A new research paper proposes a way to turn long, messy video of people working - captured from both a worker's own camera and overhead workplace cameras - into a compact, searchable record of what actually happened.
The system breaks continuous footage into segments whenever something meaningful changes: the visual scene, the person's location, their motion, what they're saying, or what object they're interacting with. Each segment becomes an "event card" noting who did what, when, where, with which tools, and what state things were in before and after, along with a confidence score and a link back to the source footage. These cards accumulate into what the researchers call a Work Environment Model, built using frozen DINOv2 and VJEPA-2 vision encoders paired with a local language model. Everything runs on-premise, so raw video never has to leave the building.
The real contribution isn't the cameras, it's the compression. Industrial sites, warehouses, and labs already drown in video they can't afford to store, review, or search; this approach promises an auditable paper trail instead of a server full of footage nobody watches. If it works as described, it's a more honest alternative to AI-monitoring pitches that quietly imply constant human surveillance.
It's worth noting this is a framework paper with proposed evaluation criteria, not benchmark results, so the real test, how well it compresses and retrieves across long, messy shifts, is still ahead of it.