C++	Kafka consumer that aggregates detections in real-time using sliding window, writes hot data to Redis
batch-processor	Python	Consumes detection events from Kafka, writes to PostgreSQL, runs scheduled hourly aggregation and cleanup
api	Python (FastAPI)	REST API serving real-time data from Redis and historical data from PostgreSQL
frontend	React/TypeScript	Live video feed via MJPEG, real-time detection charts, historical data queries with date range picker

Data Flow

capture-inference produces two independent outputs:

• Annotated frames are served directly as an MJPEG stream — the frontend displays these via a simple <img> tag with no WebSocket or API proxy needed.
• Detection events (small JSON: timestamp, class, confidence, bbox coordinates) flow through Kafka to two consumers:
  • stream-processor aggregates in real-time (sliding window counts, recent events) and writes to Redis
  • batch-processor writes every event to PostgreSQL and runs scheduled hourly aggregation

The API reads from both Redis (real-time) and PostgreSQL (historical) to serve the frontend. These paths are independent — if Kafka lags, video still plays. If the video stream drops, charts still update.

Infrastructure

The full stack runs in Docker Compose with Apache Kafka for event streaming, PostgreSQL for historical storage, and Redis for real-time caching. Kafka and PostgreSQL have health checks, and the services that depend on them wait until they report healthy before starting. Every container restarts automatically if it exits.

Tech Stack

Languages: C++17, Python, TypeScript ML: YOLOv8, ONNX Runtime Streaming: Apache Kafka (librdkafka) Backend: FastAPI Frontend: React, TypeScript Databases: PostgreSQL, Redis Infrastructure: Docker, Docker Compose Build: CMake, Google Test

Links

• GitHub