Analyse CCTV footage with an LLM — motion-only frames, audit trail

Steps

1. Install peepshow

   npm install -g peepshow

2. Run with motion-only preset

   Scene detection + perceptual-hash dedup drops static noise.

   peepshow ./overnight.mp4 --strategy scene --max 30 --dedup perceptual

3. Push to a durable sink

   SQLite or S3 keeps the frame archive searchable.

   peepshow ./overnight.mp4 --sink sqlite --sink s3

4. Hand frames to the LLM

   Drag the frame folder into Claude / ChatGPT for a per-frame review.

   peepshow ./overnight.mp4 --emit json | claude-cli -p 'List motion events'

Why it works

Surveillance footage has a poor signal-to-noise ratio. Native video upload to Gemini would bill ~258 tokens/second — an overnight 12-hour clip = 11M tokens, well over context limits. Scene detect picks frames where the visual entropy spikes; perceptual dedup drops the duplicates of an empty hallway. Result: hours of footage compressed to a handful of frames an LLM can actually reason about.

When it helps

• Overnight security camera review without re-watching at 1×.
• Dashcam incident extraction.
• Industrial / lab timelapses where the interesting moments are sparse.

Pitfalls

• Scene threshold is configurable (`--scene-threshold 0.3` default); raise for noisier footage.
• Perceptual dedup uses pHash by default — works on luma changes, less reliable on pure colour shifts.
• Long clips: chunk via `--start` / `--duration` if extracting in shifts.

Works with these LLMs

Pairs with these sinks