PhD forum: Non supervised learning of human activities in Visual Sensor Networks

Abstract

We outline how Human Activity Recognition systems based on Dynamic Bayesian Networks using a single camera may be adapted to be used in Visual Sensor Networks. It is assumed that current activity generates independent observations on some cameras in the network. Then, the activity is inferred by the accumulation of the evidences provided by the observations gathered. At the same time, some activities never produce observations on some cameras. Baum-Welch algorithm is modified to deal with this situation, providing some examples of when it converges.