[OSCE Lunch meetup] From Public Video to Open Mobility Data: Reproducible Global Traffic Research with YouTube and AI

Register here: https://forms.cloud.microsoft/e/2zQ2wx4L6A

Public video platforms contain an enormous amount of naturalistic mobility data, but turning this material into reproducible research requires much more than simply downloading videos. In this talk, I will present our work on building scalable, transparent, and reusable mobility datasets from publicly available YouTube footage.

I will introduce CROWD, a global dataset of continuous urban driving covering more than 33963.47 hours of footage across 9,868 localities in 238 countries and territories, together with structured metadata and automatically generated object detections and tracks. I will then show how this resource can support global behavioural research through our Pedestrian Planet study, which analyses pedestrian crossing behaviour across countries using measures such as crossing speed and crossing initiation time.
Finally, I will discuss our more recent work on automated and continuously updateable dataset construction, where vision language models, temporal segmentation, geographic grounding, provenance tracking, and versioned outputs are combined to transform public video into structured research data.

Throughout the talk, I will focus on the open science challenges behind this work, including reproducibility, provenance, dataset versioning, copyright and privacy, transparent automated annotation, and how public web data can be made useful for scientific research without redistributing the underlying videos.