Real-time Video Analysis Inference Server Development

Developed an inference server running object detection and pose estimation with trained YOLO models for a public security agency's gait pattern analysis system. Targeted 30 FPS, decoupling ingestion from inference via Kafka and storing results in MongoDB for query. The project stopped at the PoC stage.

Architecture

SEO Inc. Public Sector

2022.09 - 2023.01

Public Security Agency

Tech Stack

Python YOLO Kafka Docker OpenCV MongoDB

Highlights

Object detection and pose estimation serving with YOLO v5/v7
Pipeline decoupling ingestion from inference via Kafka (30 FPS target)
GPU Docker container-based serving architecture
MongoDB result storage and query REST API

Key Tasks

Inference Engine Implementation

YOLOv5, YOLOv7 model serving environment setup
Object Detection + Human Skeleton inference
GPU optimization and batch processing

Distributed Processing Pipeline

Kafka Producer/Consumer implementation
Message queue-based load balancing
30 FPS target

Serving Environment

GPU Docker container-based serving architecture

Result Storage & API

MongoDB schema design
Analysis result storage pipeline
REST API development

삶을 기록하는 두 축

Architecture