Real-Time Video Monitoring & Detection

A real-time video pipeline that watches a live camera feed, finds what it cares about, and decides whether what it's looking at is normal.

It started as straightforward object detection — Haar Cascade classifiers over an OpenCV capture loop, with a lot of preprocessing tuning to get reliable hits under changing light. The interesting part came second: swapping "is there a thing here" for "does this look right," using PatchCore via Anomalib on top of PyTorch. The concrete test case was a bed. Made, or not made?

That turns out to be a genuinely good anomaly-detection problem. There's no clean bounding box to draw, the difference is textural, and "unmade" has infinite variations while "made" has only a few. So the model learns what made looks like and scores everything by how far it deviates.

Getting it onto a Pi

Training on a laptop is one thing; inference on a Raspberry Pi is another. The work there was:

• Collecting and labeling my own dataset, since nothing off the shelf covers this.
• Feature extraction and threshold calibration so the anomaly score maps onto an actual decision rather than a number nobody can interpret.
• Trimming the inference path to fit the Pi's memory and CPU budget.
• Encrypting the video stream in transit and at rest — it's a camera in a bedroom, so that isn't optional.

Structure

I split it into data ingestion, preprocessing, inference, and alerting as separate modules. That sounds like over-engineering for a personal project until you're three weeks in trying to work out whether a bad result came from the camera, the crop, or the model. Being able to test each stage on its own is what made it debuggable.

A later version added a two-run confirmation step: it only alerts when two consecutive checks agree, which killed nearly all of the false positives from someone just walking through frame.