Project Overview
EdgeGrid is a secure, entirely offline, on-premise RTSP camera surveillance system built for a confidential military installation in Bangladesh. Running on an NVIDIA Jetson Orin Nano, it pulls feeds from Hikvision IP cameras over an isolated LAN and uses custom YOLO computer vision models to detect and classify tactical threats in real time - distinguishing allies from armed civilians, triggering local alarms, and autonomously deploying a UAV for high-priority incidents. All video processing, recording, and incident logging happens entirely on the edge. No cloud. No server room. No internet required.
Challenges & What TDI Built
Remote Terrain
Critical zones had to be monitored continuously without dependable internet access.
Limited Footprint
The site needed edge intelligence without a cloud pipeline or traditional server room.
Autonomous Response
TDI paired local AI detection with alarms, recording, and UAV dispatch for immediate action.
The Challenges
Manual perimeter monitoring across rugged terrain - slow, error-prone, and impossible to sustain 24/7.
Too few personnel to cover all critical zones and blind spots simultaneously.
Remote location with weak or no broadband - cloud-based solutions completely ruled out.
No physical space for a dedicated server room or traditional desktop infrastructure.
Zero automated night detection - threats could go entirely unnoticed during off-hours.
What TDI Built
On-edge YOLO models auto-detect threats on tactical walking paths - no human in the loop.
Centralized PyQt multi-camera grid - one operator covers every zone from a single screen.
100% air-gapped architecture - all processing and logging runs on an isolated local LAN.
Entire pipeline on a single NVIDIA Jetson Orin Nano - compact, low-power, no server room needed.
Instant local alarm triggers - sirens and light panels fire the moment a threat is identified.
Business Impact
100% Automated Detection
Replaced manual surveillance on all high-risk walking paths with local YOLOv8 frame processing - threats flagged instantly, no operator required.
Multiplied Personnel Reach
Limited on-site staff now cover every critical blind spot simultaneously from a single unified display - dramatically reducing manpower gaps.
Zero Internet Dependency
Fully operational in remote locations with no broadband. All threat data and video logs secured inside the local environment with zero cloud-leak risk.
Zero Server Room Footprint
The entire surveillance pipeline runs on a single low-power edge device - no racks, no cooling, no dedicated room required.
Autonomous UAV Response
On high-priority threat detection, a tactical drone auto-deploys and tracks the target across rugged terrain - continuous visual reconnaissance without human dispatch.
12-15 Days New Features
Modular architecture allows new detection capabilities to be deployed within 12-15 working days, making the system continuously improvable in the field.
Core Technologies
Team Involvement
Development Phases
Data Collection & Annotation
Collected, pre-processed, and annotated the on-site tactical dataset across 8 weeks - building the foundation for accurate on-edge threat detection in real military environments.
Model Training & Evaluation
Trained, rigorously evaluated, and finalized the YOLOv8-based threat detection model - validated against real tactical footage to ensure reliable ally vs. threat classification.
Architecture & Pipeline Design
Designed the full system blueprint, RTSP streaming pipeline, and inference architecture before a single line of code was written - ensuring a clean, scalable foundation.
Core System Development
Built the full codebase, multi-camera PyQt UI, threat severity engine, and real-time distance measurement system - transforming the architecture blueprint into a working surveillance platform.
IoT Integration & On-Site Deployment
Connected AI outputs to physical hardware - sirens, light panels, and the autonomous UAV system. On-site engineers mounted cameras on trees and ran cable through rugged terrain to complete the live deployment.
Core Features
AI Threat Detection
On-edge YOLO models scan tactical walking paths and distinguish allies from armed threats in real time - dynamically scoring each incident as High, Medium, or Low based on weapon lethality and proximity.
Virtual Fences
Operators draw digital tripwires directly on live RTSP feeds - any person or vehicle crossing the line triggers an immediate alert, no manual monitoring required.
Crowd Density Monitoring
Detects sudden surges in people within a zone and flags anomalies that frequently precede physical altercations, stampedes, or coordinated security breaches.
Autonomous UAV Dispatch
On a high-priority threat alert, a tactical drone auto-launches and receives live threat coordinates - tracking and maintaining continuous visual reconnaissance across rugged terrain without human dispatch.
Blind-Spot Elimination Grid
A low-latency PyQt multi-camera matrix consolidates every critical zone into a single unified display - allowing limited personnel to monitor the full perimeter simultaneously.
Air-Gapped Recording
Continuous secure local storage with automatic management - all video logs and incident data stay entirely on-premise with absolute zero cloud exposure or external network risk.
Future Prospects
Loitering detection - flags any person or vehicle remaining in a restricted zone beyond a predefined time threshold
Abandoned object detection - identifies static items not part of the original scene background, ideal for spotting potential security risks
RTSP blindness alert - auto-detects if a camera feed goes black or a lens is deliberately covered, a common precursor to a security breach
Multi-tiered behavioral threat scoring - cross-references visual cues with temporal behavior patterns like loitering, sudden movement toward entry points, or static aiming
Scalable rollout across additional installations - modular architecture enables new detection features to be deployed within 12-15 working days at any new site