Secai
BetaCV + LLM vision for security camera analysis
Automated security camera video analysis combining computer vision with LLM verification. Detects objects of interest in surveillance footage with 99.9% accuracy across thousands of recordings.
About
Secai analyzes security camera footage to detect objects of interest using a two-stage pipeline. First, OpenCV does the heavy lifting — motion isolation via background subtraction, then template matching against known targets. Detections that land in a confidence gray zone get escalated to LLM vision (Claude or GPT) for a second opinion.
The motion isolation step is what makes it work. Subtracting the static background from each frame and masking the moving object gives template matching something clean to work with. Without it, accuracy drops off a cliff.
Runs against Tapo Cloud cameras, processes video with FFmpeg, parallelizes across 10 workers. Scans 1,200+ videos in about 20 minutes. Results get cached as JSON so historical queries are instant.