StationWatch

Computer vision · runs in your browser

Catch stopped escalators in CCTV footage

StationWatch watches a video of an escalator and tells you, second by second, whether it is working, stopped with people on it or idle, with an event log and snapshots of every fault.

  • Private: video never leaves your device
  • YOLO11 + optical flow
  • Free and open source
Synthetic test scene, annotated by the monitor: the belt stops with people on it and a fault is raised.

Try it

Three steps. Heavy lifting (about 30 MB of models and libraries) happens on your machine and is cached after the first run.

1Choose a video

How it works

Classical computer vision on top of a pretrained detector: no training data, and every decision can be traced back to numbers on screen.

  1. Find the people

    YOLO11n detects people inside the marked region every few frames, and a small tracker keeps their boxes in between.

  2. Measure the surface

    Dense optical flow measures how the handrails and steps move, with people masked out, so someone walking on a stopped escalator does not count.

  3. Sanity checks

    Real escalator motion is mostly vertical and both handrails move the same way. Anything else is rejected.

  4. A stable decision

    A state machine with separate enter and exit thresholds turns about two seconds of evidence into a steady state, without flicker.

Working the surface is moving, with or without passengers.
Stopped / fault people are on it but it is not moving.
Idle nothing moves and nobody is there, e.g. an energy-saving stop.

The full Python version adds live RTSP cameras, webhook alerts, a command-line tool and a test suite. See the source on GitHub.