ThreatSnap

ThreatSnap monitors CCTV footage to detect suspicious/dangerous human activity using AI. It can run locally, provides analytics, sends threat alerts, and can be connected to an alarm system.

Comment

Inspiration

I wanted to build something under the Security & Online Safety theme since I hadn’t explored that domain before. The idea for ThreatSnap came while brainstorming how the GPT-4o model could be used practically in surveillance and security. I realized combining it with real-time computer vision could lead to a powerful tool — and that’s how ThreatSnap was born.

What It Does

ThreatSnap is an AI-assisted video surveillance system that analyzes CCTV/video footage for human movement. When motion is detected:

  • It captures a screenshot of the scene.
  • The screenshot is passed to GPT-4o, which analyzes the image for threat indicators like posture, weapons, or dangerous behavior.
  • If a threat is detected, it:
    • Sends an email alert with analysis.
    • Triggers an external alarm system (if configured).
    • Logs the detection with structured metadata.
    • Updates a dashboard with analytics such as threat levels, timelines, and object profiles.

How I Built It

  • Backend: Python with Flask.
  • Computer Vision: YOLOv8 (ultralytics) for detecting people and tracking movement.
  • AI Integration: GPT-4o for image-based threat analysis via OpenAI API.
  • Email Alerts: Configurable SMTP email alerts with screenshots attached.
  • Data Handling: JSON-based logs with timestamped screenshots.
  • Dashboard: HTML/CSS frontend with live feed, logs, and analytics.
  • Modular Architecture: Components like detector.py, emailer.py, processor.py are well-separated.

Challenges We Ran Into

  • Learning how to use YOLOv8 effectively in a resource-optimized way.
  • Ensuring screenshot quality and relevance — filtering only meaningful frames.
  • Managing performance and cost while using a cloud-based LLM (GPT-4o).
  • Handling screenshot analysis rate and preventing false positives via zone filtering and movement thresholds.

Accomplishments That We're Proud Of

  • Built a fully working MVP with real-time movement detection, AI-driven threat assessment, and a functional web dashboard.
  • Integrated an option for external system support (alarms), which shows potential for deployment in real-world CCTV setups.

What We Learned

  • Gained hands-on experience with YOLOv8, threaded detection loops, and video frame handling.
  • Learned to use GPT-4o multimodal capabilities for visual reasoning and structured JSON responses.

What's Next for ThreatSnap

  • Run completely offline using a local LLM and YOLO models for deployment in sensitive or remote areas.
  • Add real-time CCTV support instead of static videos.
  • Optimize detection to only capture frames that suggest danger, reducing API calls and noise.
  • Add user authentication, mobile support, and role-based access for multi-user environments.
  • Create a custom hardware integration pipeline for embedded systems and smart surveillance hardware.

Built With

Share this project:

Updates