VisionHub: Automatic Car Number Plate, Type, Make and Color Detection and Recognition System

VisionHub: Realtime Monitoring System

VisionHub is an intelligent monitoring system that addresses critical challenges in urban security and surveillance through innovative integration of cutting-edge AI technologies. Developed through collaboration between the Lebanese American University’s School of Engineering and InMind.ai, with support from the LAU Industrial Hub and partial funding from the European Bank for Reconstruction and Development (EBRD), VisionHub represents a significant advancement in real-time video analytics.

Project Motivation

In today’s rapidly urbanizing world, traditional surveillance methods struggle to process the overwhelming volume of visual data generated by omnipresent cameras. This creates dangerous gaps in threat detection and response capabilities. VisionHub emerged as a direct response to this challenge, seeking to transform passive monitoring into proactive intelligence through AI-powered analysis.

Technical Approach and Implementation Highlights

VisionHub delivers a robust end-to-end surveillance analytics platform designed with performance, scalability, and usability in mind. Built on NVIDIA’s DeepStream SDK and deployed across a hybrid infrastructure of NVIDIA Jetson edge devices and GPU-accelerated servers, VisionHub efficiently handles multiple concurrent video streams through a modular four-layer architecture:

Source Layer – Ingests live video feeds from CCTV cameras or pre-recorded footage.
Data Processing Layer – Manages real-time stream processing and inference through two services:
- The Stream Service, built with DeepStream, supports parallel and scalable video ingestion.
- The Inference Service, powered by Triton Inference Server and TensorRT, executes optimized models for object detection, tracking, and scene understanding.
Data Access Layer – Utilizes Apache Kafka for logging model outputs, events, and metadata in a standardized and efficient manner, ensuring scalable communication and persistent storage.
Information Retrieval Engine – An LLM-powered interface that enables users to query surveillance data in natural language (e.g., “Find the black car with this license plate between 2–3 PM at Entrance A”), dramatically enhancing accessibility.

Key implementation strengths include:

GPU-accelerated parallel processing for real-time, high-throughput inference.
Edge–cloud hybrid deployment for flexible scalability: lightweight Jetson devices at distributed camera sites, and high-performance servers for centralized processing.
Modular pipeline design enabling independent component upgrades and customizable workflows.
Privacy-first engineering with adherence to security and data compliance standards.
Intelligent analytics capabilities, including cross-camera vehicle tracking, attribute-based person recognition, video upscaling, unattended object detection, and smart alerting.

Together, these components form a scalable and intelligent video analytics platform. VisionHub’s natural language query interface exemplifies how cutting-edge AI can make surveillance data both actionable and accessible, transforming the traditional paradigm of passive monitoring into an interactive intelligence system.

Technical Innovations

VisionHub integrates several state-of-the-art technologies to deliver comprehensive monitoring capabilities:

Multi-stream processing using NVIDIA DeepStream’s optimized pipeline architecture
Vehicle analytics including make/model recognition, color classification, and license plate reading
People attribute detection that identifies characteristics such as clothing type and color
Suspicious package detection through object association and monitoring
Video enhancement using AI-powered upscaling to improve detection accuracy
Natural language querying through LLM integration (GPT-4o-mini) for intuitive data exploration

The system employs optimized computer vision models, including YOLOv8/v9 variants for detection tasks and specialized models for license plate recognition, upscaling, and attribute classification. Model optimization through ONNX and TensorRT ensures efficient inference on both edge and server deployments.

Project Details

Student(s): Ali Zayat, Amir Amine
Advisor(s): Dr. Joe Tekli
Year: 2024-2025

[photo]