CCTV Footage Analysis AI

• 

  Set Location for the video

• 

  Sync CCTV video to 2D Google map, timeline + report

• 

  Pick points for homography

• 

  Doppler Building

• 

  Report PDF Export

• 

  Architecture Diagram

Inspiration

I watched how expensive and slow current “video understanding” models run, and after watching some car accident lawyer report generation tool, I decided to create an affordable and computation-cheap and agentic way to generate the same report from just video.

Similar manual tool: https://pc-crash.com

Inspiration

Most “video understanding” AIs are expensive, slow, and impractical for real-world use. After seeing how accident reconstruction tools and lawyer reports still rely on manual video analysis, I wanted to build a faster, cheaper, and agentic way to generate the same quality of insight — directly from raw CCTV footage.

(Similar manual tool: PC-Crash, which inspired me to build an AI-driven alternative that works on everyday hardware.)

What it does

A tool that turns a short CCTV clip into a structured report:

• Tracks vehicles and estimates speed in mph (bounding boxes, speed smoothing)
  
• Generates a 2D Google Map replay that visualizes trajectories
  
• Uses Claude (via AWS Bedrock) as an orchestrator to decide what tools to call to narrate the event timeline
  
• Exports the full analysis as a professional-looking PDF report
  

How I built it

• Designed a reliable visual workflow locally using YOLOv8 + ByteTrack + classic vision algorithms
  
• Built a full-stack web app where users can log in, upload footage, and view results
  
• Hosted the heavy visual workflows on async workers for scalable background processing
  
• Integrated Claude (via AWS Bedrock) as the “agent” to coordinate the end-to-end pipeline
  

Challenges we ran into

• Out-of-the-box detection/tracking models struggle on blurred, low-res, or night-time traffic footage
  
• Mapping angled CCTV footage onto a 2D Google Map using homography
  
• Creating a readable, time-aligned event timeline from raw frame data
  
• Designing an intuitive way for users to explore and verify visual results
  

Accomplishments that I'm proud of

• Combined and ranked multiple visual algorithms (image enhancement, boundary detection) to improve tracking accuracy and speed estimation
  
• Created an intuitive calibration UI to easily map CCTV footage to 2D Google Maps
  
• Delivered a working prototype that can reconstruct a crash timeline in minutes, not hours
  

What I learned

• Many “hard” (hard to me) AI problems can be solved well with existing algorithms when combined intelligently
• Breaking complex tasks into smaller, deterministic steps is far more effective than chasing “end-to-end” magic (I'm learning my lessons a few times after seeing over-promising AI demos)

What's next for CCTV Footage Analysis AI

• Automate homography calibration using AI + image matching to make the tool fully hands-off
  
• Add data correction and adjustment tools to increase precision for professional users
  
• Extend the system to other use cases — industrial safety, sports analysis, and traffic optimization
  

Built With

• agentcore
• bedrock
• celery
• codex
• cursor
• docker
• fastapi
• postgresql
• python
• react
• s3
• supervision
• weatherapi