VisionGaurd.ai

Turn Your CCTV into a Real-Time Privacy Consious Theft Detection Solution

Comment

Inspiration

We own a grocery shop. One day, while reviewing our CCTV footage, I noticed a guy casually walk in, pick up a few items, slip them into his pockets and bag, and walk out without paying.

We have CCTV cameras. The cameras had captured everything — yet they did nothing.
They just watched.

That’s when it struck me: traditional CCTV cameras are incredibly dumb.
They record, but they don’t understand.

As we dug deeper, we realized this wasn’t just our problem.
Small and medium enterprises (SMEs) worldwide lose ~0.8% of total sales to theft, adding up to over \$200 billion annually, with total retail shrinkage nearing \$1 trillion.

Possible solutions?

  • Hiring more security guards → expensive and old-school
  • Buying smart AI cameras → effective, but extremely costly
  • Replacing existing CCTV systems with AI cameras → even more expensive due to hardware and opportunity cost

So we asked ourselves:
What if the cameras people already have could actually understand what’s happening?

That question led to VisionGuard.ai.

What it does

VisionGuard.ai turns existing CCTV cameras into a real-time, intelligent theft and anomaly detection system—without replacing hardware and without using facial recognition.

Instead of analyzing raw video frames, VisionGuard.ai focuses on human behavior.
It learns what normal shopping behavior looks like and flags statistically abnormal actions—such as suspicious movement or item concealment—in real time.

Core features:

  • Works on existing CCTV systems
  • Privacy-preserving pose-based behavior analysis
  • Real-time anomaly alerts via dashboard and Telegram
  • Multi-shop and role-based access (Owner / Manager)
  • After-hours intrusion detection (Owl-Eye Mode)
  • Human-in-the-loop validation to reduce false alarms

How we built it

VisionGuard.ai is designed to be cost-aware and deployment-friendly.

We split computation intelligently:

  • Client-side
    • Person detection and tracking
    • Human pose estimation
    • Temporal pose sequence generation
  • Server-side
    • Anomaly scoring using a Spatio-Temporal Graph Neural Network with Normalizing Flow (STG-NF)
    • Incident management and alert delivery

The simplified pipeline: [ \text{CCTV Stream} \rightarrow \text{Person Detection} \rightarrow \text{Tracking} \rightarrow \text{Pose Estimation} \rightarrow \text{STG-NF Anomaly Scoring} \rightarrow \text{Alerts} ]

By modeling behavior instead of pixels, we reduce server cost, improve privacy, and remain robust across different camera setups.

Challenges we ran into

  • Making AI accurate yet affordable for SMEs
  • Reducing false positives from normal shopping behavior
  • Maintaining real-time performance without GPUs
  • Pose estimation errors due to occlusion and camera distance
  • Integrating streaming, AI inference, alerts, and dashboards into one stable system

Accomplishments that we're proud of

  • Built a fully working real-time system, not just a research model
  • Enabled AI surveillance without camera replacement
  • Eliminated the need for facial recognition
  • Reduced server costs through smart compute distribution
  • Delivered instant alerts with evidence storage
  • Designed a system usable by non-technical shop owners

What we learned

  • Real-world AI is about constraints, not just accuracy
  • Behavior-based representations are powerful and privacy-friendly
  • Alert quality matters more than alert quantity
  • System architecture choices directly impact business viability
  • End-to-end integration is often harder than model training

What's next for VisionGuard.ai

  • Adaptive per-shop anomaly thresholds
  • Continuous learning from validated incidents
  • Better robustness in crowded and occluded scenes
  • Edge-device deployment for even lower latency
  • Expansion beyond retail to warehouses and campuses

Business Model

VisionGuard.ai follows a monthly subscription model, making advanced AI surveillance affordable for small and medium businesses.

Our go-to-market strategy includes partnering with local CCTV vendors, allowing VisionGuard.ai to be offered as an add-on service to existing installations.
This ensures fast adoption, smooth installation, and accessibility for non-technical users.

VisionGuard.ai doesn’t just watch.
It understands.

Built With

Share this project:

Updates