LMM-integration inom videoövervakning : Förutsättningar för ekonomisk hållbarhet och flexibel arkitektur

Sammanfattning

Video surveillance is undergoing a shift as Large Multimodal Models (LMMs) enable systems to reason about events and understand context in natural language rather than only detecting motion or classifying objects. However, there is a lack of understanding of which use cases are economically and practically viable and how to design systems that handle different needs across environments. This study explores this through semi-structured interviews with five respondents representing different parts of the ecosystem, including a VMS platform CEO, a product manager at a security company, an IT technician in smart public transport, an elderly care worker, and a construction engineer. The results show that the economic viability of LMM-based video analysis is primarily determined by the relationship between trigger frequency and the cost of a missed event. Sterile environments with few triggers and environments where a single miss carries high costs appear as the most viable use cases. Based on the interviews and principles from Clean Architecture, a requirements model is proposed where technical details can be delayed and configured per use case.