What it means
Classic video analytics detect a fixed list of classes, such as car, truck, helmet or seatbelt. A VLM accepts open questions. An operator can type a description, such as a white truck with a damaged side or a person in a red cap, and get matching frames without anyone having trained a model for that exact class.
VLMs are large, so running one on a roadside camera needs a capable edge processor and careful model optimisation. They also need limits. A search result is a lead for an investigator to check, not evidence on its own.
How it applies in India
Indian roads add text in several scripts on vehicles, shop signs and number plates. A VLM that understands English, Hindi and Marathi lets an investigator search for text written on a vehicle, not only for its plate. Stellarview AI Search uses a VLM on the Recon camera's edge processor and supports queries in English, Hindi and Marathi.