Privacy-preserving computer vision for near-miss detection at crossings, surfacing dangerous intersections before they produce a casualty record.
Overview
We usually learn an intersection is dangerous from its crash history — a grim, lagging indicator. This project built a privacy-preserving vision pipeline that detects near misses, giving planners a leading signal and the chance to intervene before harm occurs.
Key Ideas
- On-device processing; no identifiable imagery ever leaves the sensor.
- Near-miss severity scoring calibrated against expert review.
- A ranked watch-list of intersections for proactive redesign.
Findings
Near-miss frequency correlated strongly with subsequent incident rates, supporting its use as an early-warning metric. Two flagged intersections were redesigned ahead of any recorded casualty.
Collaborators on this project
PS
Prof. Sarah Chen
Massachusetts Institute of Technology
DA
Dr. Amara Diallo
Stanford University