Chronos: Real-time Video Analytics Pipeline

Chronos is an event-driven video ingestion and analytics engine. It slices live video streams, processes them with Computer Vision (YOLO) in real-time, stores semantic embeddings in a Vector Database (Qdrant), and allows for natural language search (e.g., "Find me a clip of a bird").

Architecture

Ingestion (Go) -> Streaming (Kafka) -> Inference (Python/YOLO) -> Storage (Qdrant)

1. Stream Slicer (Go): Downloads video streams, segments them into 5-second chunks, and publishes metadata to Kafka.
2. Message Broker (Redpanda/Kafka): Decouples ingestion from processing to handle high-throughput backpressure.
3. ML Worker (Python): Consumes messages, runs YOLOv8 object detection, and generates text descriptions (e.g., "A video containing: person, car").
4. Vector Store (Qdrant): Stores the semantic embeddings of the descriptions.
5. Search API: Allows querying the video library using natural language.

Tech Stack

• Language: Go (Ingestion), Python 3.13 (ML/Inference)
• AI Models: Ultralytics YOLOv8 (Vision), SentenceTransformers (Embeddings)
• Infrastructure: Docker, Redpanda (Kafka compatible), Qdrant (Vector DB)
• Libraries: confluent-kafka, requests (Direct API implementation)