An Adaptive Video Surveillance Architecture for Behavior Analysis

Abstract

Adaptivity to scene changes is a main requirement for video analysis. The interpretation of video streams can be dealt by triggering different techniques depending on the scene properties. We present a work-on-progress for the design of a video surveillance architecture where different tasks in the context of behavior analysis are addressed, depending on the crowd level. A coarse estimation of the scene occupancy allows us to focus on single person or groups, adopting appropriate strategies to model the dynamic information. This paper focuses in particular on the crowd estimation problem: we propose a solution to detect and localize groups of people, able to provide an estimate of the number of people in the scene.