This is the real output of Crossing Watch on our college gate: every near-miss candidate SAM 3.1 tracking found in 8.6 minutes. Move the threshold and switch Muse Spark's review off and on, and the counts change by the same rules the pipeline uses. Click any event to replay it on the road map. Most candidates are the motorcycle's own rider, whom SAM also tracks as a person; Muse Spark looks at each one and removes them.
Try it: switch Muse Spark review off and the 5 near-misses become 75 alarms; switch it back on.
The measured stretch seen from above, built from the 4 marked corners (we used A4 sheets). Cyan: the person's path. Red: the vehicle's. The ring marks the shared spot.
| # | At | Tier now | PET | Vehicle | Speed | Muse Spark |
|---|
Python 3.11+ and ffmpeg. The self-checks and the synthetic sample need no key; your own video needs a Meta Model API key.
git clone https://github.com/shaunmortal/crossing-watch && cd crossing-watch python3 test_cw.py # self-checks, under a minute python3 pack.py sample sample/site.json # the synthetic sample, no key python3 monday.py your-video.mp4 sites/your-site.json --budget 14 --yes # your crossing
This page replays recorded results; it calls no model. Rules: serious under the threshold you set, close pass under 3.0 s; a track Muse Spark calls a rider or not a person is removed from every event, as pack.confirmed does.