Skip to Main Content

NearMiss

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

NearMiss — Demo video

NearMiss measures near-misses between vehicles and pedestrians at a street crossing from an ordinary phone video. Road crashes kill 1.19 million people a year, a quarter of them pedestrians, mostly in low- and middle-income countries, and crossings usually get fixed only after a death. Near-misses are the early warning, but measuring them normally takes special cameras and traffic consultants. NearMiss tracks every road user with SAM 3.1 and turns each one's contact with the road into metres. The scale comes from the zebra stripes or, where there are none, from the pedestrians themselves. It then computes the standard safety measures: post-encroachment time and time to collision. Muse Spark names the road users to track and reviews every flagged moment from a highlighted clip, confirming it or rejecting it as a measurement error. A web app replays each analysis with SAM's outlines, a to-scale map, and the evidence for each near-miss. A 30-second clip costs about $0.20 to analyse.

Team