‹ All Projects
Project Thread

LibraryFlow — Enterprise Multi-Camera Edge AI & Spatial Analytics Platform

🐙 View GitHub

Overview

A distributed "Vision-as-a-Service" pipeline that transforms CCTV feeds into a Security Operations Center (SOC). It tracks facility occupancy and performs advanced human re-identification.

Tech Stack

Python (Ultralytics, Supervision)DeepFace (RetinaFace backend)PostgreSQL (Relational Data Arch)Streamlit (BI Dashboard)Docker

Features

  • Dual-Lens Entrance Node: Simultaneous facial recognition and clothing attribute extraction (HSV Colorimetry).
  • Mathematical Re-ID: Cross-camera identity matching using color vector histograms.
  • Temporal State Filtering: 3-second memory threshold to eliminate AI flickering.
  • Security Synonym Engine: NLP-powered search allowing staff to find suspects by description.

Key Decisions

01

Prioritized HSV Colorimetry over RGB for Re-ID to ensure stability under the library's varying fluorescent and natural lighting conditions.

02

Implemented a 3-second Temporal Filter to prevent "Ghost Identities" caused by temporary camera occlusions in high-traffic areas.

Learnings

→Developed a custom spatial mapping logic to translate 3D camera coordinates into 2D floor plan heatmaps.
→Learned to manage high-volume video stream ingestion by decoupling the YOLO detection frequency from the UI rendering rate.
EmailGitHubLinkedInResume