A stochastic approach to tracking objects across multiple cameras

Abstract

This paper is about tracking people in real-time as they move through the non-overlapping fields of view of multiple video cameras. The paper builds upon existing methods for tracking moving objects in a single camera. The key extension is the use of a stochastic transition matrix to describe peoples observed patterns of motion both within and between fields of view. The parameters of the model for a particular environment are learnt simply by observing a person moving about in that environment. No knowledge of the environment or the configuration of the cameras is required.