Person tracking in camera networks using graph-based bayesian inference

In this paper, a probabilistic approach for tracking multiple persons through a network of distributed cameras is presented. The approach deals with the main problems associated with the tracking of persons through wide area networks-bridging large observation gaps between camera views and re-identifying persons-by building on robust and view-invariant high-level features as well as a highly error-tolerant probabilistic filtering of person locations. The extraction quality and discriminative power of the proposed features is evaluated on realistic data including well-known and established benchmark datasets. A comparative performance analysis is then made to assess the accuracy of the probabilistic inter-camera tracking method given a number of different simulated and real quality levels of the underlying person detection and feature extraction components. The experiments are made using a simulated virtual environment involving multiple persons in an indoor surveillance scenario.